Mario plus Red Dead Redemption?! Reverse Engineering AI

1
00:00:02,540 --> 00:00:03,840
Welcome everyone to the Gen
2
00:00:04,480 --> 00:00:05,260
Meetup podcast.
3
00:00:05,440 --> 00:00:06,600
I'm one of your hosts, Shashank,
4
00:00:06,800 --> 00:00:07,860
and I have Mark with me
5
00:00:07,900 --> 00:00:08,100
here.
6
00:00:08,580 --> 00:00:10,800
We're both ex senior software engineers
7
00:00:10,900 --> 00:00:12,000
at faang companies
8
00:00:12,400 --> 00:00:13,840
at Google and Amazon and now
9
00:00:13,880 --> 00:00:15,260
we're doing our own thing.
10
00:00:15,560 --> 00:00:17,020
Mark is now the founder
11
00:00:17,560 --> 00:00:19,680
of an AI native video editing
12
00:00:19,740 --> 00:00:20,920
startup called Nova Cut.
13
00:00:21,300 --> 00:00:21,720
Check it out,
14
00:00:21,860 --> 00:00:22,680
links will be below.
15
00:00:23,120 --> 00:00:24,280
And I'm working on a bunch
16
00:00:24,320 --> 00:00:24,960
of ideas.
17
00:00:25,060 --> 00:00:26,700
Check out our substack where I
18
00:00:26,760 --> 00:00:29,980
post our experiments and uh,
19
00:00:30,220 --> 00:00:30,800
our findings.
20
00:00:31,020 --> 00:00:31,600
But today
21
00:00:32,060 --> 00:00:33,060
we have a bunch of news
22
00:00:33,120 --> 00:00:33,960
that we're going to talk about
23
00:00:34,020 --> 00:00:35,120
since our last episode.
24
00:00:35,700 --> 00:00:35,840
Uh,
25
00:00:35,980 --> 00:00:39,180
from new models to bunch uh,
26
00:00:39,181 --> 00:00:41,000
of controversy with the anthropic
27
00:00:41,380 --> 00:00:42,280
going to IPO at a $2
28
00:00:42,340 --> 00:00:44,280
trillion valuation, uh,
29
00:00:44,480 --> 00:00:45,860
claiming that they're profitable,
30
00:00:46,060 --> 00:00:48,540
but actually not according to standard
31
00:00:48,600 --> 00:00:49,740
gap reporting principles.
32
00:00:50,380 --> 00:00:50,860
Um,
33
00:00:51,580 --> 00:00:52,080
and
34
00:00:52,920 --> 00:00:54,580
personally something that I'm pretty excited
35
00:00:54,640 --> 00:00:57,680
about is how AI is affecting
36
00:00:58,180 --> 00:01:00,120
software engineering with the ability to
37
00:01:00,160 --> 00:01:01,680
reverse engineer anything.
38
00:01:02,140 --> 00:01:03,900
This new GitHub repo uh,
39
00:01:04,020 --> 00:01:05,460
that everyone's been playing around with
40
00:01:05,540 --> 00:01:05,960
on Twitter
41
00:01:06,480 --> 00:01:08,940
and reverse engineering literally everything that
42
00:01:08,980 --> 00:01:10,040
they can get their hands on.
43
00:01:10,400 --> 00:01:11,240
If it's on your laptop,
44
00:01:11,320 --> 00:01:12,500
if it's on your phone,
45
00:01:12,680 --> 00:01:13,740
you can reverse engineer it,
46
00:01:14,120 --> 00:01:14,900
you can clone it,
47
00:01:15,160 --> 00:01:16,140
build your own copy,
48
00:01:16,640 --> 00:01:19,160
modify it and take just basically
49
00:01:19,220 --> 00:01:20,520
get that functionality for free.
50
00:01:21,920 --> 00:01:22,280
Um,
51
00:01:23,100 --> 00:01:25,540
there's this really popular repo on
52
00:01:25,600 --> 00:01:27,380
GitHub which is racked up.
53
00:01:28,180 --> 00:01:29,000
Wow, it was uh,
54
00:01:29,260 --> 00:01:32,200
100,000 stars a little while back.
55
00:01:32,320 --> 00:01:32,800
Now it's
56
00:01:33,780 --> 00:01:35,860
337,000 stars that it's racked up
57
00:01:35,920 --> 00:01:36,460
on GitHub
58
00:01:36,860 --> 00:01:38,860
which allows you to reverse engineer
59
00:01:39,340 --> 00:01:41,580
any app binary that you download
60
00:01:41,640 --> 00:01:43,000
from the App Store on your
61
00:01:43,040 --> 00:01:44,140
laptop, wherever.
62
00:01:44,640 --> 00:01:45,640
But honestly you don't even need
63
00:01:45,700 --> 00:01:46,340
to use this.
64
00:01:46,500 --> 00:01:48,820
Just point your agent at uh,
65
00:01:48,860 --> 00:01:51,080
Microsoft and have it reverse engineer
66
00:01:51,140 --> 00:01:52,180
that entire software,
67
00:01:52,640 --> 00:01:53,800
which is what people have been
68
00:01:53,840 --> 00:01:54,240
doing.
69
00:01:54,780 --> 00:01:55,980
Um, so I think like ah,
70
00:01:56,220 --> 00:01:58,160
one of the popular YouTubers, uh,
71
00:01:58,240 --> 00:02:00,000
maybe like Linus from Tech Tips
72
00:02:00,040 --> 00:02:01,700
or someone has
73
00:02:02,700 --> 00:02:03,920
had their entire team
74
00:02:04,380 --> 00:02:06,760
test out the Photoshop clone and
75
00:02:06,880 --> 00:02:07,780
actual Photoshop
76
00:02:08,240 --> 00:02:09,320
and see how it compares.
77
00:02:09,360 --> 00:02:11,200
And it's actually pretty decent.
78
00:02:11,580 --> 00:02:13,040
I mean obviously it isn't perfect
79
00:02:13,100 --> 00:02:14,940
but it gets you 80% of
80
00:02:15,000 --> 00:02:15,580
the way there,
81
00:02:16,040 --> 00:02:16,420
which is
82
00:02:17,000 --> 00:02:18,500
good enough for most people who
83
00:02:18,580 --> 00:02:20,540
are just up to their wits
84
00:02:20,620 --> 00:02:21,460
end with uh,
85
00:02:21,640 --> 00:02:24,960
Adobe and their ridiculous cancellation fees.
86
00:02:25,340 --> 00:02:26,580
So I remember signing up for
87
00:02:26,640 --> 00:02:27,720
Adobe Photoshop like
88
00:02:28,300 --> 00:02:29,540
several years ago right when I
89
00:02:29,580 --> 00:02:30,500
was right out of college.
90
00:02:30,900 --> 00:02:31,420
I didn't have a lot of
91
00:02:31,480 --> 00:02:32,600
money and I just wanted to
92
00:02:32,700 --> 00:02:34,040
use something for some
93
00:02:34,700 --> 00:02:35,840
project that I was working on
94
00:02:36,480 --> 00:02:38,020
and I canceled it after a
95
00:02:38,060 --> 00:02:38,960
month and they charged me like
96
00:02:39,320 --> 00:02:39,940
$67
97
00:02:40,560 --> 00:02:41,740
which was a lot for For
98
00:02:41,780 --> 00:02:43,580
a College student pre AI,
99
00:02:43,840 --> 00:02:45,880
when SaaS subscriptions were pretty cheap
100
00:02:46,280 --> 00:02:47,560
and people weren't spending that much
101
00:02:47,620 --> 00:02:48,340
on software.
102
00:02:48,860 --> 00:02:50,440
And so it's,
103
00:02:50,480 --> 00:02:51,540
it's kind of become like this
104
00:02:51,640 --> 00:02:52,860
comcast of uh,
105
00:02:53,000 --> 00:02:55,400
SaaS companies where they're gouging as
106
00:02:55,440 --> 00:02:56,480
much as they can out of
107
00:02:56,540 --> 00:02:58,420
you for their subscription based model,
108
00:02:58,480 --> 00:02:58,660
which
109
00:02:59,280 --> 00:03:00,600
it shouldn't be a subscription because
110
00:03:00,660 --> 00:03:03,040
it's like a fixed software which
111
00:03:03,100 --> 00:03:05,660
doesn't have any um, uh,
112
00:03:05,700 --> 00:03:08,100
like negligible ongoing costs.
113
00:03:08,220 --> 00:03:08,940
And so people have been
114
00:03:09,940 --> 00:03:12,100
building their own versions of Photoshop
115
00:03:12,220 --> 00:03:13,520
and Microsoft Office.
116
00:03:14,120 --> 00:03:14,420
Um,
117
00:03:14,600 --> 00:03:15,280
and I feel like this is
118
00:03:15,320 --> 00:03:15,980
the new normal.
119
00:03:16,440 --> 00:03:17,780
Like what do you think Mark?
120
00:03:18,760 --> 00:03:18,960
Yeah,
121
00:03:19,040 --> 00:03:20,540
there's a lot to unpack there.
122
00:03:20,760 --> 00:03:22,160
So uh, speaking of,
123
00:03:22,340 --> 00:03:23,760
if you're looking for a better
124
00:03:24,100 --> 00:03:26,160
video editor uh, than Premiere Pro,
125
00:03:26,300 --> 00:03:27,440
you could check out Nova Cat,
126
00:03:27,520 --> 00:03:28,220
that AI.
127
00:03:28,520 --> 00:03:28,640
Um,
128
00:03:28,760 --> 00:03:30,040
so that is what I am
129
00:03:30,080 --> 00:03:31,380
building and you can think of
130
00:03:31,440 --> 00:03:32,940
it kind of like uh,
131
00:03:33,440 --> 00:03:35,820
Premiere Pro but with AI.
132
00:03:35,920 --> 00:03:36,620
So you can just throw in
133
00:03:36,680 --> 00:03:37,780
all your videos and have it
134
00:03:37,840 --> 00:03:38,700
edited for you.
135
00:03:39,120 --> 00:03:40,620
It does have a subscription but
136
00:03:40,720 --> 00:03:41,560
I will not charge you a
137
00:03:41,620 --> 00:03:42,660
cancellation fee if you want to
138
00:03:42,720 --> 00:03:44,740
cancel. So um, there we go.
139
00:03:45,180 --> 00:03:47,600
Anyways, with that being said, um,
140
00:03:47,840 --> 00:03:47,980
yeah,
141
00:03:48,820 --> 00:03:52,840
Reverse Engineer anything is incredibly interesting.
142
00:03:53,360 --> 00:03:53,640
So
143
00:03:55,740 --> 00:03:58,620
reverse engineering from a binary is
144
00:03:58,720 --> 00:03:59,360
super hard.
145
00:03:59,520 --> 00:04:00,460
So for those who don't know
146
00:04:00,500 --> 00:04:01,140
what we're talking about,
147
00:04:01,400 --> 00:04:03,790
a binary is the
148
00:04:04,270 --> 00:04:06,430
thing that you would use to
149
00:04:06,470 --> 00:04:07,210
run a program.
150
00:04:07,310 --> 00:04:08,210
So for example,
151
00:04:08,430 --> 00:04:10,670
if I go and download
152
00:04:11,150 --> 00:04:13,850
Photoshop or I don't know,
153
00:04:13,970 --> 00:04:15,190
like Chrome or whatever,
154
00:04:15,590 --> 00:04:16,910
oftentimes it's going to be in
155
00:04:17,010 --> 00:04:18,770
a single file that you can
156
00:04:18,830 --> 00:04:19,570
click on and run.
157
00:04:19,830 --> 00:04:23,770
And that file is uh,
158
00:04:24,210 --> 00:04:25,471
created from all,
159
00:04:25,530 --> 00:04:27,750
oftentimes millions of lines of source
160
00:04:27,810 --> 00:04:28,070
code.
161
00:04:28,390 --> 00:04:29,950
But like you can't actually read
162
00:04:30,010 --> 00:04:30,470
that code,
163
00:04:30,550 --> 00:04:32,170
it's just kind of um.
164
00:04:32,450 --> 00:04:35,710
I wouldn't say it's uh, encrypted,
165
00:04:35,830 --> 00:04:38,350
it's not encrypted but it's compressed.
166
00:04:38,590 --> 00:04:38,910
So
167
00:04:39,450 --> 00:04:40,650
for all intents and purposes you
168
00:04:40,750 --> 00:04:41,670
can't actually
169
00:04:42,190 --> 00:04:44,210
read it directly and going it's
170
00:04:44,230 --> 00:04:46,330
not human readable.
171
00:04:46,490 --> 00:04:48,050
Computers can read it but it's
172
00:04:48,130 --> 00:04:49,530
really difficult for humans to make
173
00:04:49,610 --> 00:04:50,390
out what it means.
174
00:04:50,890 --> 00:04:51,070
Yeah,
175
00:04:51,530 --> 00:04:53,150
it's like almost written in a
176
00:04:53,190 --> 00:04:53,910
different language.
177
00:04:54,430 --> 00:04:56,130
It's not binary but it's close
178
00:04:56,170 --> 00:04:56,790
to binary.
179
00:04:57,710 --> 00:04:59,530
Ah, really low level, uh,
180
00:04:59,650 --> 00:05:00,430
system instructions.
181
00:05:00,470 --> 00:05:01,290
So um,
182
00:05:01,770 --> 00:05:03,490
it's very hard to undo.
183
00:05:03,650 --> 00:05:04,550
It's like once it's in that
184
00:05:04,610 --> 00:05:05,810
binary it's very hard to go
185
00:05:05,850 --> 00:05:06,870
back to the source code.
186
00:05:06,970 --> 00:05:07,230
And
187
00:05:07,750 --> 00:05:08,850
if a human were going to
188
00:05:08,910 --> 00:05:09,170
do that,
189
00:05:09,230 --> 00:05:10,710
it's possible but
190
00:05:11,270 --> 00:05:11,790
at that point you
191
00:05:12,190 --> 00:05:13,330
be better off just trying to
192
00:05:13,390 --> 00:05:15,790
rewrite the entire software from scratch
193
00:05:15,870 --> 00:05:16,510
most likely.
194
00:05:16,590 --> 00:05:17,350
As opposed to trying to.
195
00:05:17,410 --> 00:05:19,910
People have been trying to decompile
196
00:05:20,030 --> 00:05:21,030
apps like uh,
197
00:05:22,390 --> 00:05:24,850
the PS1 OS or something.
198
00:05:25,030 --> 00:05:25,310
There's
199
00:05:26,010 --> 00:05:28,310
groups of people who spend years
200
00:05:28,390 --> 00:05:29,890
to try to decompile, you know,
201
00:05:29,970 --> 00:05:30,790
like the GameCube
202
00:05:31,230 --> 00:05:33,250
OS to run their favorite games
203
00:05:33,290 --> 00:05:34,050
that they played when they were
204
00:05:34,110 --> 00:05:35,310
a kid for nostalgia.
205
00:05:35,670 --> 00:05:36,390
And it takes a lot of
206
00:05:36,430 --> 00:05:36,710
work.
207
00:05:36,750 --> 00:05:38,130
It takes years and years and
208
00:05:39,270 --> 00:05:39,531
lots, uh,
209
00:05:39,531 --> 00:05:41,310
and lots of people combined effort.
210
00:05:41,710 --> 00:05:44,230
Yeah, yeah, it's very tedious,
211
00:05:44,590 --> 00:05:45,210
very hard.
212
00:05:45,330 --> 00:05:45,750
But now,
213
00:05:45,990 --> 00:05:46,850
as you were about to say,
214
00:05:46,890 --> 00:05:47,230
Shoshana,
215
00:05:47,590 --> 00:05:49,030
you could just throw an LLM
216
00:05:49,070 --> 00:05:51,370
at the problem and they will
217
00:05:51,450 --> 00:05:53,970
do the decompilation for you.
218
00:05:54,730 --> 00:05:56,070
You just pay a bit of
219
00:05:56,110 --> 00:05:56,970
money in tokens
220
00:05:57,550 --> 00:05:59,350
and the LLM will be able
221
00:05:59,490 --> 00:06:00,350
to figure it out.
222
00:06:00,450 --> 00:06:01,211
And you've, uh,
223
00:06:01,211 --> 00:06:03,290
been seeing some really interesting things.
224
00:06:03,330 --> 00:06:05,090
Shashank has shared his screen here
225
00:06:05,650 --> 00:06:06,130
that uh,
226
00:06:06,230 --> 00:06:09,850
people have made duplicates of
227
00:06:11,150 --> 00:06:14,650
Microsoft Word, uh, PowerPoint, Excel.
228
00:06:14,870 --> 00:06:16,990
I've seen people have been using
229
00:06:17,250 --> 00:06:19,190
Adobe, uh, tools where um,
230
00:06:19,290 --> 00:06:20,390
people have been uh,
231
00:06:21,050 --> 00:06:22,290
duplicating things like
232
00:06:22,670 --> 00:06:25,910
Photoshop or Premiere Pro, um,
233
00:06:26,490 --> 00:06:27,850
all kinds of stuff and just
234
00:06:27,930 --> 00:06:28,130
open,
235
00:06:28,210 --> 00:06:30,050
sourcing it and making it free,
236
00:06:30,650 --> 00:06:31,790
uh, for people to use.
237
00:06:32,010 --> 00:06:32,090
Uh,
238
00:06:32,310 --> 00:06:33,390
so it's
239
00:06:34,050 --> 00:06:35,670
really probably going to be a
240
00:06:35,710 --> 00:06:39,350
very scary time for Adobe because
241
00:06:39,430 --> 00:06:41,010
their moat is drying up and
242
00:06:41,070 --> 00:06:42,210
drying up very, very fast.
243
00:06:44,670 --> 00:06:44,790
Yeah,
244
00:06:45,250 --> 00:06:46,990
I think this is pretty cool,
245
00:06:47,430 --> 00:06:50,970
um, for the average consumer, um,
246
00:06:51,090 --> 00:06:52,370
but I honestly don't know what
247
00:06:52,390 --> 00:06:53,410
this means for
248
00:06:54,650 --> 00:06:56,350
the business owner because
249
00:06:57,030 --> 00:06:58,210
let me look at Adobe stock.
250
00:06:58,250 --> 00:07:00,570
I don't think it's reacted, uh,
251
00:07:01,110 --> 00:07:03,130
Adobe stock price hasn't quite responded
252
00:07:03,190 --> 00:07:05,210
to this new wave of decompilation.
253
00:07:05,910 --> 00:07:09,130
Um, and so legally I think,
254
00:07:09,370 --> 00:07:10,170
uh, from what I've read,
255
00:07:10,370 --> 00:07:12,230
it's fine to do this,
256
00:07:12,710 --> 00:07:14,230
but because you are not
257
00:07:14,890 --> 00:07:18,250
copying or stealing Adobe's source code,
258
00:07:18,910 --> 00:07:21,030
you are taking the app,
259
00:07:21,310 --> 00:07:22,790
the finished product that they've built,
260
00:07:23,350 --> 00:07:25,070
and kind of like reverse engineering
261
00:07:25,130 --> 00:07:25,790
what they've done.
262
00:07:26,110 --> 00:07:26,930
You look at uh,
263
00:07:27,210 --> 00:07:29,050
their final product, you look at,
264
00:07:29,390 --> 00:07:29,510
well,
265
00:07:29,590 --> 00:07:30,270
you can kind of look at
266
00:07:30,310 --> 00:07:31,070
their lines of code,
267
00:07:31,430 --> 00:07:33,650
but essentially you are rewriting that
268
00:07:33,710 --> 00:07:34,390
functionality
269
00:07:34,890 --> 00:07:36,290
and you're delegating that to an
270
00:07:36,410 --> 00:07:37,110
LLM, of course.
271
00:07:37,210 --> 00:07:37,350
But
272
00:07:37,710 --> 00:07:38,970
that LLM is really good at
273
00:07:39,030 --> 00:07:40,410
tedious things as we've seen in
274
00:07:40,430 --> 00:07:41,410
the way it's solved, uh,
275
00:07:41,890 --> 00:07:43,410
some of these millennium math problems
276
00:07:43,430 --> 00:07:44,890
which take um,
277
00:07:45,090 --> 00:07:46,650
like dozens of pages of proofs.
278
00:07:46,730 --> 00:07:48,370
It's very good at tedious,
279
00:07:48,870 --> 00:07:49,770
nitty gritty details
280
00:07:50,530 --> 00:07:51,410
and it's been able to
281
00:07:52,590 --> 00:07:54,690
tediously reconstruct the functionality.
282
00:07:55,170 --> 00:07:55,850
So like Mark was saying,
283
00:07:55,890 --> 00:07:56,830
this is kind of like uh,
284
00:07:57,530 --> 00:07:59,710
taking the end product,
285
00:07:59,770 --> 00:08:00,130
which is like,
286
00:08:00,390 --> 00:08:02,270
of construction or architecture,
287
00:08:02,330 --> 00:08:03,060
which is a house,
288
00:08:03,300 --> 00:08:04,340
and then trying to
289
00:08:04,740 --> 00:08:06,620
get the blueprints of how it
290
00:08:06,620 --> 00:08:07,100
was made.
291
00:08:07,100 --> 00:08:08,260
So if you look at it,
292
00:08:08,260 --> 00:08:09,500
look at a finished product of
293
00:08:09,500 --> 00:08:09,940
the house,
294
00:08:09,940 --> 00:08:11,180
you don't know how the wiring
295
00:08:11,180 --> 00:08:11,500
works,
296
00:08:11,500 --> 00:08:12,580
you don't know how the electrical
297
00:08:12,580 --> 00:08:13,720
plumb, uh,
298
00:08:13,840 --> 00:08:15,000
what kind of load bearing walls
299
00:08:15,000 --> 00:08:15,840
they used where.
300
00:08:16,160 --> 00:08:18,320
But you can do that with
301
00:08:18,320 --> 00:08:18,540
A.I.
302
00:08:18,540 --> 00:08:18,550
uh,
303
00:08:18,800 --> 00:08:20,080
in the world of software today
304
00:08:20,080 --> 00:08:22,120
by deconstructing an app and then
305
00:08:22,120 --> 00:08:23,360
rebuilding it from scratch.
306
00:08:23,680 --> 00:08:26,800
So this kind of sucks uh,
307
00:08:26,800 --> 00:08:29,550
for business owners, um,
308
00:08:30,160 --> 00:08:32,080
it's a win if you are
309
00:08:32,560 --> 00:08:34,040
at the whim of a predatory
310
00:08:34,040 --> 00:08:34,800
software company.
311
00:08:35,200 --> 00:08:37,920
But a lot of individual software
312
00:08:38,240 --> 00:08:40,240
builders who build apps
313
00:08:41,429 --> 00:08:43,509
as indie developers, publish it,
314
00:08:43,669 --> 00:08:44,949
charge some money for it,
315
00:08:45,269 --> 00:08:46,469
are out of luck because you
316
00:08:46,469 --> 00:08:48,069
can just copy their software.
317
00:08:48,309 --> 00:08:49,869
So my friend is a musician
318
00:08:49,869 --> 00:08:50,949
and he is building this
319
00:08:51,349 --> 00:08:53,869
synth software which uses a lot
320
00:08:53,869 --> 00:08:54,909
of mathematical functions,
321
00:08:54,909 --> 00:08:57,020
MIDI sounds to generate uh,
322
00:08:57,020 --> 00:08:58,629
music and package that into a
323
00:08:58,629 --> 00:09:00,109
nice tool and then hoping to
324
00:09:00,109 --> 00:09:00,629
sell it.
325
00:09:01,109 --> 00:09:02,629
Then I saw a couple other
326
00:09:02,629 --> 00:09:04,509
musicians or people who build these
327
00:09:04,509 --> 00:09:06,789
tools for musicians talk about how
328
00:09:06,789 --> 00:09:08,029
they don't really have an emote
329
00:09:08,029 --> 00:09:08,469
anymore.
330
00:09:09,040 --> 00:09:09,630
They uh,
331
00:09:10,160 --> 00:09:12,400
can find some of the best
332
00:09:12,720 --> 00:09:15,400
tools that are coveted by musicians
333
00:09:15,400 --> 00:09:17,760
that charge maybe like a hundred
334
00:09:17,760 --> 00:09:20,080
dollars for this specific synth sound.
335
00:09:20,240 --> 00:09:21,360
But then you download it,
336
00:09:21,360 --> 00:09:22,560
it's on your computer and you
337
00:09:22,560 --> 00:09:23,880
can just point Claude at it.
338
00:09:23,880 --> 00:09:25,480
You don't even need this reverse
339
00:09:25,480 --> 00:09:26,480
engineer anything tool.
340
00:09:26,480 --> 00:09:27,560
You can just point Claude at
341
00:09:27,560 --> 00:09:28,920
it and then it can reverse
342
00:09:28,920 --> 00:09:29,600
that entire
343
00:09:30,080 --> 00:09:32,240
complex musical software from scratch.
344
00:09:32,240 --> 00:09:34,080
And so where is
345
00:09:34,580 --> 00:09:35,860
the moat now anymore?
346
00:09:35,860 --> 00:09:38,580
How are indie developers protected?
347
00:09:38,580 --> 00:09:39,260
And it's,
348
00:09:39,260 --> 00:09:41,220
it's a really scary time if
349
00:09:41,220 --> 00:09:42,020
you're in that position.
350
00:09:42,870 --> 00:09:43,000
Um,
351
00:09:43,460 --> 00:09:45,020
and honestly I don't know where
352
00:09:45,020 --> 00:09:46,420
the world is headed right now.
353
00:09:48,100 --> 00:09:50,020
Yeah, it's, it's very interesting times.
354
00:09:50,410 --> 00:09:50,730
Uh,
355
00:09:50,730 --> 00:09:52,900
but I think that it's honestly
356
00:09:52,900 --> 00:09:54,580
fairly rare for
357
00:09:54,980 --> 00:09:57,540
companies to be based solely around
358
00:09:58,340 --> 00:10:00,100
software and software only.
359
00:10:00,180 --> 00:10:01,740
And I think Adobe is an
360
00:10:01,740 --> 00:10:02,660
example of,
361
00:10:02,740 --> 00:10:04,860
I'm talking about indie indie developers.
362
00:10:05,170 --> 00:10:05,780
Uh, Adobe,
363
00:10:05,780 --> 00:10:07,180
sure it's a mega corporation,
364
00:10:07,420 --> 00:10:08,420
publicly listed,
365
00:10:08,420 --> 00:10:10,300
but if you're an indie developer,
366
00:10:10,460 --> 00:10:11,650
like let's say uh,
367
00:10:11,650 --> 00:10:13,020
instead of an AI powered video
368
00:10:13,020 --> 00:10:13,340
editor,
369
00:10:13,340 --> 00:10:14,500
you were building something that was
370
00:10:14,500 --> 00:10:16,540
just downloadable as a binary.
371
00:10:18,540 --> 00:10:20,820
I'd be pretty concerned if that
372
00:10:20,820 --> 00:10:22,900
was the case, um, where,
373
00:10:22,900 --> 00:10:24,540
if it was only a pure
374
00:10:24,540 --> 00:10:25,740
binary because I think that there
375
00:10:25,740 --> 00:10:27,060
wouldn't really be much of a
376
00:10:27,060 --> 00:10:27,740
mode for that.
377
00:10:27,850 --> 00:10:28,620
Um,
378
00:10:28,940 --> 00:10:30,380
but I think that where
379
00:10:31,920 --> 00:10:33,640
something like Nova Cut really shines
380
00:10:33,640 --> 00:10:34,360
is not.
381
00:10:34,360 --> 00:10:35,480
I mean the software I think
382
00:10:35,480 --> 00:10:36,800
is pretty good and I spent
383
00:10:36,800 --> 00:10:37,760
a lot of time working on
384
00:10:37,760 --> 00:10:37,960
it.
385
00:10:37,960 --> 00:10:40,010
But really where um,
386
00:10:41,280 --> 00:10:43,480
people will get the value of
387
00:10:43,480 --> 00:10:45,270
paying for something is paying uh,
388
00:10:45,560 --> 00:10:46,320
for the compute.
389
00:10:46,560 --> 00:10:48,590
So I am um,
390
00:10:48,830 --> 00:10:50,080
uh Offering a lot of like
391
00:10:50,080 --> 00:10:52,240
AI compute for you on your
392
00:10:52,240 --> 00:10:52,720
behalf.
393
00:10:53,030 --> 00:10:53,180
Um,
394
00:10:53,360 --> 00:10:54,960
so you're paying for that plus
395
00:10:54,960 --> 00:10:57,200
the Nova Cut um, wrapper.
396
00:10:57,500 --> 00:10:58,000
Um,
397
00:10:58,130 --> 00:10:59,010
essentially where
398
00:10:59,570 --> 00:11:01,210
something like Nova Cut would be
399
00:11:01,210 --> 00:11:04,320
able to do the um,
400
00:11:05,570 --> 00:11:07,160
video uh, editing, uh,
401
00:11:07,160 --> 00:11:08,770
but then also you're getting access
402
00:11:08,850 --> 00:11:10,260
to uh,
403
00:11:10,530 --> 00:11:13,729
different LLMs and GPUs for web
404
00:11:13,729 --> 00:11:15,670
rendering and uh,
405
00:11:15,890 --> 00:11:18,120
audio cleanup and uh,
406
00:11:19,010 --> 00:11:21,170
sound generation and all kinds of
407
00:11:21,170 --> 00:11:22,840
different things where uh,
408
00:11:23,420 --> 00:11:24,260
if you wanted to do that
409
00:11:24,260 --> 00:11:26,060
on your own you could not
410
00:11:26,140 --> 00:11:27,880
because it'd be uh,
411
00:11:27,880 --> 00:11:28,900
too prohibitively expensive.
412
00:11:28,900 --> 00:11:29,820
And I just kind of wrap
413
00:11:29,820 --> 00:11:31,460
all the services up together and
414
00:11:31,460 --> 00:11:32,700
make it easy for you to
415
00:11:32,700 --> 00:11:34,620
just talk to an AI and
416
00:11:34,700 --> 00:11:36,140
get uh, your video edited.
417
00:11:36,300 --> 00:11:38,140
But if it was
418
00:11:38,540 --> 00:11:41,300
purely a um,
419
00:11:41,340 --> 00:11:44,140
service that would be a downloadable,
420
00:11:44,140 --> 00:11:45,660
I think that you'd be paying
421
00:11:45,660 --> 00:11:46,380
for a couple things.
422
00:11:46,600 --> 00:11:47,180
Um,
423
00:11:47,260 --> 00:11:47,660
so
424
00:11:49,020 --> 00:11:50,380
to make the argument for why
425
00:11:50,380 --> 00:11:50,860
I think
426
00:11:51,680 --> 00:11:53,880
software isn't or like SaaS isn't
427
00:11:53,880 --> 00:11:54,560
completely dead,
428
00:11:54,960 --> 00:11:55,760
I think that
429
00:11:56,160 --> 00:11:57,840
yes it's true that you can
430
00:11:57,840 --> 00:11:59,600
go and vibe code any piece
431
00:11:59,600 --> 00:12:02,400
of software. That's, that's true.
432
00:12:02,860 --> 00:12:05,460
But I think it's the um,
433
00:12:05,740 --> 00:12:05,800
the
434
00:12:06,340 --> 00:12:07,740
things that are kind of broken
435
00:12:07,820 --> 00:12:09,579
around the edges where
436
00:12:10,480 --> 00:12:11,580
the things will
437
00:12:11,960 --> 00:12:13,560
really shine with a bigger company.
438
00:12:13,680 --> 00:12:14,720
So let's say
439
00:12:15,220 --> 00:12:16,200
you vibe coded a piece of
440
00:12:16,240 --> 00:12:16,980
software which
441
00:12:17,540 --> 00:12:18,560
you may or may not understand
442
00:12:18,580 --> 00:12:19,140
how it works.
443
00:12:19,480 --> 00:12:20,800
If you're Vicodin you probably don't
444
00:12:20,860 --> 00:12:21,940
understand how it works and for
445
00:12:22,000 --> 00:12:23,060
you it's a black box.
446
00:12:23,520 --> 00:12:24,680
The moment something breaks,
447
00:12:24,880 --> 00:12:25,760
sure you can go have an
448
00:12:25,880 --> 00:12:27,260
LLM, try to fix it.
449
00:12:27,520 --> 00:12:28,460
But if,
450
00:12:28,820 --> 00:12:31,320
let's say Adobe or Salesforce or
451
00:12:31,380 --> 00:12:32,440
Microsoft or whatever,
452
00:12:32,540 --> 00:12:34,740
or even an indie developer, um,
453
00:12:34,880 --> 00:12:35,700
they would know how it works
454
00:12:35,740 --> 00:12:37,380
and they can keep the software
455
00:12:37,460 --> 00:12:37,940
updated,
456
00:12:38,480 --> 00:12:40,160
they'll be continually iterating on the
457
00:12:40,260 --> 00:12:42,020
UI and um,
458
00:12:43,020 --> 00:12:44,420
they'll be doing security updates for
459
00:12:44,500 --> 00:12:44,700
you.
460
00:12:45,200 --> 00:12:45,460
So
461
00:12:45,860 --> 00:12:46,840
yes it's true that you can
462
00:12:46,880 --> 00:12:48,220
go and copy the software and
463
00:12:48,260 --> 00:12:49,780
that's great.
464
00:12:50,340 --> 00:12:50,500
Um,
465
00:12:50,680 --> 00:12:51,740
but honestly
466
00:12:52,300 --> 00:12:53,640
most people I don't think want
467
00:12:53,720 --> 00:12:54,980
to have some sort of vibe
468
00:12:55,040 --> 00:12:55,440
coded
469
00:12:56,540 --> 00:12:58,360
janky piece of software that has
470
00:12:58,400 --> 00:13:00,040
no updates or security updates or
471
00:13:00,080 --> 00:13:01,460
customer support or anything like that.
472
00:13:01,520 --> 00:13:03,740
I think that for um,
473
00:13:03,780 --> 00:13:04,920
most people, uh,
474
00:13:05,040 --> 00:13:06,300
they would want to have a,
475
00:13:06,840 --> 00:13:08,360
they'd be okay paying a
476
00:13:09,640 --> 00:13:11,720
moderate premium um,
477
00:13:12,140 --> 00:13:15,260
for software that is like actually
478
00:13:15,320 --> 00:13:16,460
good and that doesn't have a
479
00:13:16,500 --> 00:13:18,040
lot of bugs around the edges.
480
00:13:18,780 --> 00:13:18,940
Um,
481
00:13:19,020 --> 00:13:19,760
and I don't think you're going
482
00:13:19,761 --> 00:13:20,420
to be able to get that
483
00:13:20,460 --> 00:13:22,060
with vive coded software at this
484
00:13:22,120 --> 00:13:22,480
point.
485
00:13:24,160 --> 00:13:24,280
You know,
486
00:13:24,360 --> 00:13:25,380
I have a counter take.
487
00:13:25,560 --> 00:13:25,880
I think
488
00:13:26,480 --> 00:13:27,940
open source is really going to
489
00:13:27,980 --> 00:13:28,400
take off
490
00:13:29,040 --> 00:13:30,640
if you're able to vibe code
491
00:13:30,680 --> 00:13:32,980
things just uh, in one shot.
492
00:13:34,020 --> 00:13:35,760
I Think the bar for vibe
493
00:13:35,820 --> 00:13:36,640
coding is going up
494
00:13:37,220 --> 00:13:38,400
and the way to stay on
495
00:13:38,460 --> 00:13:39,600
top of this is by open
496
00:13:39,680 --> 00:13:40,240
sourcing it,
497
00:13:40,720 --> 00:13:42,140
having the collective consciousness
498
00:13:43,100 --> 00:13:44,880
work on it together to build
499
00:13:45,320 --> 00:13:46,220
really good software.
500
00:13:46,840 --> 00:13:49,100
Um, and some of these uh,
501
00:13:49,520 --> 00:13:51,560
Microsoft Office clones or Photoshop clones
502
00:13:51,920 --> 00:13:53,120
have been open source and people
503
00:13:53,140 --> 00:13:54,440
have been contributing to it and
504
00:13:54,480 --> 00:13:55,480
it's getting really good.
505
00:13:56,060 --> 00:13:56,240
Um,
506
00:13:56,380 --> 00:13:58,080
according to their own claims it's
507
00:13:58,140 --> 00:13:58,780
somewhere between
508
00:13:59,340 --> 00:14:00,980
60 to 80% feature parity with
509
00:14:01,020 --> 00:14:01,880
Microsoft Office,
510
00:14:02,440 --> 00:14:03,440
a set of tools which is
511
00:14:03,480 --> 00:14:06,100
like Word, Excel, PowerPoint, um,
512
00:14:06,180 --> 00:14:07,160
so I have a couple other
513
00:14:07,220 --> 00:14:08,160
examples uh,
514
00:14:08,360 --> 00:14:09,840
if folks are interested to try
515
00:14:09,920 --> 00:14:10,120
out.
516
00:14:10,739 --> 00:14:11,220
Um,
517
00:14:12,140 --> 00:14:13,940
my TV has been really sluggish,
518
00:14:14,080 --> 00:14:15,000
the one I used to have
519
00:14:15,040 --> 00:14:16,040
before I started traveling.
520
00:14:16,420 --> 00:14:17,220
So if you want to clean
521
00:14:17,280 --> 00:14:19,040
up all the ads and
522
00:14:19,520 --> 00:14:21,380
unnecessary modules and jank on your
523
00:14:21,440 --> 00:14:21,780
TV,
524
00:14:22,240 --> 00:14:23,940
you can just point your AI
525
00:14:24,020 --> 00:14:24,960
at it, um,
526
00:14:25,300 --> 00:14:26,940
give it a picture of your
527
00:14:27,000 --> 00:14:29,480
TV's model and screenshot of uh,
528
00:14:29,820 --> 00:14:31,120
what it's showing you
529
00:14:31,600 --> 00:14:32,400
and ask it to just clean
530
00:14:32,420 --> 00:14:33,700
it up and it'll figure out
531
00:14:33,720 --> 00:14:34,240
how to do it.
532
00:14:35,160 --> 00:14:36,980
Uh, it's just that good today.
533
00:14:37,720 --> 00:14:38,000
Um,
534
00:14:38,420 --> 00:14:40,040
one other example that I found
535
00:14:40,120 --> 00:14:40,840
as a developer
536
00:14:41,300 --> 00:14:42,600
really cool was
537
00:14:43,160 --> 00:14:46,400
this channel called aoships A Y
538
00:14:46,680 --> 00:14:48,260
O S H I P S.
539
00:14:48,860 --> 00:14:48,900
Uh,
540
00:14:49,000 --> 00:14:50,720
on Instagram he talks about how
541
00:14:51,360 --> 00:14:53,280
he reverse engineers some of these
542
00:14:53,320 --> 00:14:55,060
popular apps on um,
543
00:14:55,260 --> 00:14:57,940
iOS Android like blockblast or subway
544
00:14:58,000 --> 00:14:59,180
surfers or Candy Crush.
545
00:14:59,820 --> 00:15:00,080
Um,
546
00:15:00,320 --> 00:15:01,160
and I'm just going to read
547
00:15:01,320 --> 00:15:03,200
one explanation of one of these
548
00:15:03,320 --> 00:15:05,380
app which is just mind blowing.
549
00:15:05,620 --> 00:15:06,680
And none of this is new.
550
00:15:07,200 --> 00:15:07,320
Uh,
551
00:15:07,420 --> 00:15:09,080
we've all known that these uh,
552
00:15:09,260 --> 00:15:11,780
mobile ad based games are pretty
553
00:15:11,820 --> 00:15:13,100
predatory but uh,
554
00:15:13,101 --> 00:15:14,680
it's just pretty cool to see
555
00:15:14,760 --> 00:15:14,780
a
556
00:15:15,260 --> 00:15:16,920
breakdown of the actual game because
557
00:15:16,980 --> 00:15:17,680
now we can see it.
558
00:15:17,920 --> 00:15:20,560
So blockblast, um,
559
00:15:20,840 --> 00:15:23,520
this one is player performance, uh,
560
00:15:23,680 --> 00:15:23,960
based
561
00:15:24,600 --> 00:15:26,380
and it's dictated less by skill
562
00:15:26,420 --> 00:15:28,120
and more by ad revenue potential.
563
00:15:28,620 --> 00:15:29,100
And then it
564
00:15:29,560 --> 00:15:31,340
calculates everything using uh,
565
00:15:31,540 --> 00:15:32,720
your phone's model
566
00:15:33,180 --> 00:15:35,320
country ad view history and even
567
00:15:35,360 --> 00:15:37,040
whether the phone's vibration is uh,
568
00:15:37,240 --> 00:15:37,780
on or off.
569
00:15:38,400 --> 00:15:39,120
So it's
570
00:15:39,700 --> 00:15:40,960
the game secretly makes
571
00:15:42,100 --> 00:15:42,860
gameplay easier
572
00:15:43,360 --> 00:15:44,660
so that players feel like they're
573
00:15:44,700 --> 00:15:45,340
on a high run
574
00:15:46,000 --> 00:15:47,060
right up until they reach their
575
00:15:47,160 --> 00:15:47,960
average game length.
576
00:15:48,640 --> 00:15:50,000
And then they trigger this thing
577
00:15:50,040 --> 00:15:51,340
called the death algorithm.
578
00:15:51,820 --> 00:15:53,080
When an expensive ad is queued
579
00:15:53,160 --> 00:15:54,520
up for a user with a
580
00:15:54,560 --> 00:15:55,620
history of watching ads,
581
00:15:56,180 --> 00:15:57,880
the game helps them beat their
582
00:15:57,920 --> 00:15:58,500
high score
583
00:15:59,220 --> 00:16:01,040
and then triggers a death algorithm
584
00:16:01,100 --> 00:16:02,620
that deals an unwinnable turn.
585
00:16:02,920 --> 00:16:04,640
And then immediately upon losing,
586
00:16:04,640 --> 00:16:05,800
the player is offered
587
00:16:06,200 --> 00:16:08,240
a revive in exchange for watching
588
00:16:08,240 --> 00:16:09,160
that high value ad.
589
00:16:09,480 --> 00:16:11,440
And so accepting that revive secretly
590
00:16:11,440 --> 00:16:13,080
adds 10% to their high score
591
00:16:13,080 --> 00:16:14,120
to keep them hooked.
592
00:16:16,120 --> 00:16:17,960
This is so cool the way
593
00:16:17,960 --> 00:16:18,760
we were able to
594
00:16:19,160 --> 00:16:21,640
see exactly how these companies are
595
00:16:22,090 --> 00:16:22,890
being so predatory.
596
00:16:23,290 --> 00:16:24,330
And so when you feel like
597
00:16:24,330 --> 00:16:24,730
you're.
598
00:16:24,890 --> 00:16:26,250
You're on a roll and you're
599
00:16:26,250 --> 00:16:27,290
killing it on this game,
600
00:16:27,370 --> 00:16:28,490
that's not actually true.
601
00:16:28,490 --> 00:16:30,770
It just means that you're really
602
00:16:30,770 --> 00:16:32,850
valuable to the ad company behind
603
00:16:32,850 --> 00:16:37,370
the scenes. That is insane.
604
00:16:37,930 --> 00:16:41,370
What company was that? Uh, Blockblast.
605
00:16:41,930 --> 00:16:44,810
Block Blast. Okay. Wow.
606
00:16:44,970 --> 00:16:46,010
So that
607
00:16:46,890 --> 00:16:48,010
made me lose a little bit
608
00:16:48,010 --> 00:16:48,770
of faith in humanity.
609
00:16:48,770 --> 00:16:51,240
That's unfortunate that they've been doing
610
00:16:51,240 --> 00:16:51,920
this for a while.
611
00:16:52,400 --> 00:16:55,600
I mean, I'm sure. But still.
612
00:16:55,800 --> 00:16:56,010
Um.
613
00:16:56,010 --> 00:16:56,320
Yeah,
614
00:16:56,320 --> 00:16:57,560
but also another thing that's kind
615
00:16:57,560 --> 00:16:58,400
of cool about a lot of
616
00:16:58,400 --> 00:17:00,360
these reverse engineering is people were.
617
00:17:00,360 --> 00:17:01,960
Are able to mash up video
618
00:17:01,960 --> 00:17:02,320
games.
619
00:17:03,200 --> 00:17:03,600
So
620
00:17:04,240 --> 00:17:05,520
I saw that people were
621
00:17:06,000 --> 00:17:07,600
doing stuff like putting Super
622
00:17:08,000 --> 00:17:11,600
Mario into Red Dead Redemption or,
623
00:17:12,180 --> 00:17:12,670
um,
624
00:17:12,670 --> 00:17:15,200
going in and putting Spider man
625
00:17:15,520 --> 00:17:17,010
in, like,
626
00:17:17,010 --> 00:17:19,010
Mario Kart and all kinds of,
627
00:17:19,140 --> 00:17:20,290
uh, fun stuff like that.
628
00:17:20,450 --> 00:17:21,370
So, like, I mean,
629
00:17:21,370 --> 00:17:23,010
I think that's. That's pretty sweet.
630
00:17:23,170 --> 00:17:23,930
And also,
631
00:17:23,930 --> 00:17:25,690
I saw that people were going
632
00:17:25,690 --> 00:17:28,290
and putting old video games
633
00:17:28,770 --> 00:17:30,210
completely in the browser.
634
00:17:30,530 --> 00:17:32,050
So I think there was one
635
00:17:32,449 --> 00:17:33,100
with, uh,
636
00:17:33,100 --> 00:17:33,650
Modern
637
00:17:34,050 --> 00:17:36,450
Warfare 2 or Modern Warfare 3,
638
00:17:36,450 --> 00:17:37,170
something like that.
639
00:17:37,640 --> 00:17:37,820
Um,
640
00:17:37,890 --> 00:17:39,250
it was like Cloud of War.
641
00:17:39,860 --> 00:17:39,910
Uh,
642
00:17:40,130 --> 00:17:41,570
and then you could play that
643
00:17:41,730 --> 00:17:43,620
completely in the browser.
644
00:17:43,700 --> 00:17:44,500
Uh, because that used to be,
645
00:17:44,500 --> 00:17:46,740
I think, a PlayStation 2 game,
646
00:17:46,740 --> 00:17:48,180
PlayStation 3, something like that,
647
00:17:48,260 --> 00:17:49,140
and now you can just do
648
00:17:49,140 --> 00:17:49,820
it in your browser.
649
00:17:49,820 --> 00:17:51,860
I also saw Grand Theft Auto
650
00:17:52,660 --> 00:17:55,630
completely in the browser with, uh.
651
00:17:56,180 --> 00:17:56,980
Yeah, I mean,
652
00:17:56,980 --> 00:17:59,220
that's amazing where you could just
653
00:17:59,300 --> 00:18:00,300
do all these things.
654
00:18:00,300 --> 00:18:01,660
Not on a game console.
655
00:18:01,660 --> 00:18:02,580
But, I mean,
656
00:18:02,580 --> 00:18:04,820
Chrome is pretty darn powerful nowadays
657
00:18:04,820 --> 00:18:06,180
that you can just run full
658
00:18:06,180 --> 00:18:08,100
AAA games completely in the process.
659
00:18:08,430 --> 00:18:10,590
Absolutely ridiculous. Yeah.
660
00:18:10,590 --> 00:18:11,470
I think in the last episode
661
00:18:11,470 --> 00:18:12,830
we talked about this explosion of
662
00:18:12,830 --> 00:18:14,750
people making proof of concept.
663
00:18:14,910 --> 00:18:17,710
Um, blender, uh, 3D environments,
664
00:18:17,950 --> 00:18:20,030
or just toy games.
665
00:18:20,030 --> 00:18:21,210
Mark made a game and then,
666
00:18:21,210 --> 00:18:23,350
uh, showed it, I think, last,
667
00:18:23,350 --> 00:18:23,950
uh, video.
668
00:18:24,109 --> 00:18:24,950
Oh, I don't think I.
669
00:18:24,950 --> 00:18:25,430
Actually, we.
670
00:18:25,430 --> 00:18:26,190
I think I showed it to
671
00:18:26,190 --> 00:18:26,310
you,
672
00:18:26,310 --> 00:18:27,070
but I don't think I showed
673
00:18:27,070 --> 00:18:28,430
it to everybody.
674
00:18:28,690 --> 00:18:30,470
Um, but we can.
675
00:18:30,470 --> 00:18:31,150
We can pull it up in
676
00:18:31,150 --> 00:18:31,830
the meantime, if you want.
677
00:18:31,830 --> 00:18:32,950
Yeah, let me. Give me.
678
00:18:32,950 --> 00:18:33,670
Give me two minutes.
679
00:18:33,670 --> 00:18:35,600
I'll pull it up. You can.
680
00:18:36,000 --> 00:18:37,120
For a moment. Shashank.
681
00:18:37,950 --> 00:18:39,390
Uh, I think, uh,
682
00:18:39,390 --> 00:18:43,730
this whole wave of capability, uh,
683
00:18:44,240 --> 00:18:45,520
has created this whole wave of
684
00:18:45,520 --> 00:18:45,840
new
685
00:18:46,240 --> 00:18:48,000
apps and video games that are
686
00:18:48,000 --> 00:18:50,400
being created. And three js.
687
00:18:50,880 --> 00:18:51,280
The
688
00:18:51,840 --> 00:18:54,960
number of downloads has freaking skyrocketed.
689
00:18:55,280 --> 00:18:55,960
Someone was like,
690
00:18:55,960 --> 00:18:57,040
something's wrong with this graph.
691
00:18:57,040 --> 00:18:58,160
Usually it goes up into the
692
00:18:58,160 --> 00:18:58,360
right.
693
00:18:58,360 --> 00:18:59,920
Not just freaking up like a
694
00:18:59,920 --> 00:19:00,560
straight line,
695
00:19:01,090 --> 00:19:01,330
because
696
00:19:01,970 --> 00:19:04,210
the number of agents building games
697
00:19:04,210 --> 00:19:04,610
and
698
00:19:05,090 --> 00:19:07,170
creating proof of concept things in
699
00:19:07,170 --> 00:19:07,810
3js,
700
00:19:07,810 --> 00:19:09,170
which is like a browser based
701
00:19:10,740 --> 00:19:12,320
uh, game, uh,
702
00:19:12,320 --> 00:19:14,890
3d environment rendering tool has just
703
00:19:14,890 --> 00:19:15,730
exploded.
704
00:19:16,769 --> 00:19:17,910
Mark, uh,
705
00:19:18,130 --> 00:19:20,210
for audio listeners is showing
706
00:19:20,610 --> 00:19:22,810
a video of a game that
707
00:19:22,810 --> 00:19:24,290
he just vibe coded.
708
00:19:25,890 --> 00:19:30,490
Yeah, this game is pretty fun,
709
00:19:30,570 --> 00:19:31,130
I think.
710
00:19:31,290 --> 00:19:32,530
So I wanted to see like
711
00:19:32,530 --> 00:19:34,170
hey, with Opus 5.5,
712
00:19:34,330 --> 00:19:36,570
what can I build procedurally?
713
00:19:36,970 --> 00:19:38,770
And I decided that I was
714
00:19:38,770 --> 00:19:40,810
going to make a high quality
715
00:19:41,690 --> 00:19:46,590
3D generated, completely procedurally generated, ah,
716
00:19:46,590 --> 00:19:48,410
game where you can just explore
717
00:19:48,490 --> 00:19:49,130
a town.
718
00:19:49,770 --> 00:19:50,860
So um,
719
00:19:51,130 --> 00:19:52,250
I'll put the link in the
720
00:19:52,250 --> 00:19:52,770
show notes.
721
00:19:52,770 --> 00:19:54,050
So if you want you can,
722
00:19:54,050 --> 00:19:54,890
you can play it.
723
00:19:55,390 --> 00:19:57,070
It's actually weirdly fun.
724
00:19:57,510 --> 00:20:00,250
Um, but you are given um,
725
00:20:00,250 --> 00:20:01,590
like a sword and it's got
726
00:20:01,590 --> 00:20:03,170
like oh, the uh,
727
00:20:03,950 --> 00:20:04,750
you can have,
728
00:20:05,230 --> 00:20:07,110
you can go and fly and
729
00:20:07,110 --> 00:20:08,670
you can go to different towns
730
00:20:08,750 --> 00:20:10,330
and the towns are like uh,
731
00:20:10,330 --> 00:20:11,750
kind of big, um,
732
00:20:11,870 --> 00:20:13,070
of different sizes.
733
00:20:13,070 --> 00:20:14,150
They'll often have like a church
734
00:20:14,150 --> 00:20:14,750
in the middle.
735
00:20:14,750 --> 00:20:16,750
You can have um, a bounty.
736
00:20:16,750 --> 00:20:17,630
So if you go and kill
737
00:20:17,630 --> 00:20:19,250
people in the town, um,
738
00:20:19,250 --> 00:20:20,980
they'll remember you and then come
739
00:20:20,980 --> 00:20:22,220
and try to hunt you down.
740
00:20:22,460 --> 00:20:24,340
Every single house, uh,
741
00:20:24,340 --> 00:20:25,820
has a different layout.
742
00:20:26,380 --> 00:20:27,620
So you can go and then
743
00:20:27,620 --> 00:20:29,720
maybe rob the house or um,
744
00:20:30,760 --> 00:20:34,620
uh, go explore to different towns.
745
00:20:34,780 --> 00:20:34,920
Uh,
746
00:20:34,920 --> 00:20:36,220
there's a bunch of different like
747
00:20:36,220 --> 00:20:38,060
random events you can go and
748
00:20:38,060 --> 00:20:38,380
see.
749
00:20:38,840 --> 00:20:40,540
Uh, you can go climb mountains.
750
00:20:41,340 --> 00:20:41,980
I don't know.
751
00:20:41,980 --> 00:20:42,940
It's kind of like.
752
00:20:42,940 --> 00:20:45,620
What's that procedurally generated space game?
753
00:20:45,620 --> 00:20:46,590
No Man's Space Sky.
754
00:20:46,590 --> 00:20:47,430
It's kind of like that
755
00:20:47,830 --> 00:20:49,990
but within just like a single
756
00:20:49,990 --> 00:20:52,230
world where there's like towns and
757
00:20:52,470 --> 00:20:53,750
AI generated music.
758
00:20:54,240 --> 00:20:54,740
Uh,
759
00:20:54,740 --> 00:20:56,150
and like the towns have like
760
00:20:56,150 --> 00:20:57,030
many, many cities.
761
00:20:57,660 --> 00:20:58,950
Uh, anyways, it's, it's a,
762
00:20:58,950 --> 00:21:00,670
it's a fun game that I
763
00:21:00,670 --> 00:21:01,830
made in a couple hours.
764
00:21:02,710 --> 00:21:04,790
So if you are interested,
765
00:21:05,190 --> 00:21:07,080
I would be, uh,
766
00:21:07,430 --> 00:21:08,430
I would love to get your
767
00:21:08,430 --> 00:21:09,270
feedback on it.
768
00:21:10,310 --> 00:21:11,110
But anyways,
769
00:21:11,190 --> 00:21:12,390
we'll stop there because
770
00:21:12,880 --> 00:21:13,880
I could talk for a while
771
00:21:13,880 --> 00:21:14,280
about it.
772
00:21:14,280 --> 00:21:17,280
But it's a fun, ish game,
773
00:21:17,280 --> 00:21:18,000
I would say.
774
00:21:19,040 --> 00:21:20,240
All the audio playing in the
775
00:21:20,240 --> 00:21:20,900
background, I uh,
776
00:21:21,000 --> 00:21:22,200
think the listeners could hear it
777
00:21:22,200 --> 00:21:23,920
was all procedurally generated by.
778
00:21:23,920 --> 00:21:25,120
I just mean Claude.
779
00:21:25,280 --> 00:21:26,640
And it was able to create
780
00:21:26,800 --> 00:21:29,040
MIDI sounds for sound effects,
781
00:21:29,040 --> 00:21:30,480
background music, uh,
782
00:21:30,560 --> 00:21:33,160
and different game uh, events too,
783
00:21:33,160 --> 00:21:37,680
which is amazing. Yeah, um, awesome.
784
00:21:37,780 --> 00:21:37,840
Um,
785
00:21:38,160 --> 00:21:39,120
shall we move on to
786
00:21:40,090 --> 00:21:41,210
model releases?
787
00:21:41,450 --> 00:21:42,650
Yeah, let's talk about those.
788
00:21:43,930 --> 00:21:45,530
Muse has been on a roll
789
00:21:45,530 --> 00:21:46,010
lately.
790
00:21:46,010 --> 00:21:46,750
I think Meta, uh,
791
00:21:47,090 --> 00:21:48,170
has really turned around
792
00:21:48,650 --> 00:21:50,370
and has been doing an amazing
793
00:21:50,370 --> 00:21:52,890
job with Alexander Wang.
794
00:21:53,070 --> 00:21:53,770
Uh, you know,
795
00:21:53,770 --> 00:21:55,649
I thought that after they fired
796
00:21:55,649 --> 00:21:58,060
their entire Llama team, um,
797
00:21:58,060 --> 00:21:58,890
Yann Lecun,
798
00:21:58,890 --> 00:22:00,330
one of the godfathers of AI,
799
00:22:00,750 --> 00:22:02,890
uh, and they put this,
800
00:22:02,890 --> 00:22:03,450
you know,
801
00:22:04,090 --> 00:22:05,850
young kid at the head of
802
00:22:06,590 --> 00:22:08,270
Meta Superintelligence Labs,
803
00:22:08,510 --> 00:22:11,010
and anecdotally from friends, uh,
804
00:22:11,310 --> 00:22:12,030
working there,
805
00:22:12,190 --> 00:22:13,230
I heard that it was
806
00:22:13,790 --> 00:22:14,950
kind of very chaotic,
807
00:22:14,950 --> 00:22:16,230
but it seems like it's paid
808
00:22:16,230 --> 00:22:16,510
off.
809
00:22:17,070 --> 00:22:19,870
Their move fast, break things strategy,
810
00:22:20,370 --> 00:22:20,770
um,
811
00:22:20,770 --> 00:22:22,430
is kind of what we need
812
00:22:22,430 --> 00:22:23,270
right now, I think,
813
00:22:23,270 --> 00:22:25,550
because the traditional AI labs,
814
00:22:25,550 --> 00:22:26,590
even at Google,
815
00:22:26,830 --> 00:22:28,230
they took a long time and
816
00:22:28,230 --> 00:22:31,230
were really slow to really
817
00:22:32,360 --> 00:22:34,400
take action because we don't need
818
00:22:34,400 --> 00:22:35,960
that much research right now.
819
00:22:36,360 --> 00:22:38,200
We need action, we need compute,
820
00:22:38,360 --> 00:22:39,800
we need lots and lots of
821
00:22:39,800 --> 00:22:40,520
training runs,
822
00:22:40,520 --> 00:22:43,160
lots of experiments and just build
823
00:22:43,160 --> 00:22:45,240
something which Meta has done.
824
00:22:45,400 --> 00:22:47,480
And now they're number two
825
00:22:48,040 --> 00:22:50,120
after just being so far behind.
826
00:22:50,440 --> 00:22:51,600
They come out of nowhere.
827
00:22:51,600 --> 00:22:51,880
And
828
00:22:52,520 --> 00:22:53,920
this is not a fully released
829
00:22:53,920 --> 00:22:57,240
model, but their new muse, Wonderful.
830
00:22:57,930 --> 00:22:59,280
Um, still private,
831
00:22:59,600 --> 00:23:02,080
but is reportedly right behind
832
00:23:02,480 --> 00:23:03,760
Opus 5.5.
833
00:23:03,760 --> 00:23:06,160
And that is remarkable to see.
834
00:23:07,360 --> 00:23:07,840
Yeah,
835
00:23:08,720 --> 00:23:11,040
I know that Meta was behind
836
00:23:11,120 --> 00:23:12,120
for a long time,
837
00:23:12,120 --> 00:23:14,320
but it just warms my heart
838
00:23:14,320 --> 00:23:14,960
to see
839
00:23:15,520 --> 00:23:18,080
both Meta and Google kind of,
840
00:23:18,360 --> 00:23:18,450
uh,
841
00:23:18,560 --> 00:23:20,640
getting closer to the top of
842
00:23:20,640 --> 00:23:23,530
the leaderboards and yeah,
843
00:23:23,530 --> 00:23:24,850
I mean that's amazing.
844
00:23:24,850 --> 00:23:25,250
So, I mean,
845
00:23:25,250 --> 00:23:26,130
I think that you could make
846
00:23:26,130 --> 00:23:29,370
the argument that both, uh, anthropic,
847
00:23:29,770 --> 00:23:34,730
OpenAI, Google and Meta all have,
848
00:23:35,180 --> 00:23:35,420
uh,
849
00:23:35,770 --> 00:23:38,650
models that are within spitting distance
850
00:23:38,650 --> 00:23:39,770
of each other.
851
00:23:39,850 --> 00:23:41,890
I think that Anthropic is still
852
00:23:41,890 --> 00:23:43,290
kind of the top dog
853
00:23:43,690 --> 00:23:45,770
right now in terms of models,
854
00:23:46,170 --> 00:23:46,570
but
855
00:23:47,130 --> 00:23:49,730
Google and Meta are very close
856
00:23:49,730 --> 00:23:50,090
behind.
857
00:23:50,170 --> 00:23:51,570
And honestly for most people I
858
00:23:51,570 --> 00:23:52,810
don't think it really even matters.
859
00:23:52,810 --> 00:23:54,030
I mean, um,
860
00:23:54,030 --> 00:23:55,110
if you're just trying to have
861
00:23:55,110 --> 00:23:57,030
it summarize a website for you
862
00:23:57,030 --> 00:23:58,030
or you know,
863
00:23:58,030 --> 00:23:59,510
organize your email or do some
864
00:23:59,510 --> 00:24:01,080
travel booking, whatever, um,
865
00:24:01,150 --> 00:24:02,970
all these models are plenty good
866
00:24:03,030 --> 00:24:03,330
enough.
867
00:24:03,390 --> 00:24:04,450
It's only the people who are
868
00:24:04,490 --> 00:24:05,390
really stressing it,
869
00:24:05,790 --> 00:24:07,910
the models would even really notice
870
00:24:07,970 --> 00:24:10,030
this. But yes, good job Meta.
871
00:24:10,070 --> 00:24:12,390
And I think that, uh,
872
00:24:12,790 --> 00:24:14,070
it really shows that a lot
873
00:24:14,130 --> 00:24:15,910
of the money that Meta has
874
00:24:15,950 --> 00:24:18,630
been putting into hiring all of
875
00:24:18,670 --> 00:24:20,430
these people and spending money on
876
00:24:20,470 --> 00:24:23,130
compute, and has really been, um,
877
00:24:23,430 --> 00:24:24,170
working out.
878
00:24:24,350 --> 00:24:25,030
And Shashank,
879
00:24:25,190 --> 00:24:26,230
I agree with you that
880
00:24:26,770 --> 00:24:28,950
right now the
881
00:24:30,070 --> 00:24:31,770
models are built in such a
882
00:24:31,810 --> 00:24:32,850
way where yes,
883
00:24:33,090 --> 00:24:35,710
it would be nice to increase,
884
00:24:36,050 --> 00:24:38,990
uh, capability and improve on algorithms.
885
00:24:39,030 --> 00:24:39,990
And I think that there is
886
00:24:40,090 --> 00:24:42,590
a lot of room for improvement.
887
00:24:43,290 --> 00:24:44,870
But there already is a winning
888
00:24:44,930 --> 00:24:47,190
formula for AI and it's basically
889
00:24:47,230 --> 00:24:47,610
just like,
890
00:24:48,130 --> 00:24:49,790
get as much data as possible,
891
00:24:50,410 --> 00:24:52,170
get as much compute as possible,
892
00:24:53,210 --> 00:24:55,030
leverage the current state of the
893
00:24:55,130 --> 00:24:58,090
art, um, models for AI,
894
00:24:58,170 --> 00:24:59,750
which are in general,
895
00:24:59,790 --> 00:25:01,170
a lot of that stuff is
896
00:25:01,270 --> 00:25:02,590
open research and open papers.
897
00:25:02,650 --> 00:25:05,130
So people kind of know what
898
00:25:05,210 --> 00:25:06,489
that architecture looks like
899
00:25:07,230 --> 00:25:08,770
and they just throw a bunch
900
00:25:08,790 --> 00:25:10,090
of compute at it and have
901
00:25:10,130 --> 00:25:12,630
the AI help research themselves.
902
00:25:12,830 --> 00:25:14,130
And that seems to be a
903
00:25:14,170 --> 00:25:16,250
winning formula where it basically boils
904
00:25:16,290 --> 00:25:17,870
down to how much money you
905
00:25:17,950 --> 00:25:18,410
have
906
00:25:18,950 --> 00:25:21,090
can immediately kind of turn into
907
00:25:21,710 --> 00:25:22,690
how good of a model you're
908
00:25:22,750 --> 00:25:23,250
able to make.
909
00:25:23,310 --> 00:25:25,810
Now, not completely because um,
910
00:25:26,190 --> 00:25:27,070
Elon Musk has
911
00:25:27,690 --> 00:25:28,930
infinite money and
912
00:25:29,430 --> 00:25:32,310
Grok is still slightly behind.
913
00:25:32,650 --> 00:25:33,070
But
914
00:25:33,770 --> 00:25:36,050
even again I would argue that
915
00:25:36,130 --> 00:25:39,190
Grok is in within spinning distance
916
00:25:39,510 --> 00:25:39,730
of
917
00:25:40,390 --> 00:25:41,410
any of these top state of
918
00:25:41,450 --> 00:25:42,650
the art models and I would
919
00:25:42,690 --> 00:25:45,010
not be surprised if like next
920
00:25:45,050 --> 00:25:45,850
week or something,
921
00:25:46,030 --> 00:25:48,390
Grok made one that was also
922
00:25:48,450 --> 00:25:49,050
state of the art.
923
00:25:49,130 --> 00:25:49,850
So um,
924
00:25:50,270 --> 00:25:52,270
it's not something that small startups
925
00:25:52,310 --> 00:25:53,250
can really compete at at this
926
00:25:53,310 --> 00:25:53,610
point.
927
00:25:54,090 --> 00:25:54,250
Um,
928
00:25:54,410 --> 00:25:55,590
but if you've got a lot
929
00:25:55,630 --> 00:25:56,010
of money
930
00:25:56,610 --> 00:25:58,190
and you have smart people,
931
00:25:59,030 --> 00:25:59,131
you uh,
932
00:25:59,150 --> 00:26:00,790
can go and create a
933
00:26:01,730 --> 00:26:02,810
state of the art or near
934
00:26:02,850 --> 00:26:03,770
state of the art model.
935
00:26:03,990 --> 00:26:05,450
And especially if you're trying to
936
00:26:05,530 --> 00:26:07,230
lean in towards individual tasks,
937
00:26:07,310 --> 00:26:08,510
you can definitely create something that's
938
00:26:08,511 --> 00:26:08,970
state of the art.
939
00:26:09,010 --> 00:26:09,770
If you say I wanted to
940
00:26:09,850 --> 00:26:12,410
specialize in ah, one particular thing,
941
00:26:12,550 --> 00:26:14,350
so good job.
942
00:26:14,830 --> 00:26:16,390
Mark Zuckerberg and Facebook,
943
00:26:16,650 --> 00:26:18,510
I think they lost some goodwill
944
00:26:18,570 --> 00:26:18,810
but
945
00:26:19,510 --> 00:26:21,330
they've been gaining a lot of
946
00:26:21,410 --> 00:26:22,970
it back at least
947
00:26:23,890 --> 00:26:26,010
going back and creating some open
948
00:26:26,070 --> 00:26:27,330
models and uh,
949
00:26:27,870 --> 00:26:28,810
giving away a lot of tokens
950
00:26:28,830 --> 00:26:29,330
for free.
951
00:26:29,390 --> 00:26:30,110
And yes,
952
00:26:30,370 --> 00:26:31,950
I think it's true that um,
953
00:26:32,330 --> 00:26:32,970
they're maybe
954
00:26:33,630 --> 00:26:35,310
a little looser with some of
955
00:26:35,370 --> 00:26:38,210
the data sharing things, um,
956
00:26:38,790 --> 00:26:41,270
to put it mildly where um,
957
00:26:41,670 --> 00:26:43,170
yes you're getting 100 million tokens
958
00:26:43,190 --> 00:26:43,530
for free,
959
00:26:43,590 --> 00:26:45,310
but they are probably using all
960
00:26:45,350 --> 00:26:46,550
your conversations to train.
961
00:26:47,190 --> 00:26:48,710
However, we don't know.
962
00:26:48,750 --> 00:26:50,230
That's all alleged.
963
00:26:50,690 --> 00:26:51,390
Maybe not alleged,
964
00:26:51,450 --> 00:26:52,990
I don't know but um,
965
00:26:53,910 --> 00:26:54,610
I'm not going to say
966
00:26:55,090 --> 00:26:55,890
use it, don't use it,
967
00:26:56,390 --> 00:26:58,630
but just be aware that
968
00:26:59,230 --> 00:27:01,270
some of your conversations may be
969
00:27:01,350 --> 00:27:01,990
used to train.
970
00:27:02,130 --> 00:27:03,731
However they ah,
971
00:27:03,770 --> 00:27:04,490
are giving a lot away,
972
00:27:04,530 --> 00:27:05,570
a lot of free usage and
973
00:27:05,610 --> 00:27:07,490
they are incredibly good models.
974
00:27:07,830 --> 00:27:09,070
So exciting stuff.
975
00:27:10,570 --> 00:27:10,950
Yeah, they're,
976
00:27:11,250 --> 00:27:11,550
they're
977
00:27:11,910 --> 00:27:13,950
giving away free usage and
978
00:27:14,350 --> 00:27:16,130
you can use their Meta Muse
979
00:27:16,450 --> 00:27:16,810
Spark.
980
00:27:17,130 --> 00:27:17,450
Uh, sorry,
981
00:27:17,830 --> 00:27:21,030
your Meta Muse Always On Agent,
982
00:27:21,290 --> 00:27:22,610
the little um,
983
00:27:22,930 --> 00:27:24,710
plush toy looking uh,
984
00:27:24,850 --> 00:27:27,210
teddy bear that is um,
985
00:27:27,470 --> 00:27:28,650
deceptively cute.
986
00:27:28,770 --> 00:27:28,990
Uh,
987
00:27:29,150 --> 00:27:29,470
but is
988
00:27:30,130 --> 00:27:31,330
may or may not be secretly
989
00:27:31,650 --> 00:27:32,570
stealing all of your data.
990
00:27:33,190 --> 00:27:33,410
Um,
991
00:27:33,590 --> 00:27:35,110
which is probably one of the
992
00:27:35,190 --> 00:27:36,950
best kinds of use cases for
993
00:27:36,990 --> 00:27:37,810
the average person.
994
00:27:38,410 --> 00:27:38,610
Um,
995
00:27:38,790 --> 00:27:39,970
there's a couple different tools out
996
00:27:40,010 --> 00:27:40,250
there.
997
00:27:40,410 --> 00:27:41,810
The Always on agents have just
998
00:27:41,870 --> 00:27:43,110
been um,
999
00:27:43,770 --> 00:27:45,890
on a roll lately with meta's,
1000
00:27:45,930 --> 00:27:47,410
uh, muse, OpenAI's dots,
1001
00:27:47,770 --> 00:27:48,430
grokbots
1002
00:27:48,890 --> 00:27:50,650
and Instinct and so all of
1003
00:27:50,710 --> 00:27:51,830
those have been coming out
1004
00:27:52,350 --> 00:27:53,190
and Meta has done a good
1005
00:27:53,250 --> 00:27:53,530
job.
1006
00:27:54,330 --> 00:27:54,730
Um,
1007
00:27:54,990 --> 00:27:56,890
Meta has also open sourced one
1008
00:27:56,930 --> 00:27:57,990
of their smaller models.
1009
00:27:58,650 --> 00:27:58,870
Um,
1010
00:27:58,950 --> 00:27:59,950
but to be honest I feel
1011
00:27:59,970 --> 00:28:01,470
like the open source world
1012
00:28:02,130 --> 00:28:04,580
is not that exciting.
1013
00:28:04,620 --> 00:28:05,280
In the US
1014
00:28:05,640 --> 00:28:07,940
their Spark model scores like
1015
00:28:09,940 --> 00:28:12,580
30M9 or something on artificial analysis
1016
00:28:12,660 --> 00:28:13,720
where the
1017
00:28:14,780 --> 00:28:17,120
frontier is at 58 or something.
1018
00:28:17,960 --> 00:28:20,520
So the open source
1019
00:28:21,100 --> 00:28:24,600
is really being pushed forward by
1020
00:28:24,680 --> 00:28:25,780
Chinese AI, lots of.
1021
00:28:26,560 --> 00:28:28,080
Um, but hey,
1022
00:28:28,380 --> 00:28:29,600
I think a new open source
1023
00:28:29,640 --> 00:28:30,620
model is not bad.
1024
00:28:30,980 --> 00:28:32,320
People are going to learn from
1025
00:28:32,960 --> 00:28:33,640
how it works,
1026
00:28:33,820 --> 00:28:35,300
what strategies they took for training.
1027
00:28:35,880 --> 00:28:36,100
Um,
1028
00:28:36,180 --> 00:28:37,580
they've released their open weights too
1029
00:28:37,780 --> 00:28:38,740
so that's a,
1030
00:28:39,180 --> 00:28:40,480
that's a plus for the community.
1031
00:28:41,320 --> 00:28:44,080
Um, yeah, it's great.
1032
00:28:44,140 --> 00:28:45,720
And also I think we should
1033
00:28:45,780 --> 00:28:47,920
mention that there are a few
1034
00:28:47,980 --> 00:28:50,560
different Chinese model uh,
1035
00:28:50,620 --> 00:28:53,040
releases that have come out recently.
1036
00:28:53,320 --> 00:28:54,500
So um,
1037
00:28:54,900 --> 00:28:56,020
as of a couple days ago,
1038
00:28:56,320 --> 00:29:00,680
Deepseek v4.1 flash 1017.
1039
00:29:00,800 --> 00:29:02,220
So I know it's still Deep
1040
00:29:02,280 --> 00:29:04,420
Sea v4.1, um,
1041
00:29:04,640 --> 00:29:08,120
but it's the October 17th version
1042
00:29:08,160 --> 00:29:10,420
which actually has some pretty uh,
1043
00:29:10,520 --> 00:29:13,660
substantial gains from the model that
1044
00:29:13,720 --> 00:29:15,960
I think was made in September
1045
00:29:16,000 --> 00:29:16,540
or something.
1046
00:29:17,140 --> 00:29:17,440
Um,
1047
00:29:18,140 --> 00:29:19,980
it's awesome that they're releasing monthly
1048
00:29:20,020 --> 00:29:22,520
at this point. Yeah, yeah.
1049
00:29:22,560 --> 00:29:23,980
So it seems like just fine
1050
00:29:24,060 --> 00:29:25,880
tuning or something.
1051
00:29:26,000 --> 00:29:27,960
Doing additional post training.
1052
00:29:28,000 --> 00:29:28,520
Post training
1053
00:29:29,180 --> 00:29:31,080
has been able to uh,
1054
00:29:31,740 --> 00:29:36,080
give significant quality improvements and
1055
00:29:36,820 --> 00:29:38,160
so it's getting cheaper
1056
00:29:38,780 --> 00:29:41,060
and faster and smarter.
1057
00:29:41,200 --> 00:29:43,480
So all the good things, uh,
1058
00:29:43,700 --> 00:29:44,540
and it's improving.
1059
00:29:44,560 --> 00:29:45,880
And I don't know why
1060
00:29:47,080 --> 00:29:48,200
Deep Seek is
1061
00:29:49,040 --> 00:29:50,980
not increasing their model numbers.
1062
00:29:51,040 --> 00:29:52,980
Like I feel like most companies
1063
00:29:53,020 --> 00:29:54,480
would not call this a deep
1064
00:29:54,540 --> 00:29:55,820
seq v4.1.
1065
00:29:56,300 --> 00:29:57,060
They would call it maybe a
1066
00:29:57,120 --> 00:29:59,220
deep seq v4.2,
1067
00:29:59,740 --> 00:30:01,640
but instead they're just doing another
1068
00:30:01,700 --> 00:30:03,520
release and changing the date.
1069
00:30:03,600 --> 00:30:04,480
So fine,
1070
00:30:04,540 --> 00:30:05,340
whatever they want to do.
1071
00:30:05,400 --> 00:30:06,460
I don't know if it's like
1072
00:30:06,600 --> 00:30:09,060
regulatory reasons or they're just so
1073
00:30:09,120 --> 00:30:10,600
humble that they don't want to
1074
00:30:11,000 --> 00:30:11,520
say something.
1075
00:30:11,600 --> 00:30:13,220
But again, Deep Seek,
1076
00:30:13,660 --> 00:30:15,540
even though the model number hasn't
1077
00:30:15,600 --> 00:30:16,400
changed, uh,
1078
00:30:16,560 --> 00:30:19,700
the quality has increased by actually
1079
00:30:19,820 --> 00:30:21,020
quite a bit, um,
1080
00:30:21,200 --> 00:30:22,880
with just some additional post training.
1081
00:30:22,940 --> 00:30:24,480
So good job Deep Seek.
1082
00:30:24,780 --> 00:30:26,880
Also, not to be left out,
1083
00:30:27,880 --> 00:30:30,800
Minimax, uh, just today came out,
1084
00:30:31,360 --> 00:30:32,780
or actually yesterday came out with
1085
00:30:32,840 --> 00:30:34,800
the M Mini max M M
1086
00:30:35,060 --> 00:30:37,600
3.1 flash,
1087
00:30:38,600 --> 00:30:40,560
which seems to be doing very
1088
00:30:40,620 --> 00:30:43,060
well on some benchmarks.
1089
00:30:43,580 --> 00:30:45,740
Um, it seems to be uh,
1090
00:30:45,980 --> 00:30:47,960
kind of right along uh,
1091
00:30:48,420 --> 00:30:51,140
in terms of raw capability, uh,
1092
00:30:52,100 --> 00:30:56,160
around GLM 5.3.
1093
00:30:57,120 --> 00:30:59,060
Uh, so it's pretty good.
1094
00:30:59,260 --> 00:31:00,360
Or you could say it's around
1095
00:31:00,380 --> 00:31:03,320
the same level as Haiku, Um,
1096
00:31:03,960 --> 00:31:06,320
but it's also open weight.
1097
00:31:06,460 --> 00:31:09,440
So I support uh, it, um,
1098
00:31:10,160 --> 00:31:11,660
but a lot of people are
1099
00:31:11,740 --> 00:31:13,300
claiming that there may be a
1100
00:31:13,340 --> 00:31:16,140
bit of benchmaxing going on.
1101
00:31:16,260 --> 00:31:16,580
I don't know,
1102
00:31:16,660 --> 00:31:17,980
I haven't tried the model yet.
1103
00:31:18,580 --> 00:31:20,380
But I love to see that
1104
00:31:20,440 --> 00:31:22,820
there are new releases coming out
1105
00:31:22,980 --> 00:31:23,200
um,
1106
00:31:23,520 --> 00:31:25,700
from more and more companies all
1107
00:31:25,780 --> 00:31:28,100
at a very affordable rates for
1108
00:31:28,140 --> 00:31:29,460
their API.
1109
00:31:29,500 --> 00:31:30,700
And I think that what is
1110
00:31:30,780 --> 00:31:32,880
also very exciting is when the
1111
00:31:32,980 --> 00:31:34,460
input and output tokens
1112
00:31:34,860 --> 00:31:36,660
and also the caching tokens go
1113
00:31:36,720 --> 00:31:38,260
so cheap, um,
1114
00:31:38,440 --> 00:31:39,720
I think it's very reasonable to
1115
00:31:39,760 --> 00:31:43,140
just spend on the API directly
1116
00:31:43,240 --> 00:31:45,260
as opposed to getting uh,
1117
00:31:45,560 --> 00:31:47,200
these subscription plans.
1118
00:31:47,260 --> 00:31:47,460
Because
1119
00:31:48,840 --> 00:31:50,900
even if you use a large
1120
00:31:50,980 --> 00:31:52,700
amount of input or output tokens,
1121
00:31:53,140 --> 00:31:54,320
oftentimes a lot of people are
1122
00:31:54,380 --> 00:31:56,760
using caching quite heavily.
1123
00:31:56,900 --> 00:31:59,400
So if you're 99% cache then
1124
00:31:59,460 --> 00:32:00,840
your cache input and output
1125
00:32:01,360 --> 00:32:02,820
might only cost like a penny
1126
00:32:02,860 --> 00:32:03,860
or less than a penny.
1127
00:32:04,080 --> 00:32:04,160
Um,
1128
00:32:04,500 --> 00:32:07,940
so oftentimes it's not
1129
00:32:08,660 --> 00:32:11,060
yet cheaper than a subscription
1130
00:32:11,460 --> 00:32:11,860
but
1131
00:32:12,420 --> 00:32:14,060
for a lot of these models
1132
00:32:14,060 --> 00:32:16,180
it's becoming almost
1133
00:32:16,820 --> 00:32:19,340
cheaper just or like as cheap
1134
00:32:19,340 --> 00:32:22,100
as the subsidized subscriptions to just
1135
00:32:22,120 --> 00:32:22,270
uh,
1136
00:32:22,270 --> 00:32:24,730
pay for the API uh directly.
1137
00:32:24,730 --> 00:32:26,650
Which I mean I think overall
1138
00:32:26,650 --> 00:32:28,010
is a good thing.
1139
00:32:28,380 --> 00:32:30,180
Um, so that like uh,
1140
00:32:30,180 --> 00:32:31,050
these companies
1141
00:32:31,450 --> 00:32:32,850
won't be losing money but they
1142
00:32:32,850 --> 00:32:35,370
can actually ah, price at a,
1143
00:32:36,090 --> 00:32:37,690
something where they won't be losing
1144
00:32:37,690 --> 00:32:37,970
money.
1145
00:32:37,970 --> 00:32:38,970
Because I don't,
1146
00:32:39,290 --> 00:32:40,690
honestly I don't want to see
1147
00:32:40,690 --> 00:32:41,650
these companies losing money.
1148
00:32:41,650 --> 00:32:43,130
I want to see these companies
1149
00:32:43,290 --> 00:32:43,690
being
1150
00:32:44,250 --> 00:32:45,930
profitable so that they can continue
1151
00:32:46,090 --> 00:32:47,270
to uh,
1152
00:32:47,270 --> 00:32:49,290
iterate and improve on their models.
1153
00:32:49,450 --> 00:32:51,490
And I think that the like,
1154
00:32:51,490 --> 00:32:52,730
I'm okay with them making
1155
00:32:53,210 --> 00:32:53,850
a profit.
1156
00:32:53,850 --> 00:32:54,970
Like I don't want them to
1157
00:32:54,970 --> 00:32:57,290
be losing money acquiring market share,
1158
00:32:57,290 --> 00:32:58,570
putting everybody else out of business
1159
00:32:58,730 --> 00:33:00,970
and then um, just being,
1160
00:33:01,050 --> 00:33:02,250
and then jacking up their price
1161
00:33:02,250 --> 00:33:02,570
later.
1162
00:33:02,730 --> 00:33:03,730
We don't want to see that.
1163
00:33:03,730 --> 00:33:04,770
Like I would love to see
1164
00:33:04,770 --> 00:33:06,730
them making a healthy margin,
1165
00:33:06,730 --> 00:33:08,970
maybe 20% or whatever it is
1166
00:33:09,210 --> 00:33:10,990
and then uh,
1167
00:33:11,210 --> 00:33:12,930
building a sustainable company on that
1168
00:33:12,930 --> 00:33:14,210
and then people just use an
1169
00:33:14,210 --> 00:33:14,810
API,
1170
00:33:15,130 --> 00:33:16,210
you just pay for what you
1171
00:33:16,210 --> 00:33:16,410
use.
1172
00:33:16,410 --> 00:33:17,450
I think that'd be a
1173
00:33:18,040 --> 00:33:19,680
really great sustainable business model going
1174
00:33:19,680 --> 00:33:20,040
forward.
1175
00:33:21,720 --> 00:33:23,480
Yeah, um, I,
1176
00:33:23,480 --> 00:33:24,790
I don't think ah,
1177
00:33:24,790 --> 00:33:26,200
it is good for the world
1178
00:33:26,200 --> 00:33:27,960
at large if these companies are
1179
00:33:28,200 --> 00:33:29,840
operating at a loss because some
1180
00:33:29,840 --> 00:33:31,360
of these debt numbers that these
1181
00:33:31,360 --> 00:33:32,520
companies are taking on is,
1182
00:33:32,680 --> 00:33:33,960
is just outrageous.
1183
00:33:34,290 --> 00:33:36,460
Um, OpenAI has uh,
1184
00:33:36,460 --> 00:33:38,240
taken on reportedly like $50 billion
1185
00:33:38,560 --> 00:33:40,290
of debt for their um,
1186
00:33:40,290 --> 00:33:41,480
Broadcom spending.
1187
00:33:41,480 --> 00:33:43,170
Broadcom chip, ah,
1188
00:33:43,170 --> 00:33:45,640
space SpaceX reportedly is taking on
1189
00:33:45,640 --> 00:33:46,680
$40 billion.
1190
00:33:47,270 --> 00:33:47,280
Uh,
1191
00:33:47,650 --> 00:33:50,130
Nvidia has $500 billion in commitments.
1192
00:33:50,370 --> 00:33:51,410
And so a lot of these
1193
00:33:51,410 --> 00:33:53,330
are debt financing which
1194
00:33:53,730 --> 00:33:55,490
is money that they don't have
1195
00:33:55,810 --> 00:33:58,210
to serve users at a subsidy
1196
00:33:58,210 --> 00:33:59,290
which uh,
1197
00:33:59,410 --> 00:34:01,210
is just a recipe for disaster
1198
00:34:01,210 --> 00:34:02,610
for the global financial market.
1199
00:34:03,290 --> 00:34:03,790
Um,
1200
00:34:03,890 --> 00:34:04,690
so it's good
1201
00:34:05,250 --> 00:34:07,010
that some of these companies are
1202
00:34:07,010 --> 00:34:07,730
making money.
1203
00:34:08,170 --> 00:34:08,520
Um,
1204
00:34:10,210 --> 00:34:13,250
but moving on to the model
1205
00:34:13,250 --> 00:34:16,870
launches, um for our audio listeners,
1206
00:34:16,870 --> 00:34:17,880
I showed some of uh,
1207
00:34:18,230 --> 00:34:18,630
the
1208
00:34:19,430 --> 00:34:22,590
graphs from artificial analysis and it
1209
00:34:22,590 --> 00:34:24,870
clearly shows that the Chinese models
1210
00:34:25,190 --> 00:34:26,910
are at the top of the
1211
00:34:26,910 --> 00:34:29,270
open source leaderboard.
1212
00:34:29,350 --> 00:34:30,990
But we still have some cool
1213
00:34:30,990 --> 00:34:32,550
models from the US companies coming
1214
00:34:32,550 --> 00:34:32,870
out.
1215
00:34:33,310 --> 00:34:33,670
Um,
1216
00:34:33,670 --> 00:34:36,070
Nematron 3.5 from Nvidia has come
1217
00:34:36,070 --> 00:34:36,390
out.
1218
00:34:36,870 --> 00:34:40,240
Reflection AI has introduced Beam,
1219
00:34:40,400 --> 00:34:41,650
which is uh,
1220
00:34:41,650 --> 00:34:44,120
one of their open weight model.
1221
00:34:44,120 --> 00:34:46,170
Which is awesome because, uh,
1222
00:34:46,170 --> 00:34:47,320
similar to how we were talking
1223
00:34:47,320 --> 00:34:47,600
about
1224
00:34:48,000 --> 00:34:50,320
people just reverse engineering anything and
1225
00:34:50,320 --> 00:34:52,400
like, you know, mixing and mashing,
1226
00:34:53,100 --> 00:34:54,220
uh, games, uh,
1227
00:34:54,240 --> 00:34:55,640
some from some of their favorite
1228
00:34:55,640 --> 00:34:58,880
nostalgic childhood experiences, um,
1229
00:34:58,880 --> 00:35:00,240
I could imagine people
1230
00:35:01,040 --> 00:35:02,640
mixing and matching models.
1231
00:35:03,060 --> 00:35:03,150
Uh,
1232
00:35:03,520 --> 00:35:05,320
that's still kind of like black
1233
00:35:05,320 --> 00:35:05,840
magic.
1234
00:35:05,840 --> 00:35:06,880
And uh,
1235
00:35:07,560 --> 00:35:10,360
it's a very esoteric practice to
1236
00:35:10,360 --> 00:35:11,320
mix models.
1237
00:35:11,480 --> 00:35:13,240
But you could fine tune them,
1238
00:35:13,240 --> 00:35:14,560
you could like tailor them to
1239
00:35:14,560 --> 00:35:16,330
your own needs, um,
1240
00:35:16,600 --> 00:35:18,120
you could modify them and I
1241
00:35:18,120 --> 00:35:19,320
feel like that would be
1242
00:35:19,800 --> 00:35:22,040
a cool new wave.
1243
00:35:22,760 --> 00:35:22,900
Um,
1244
00:35:22,920 --> 00:35:24,120
if we were to see that.
1245
00:35:24,280 --> 00:35:25,360
I think we're still a little
1246
00:35:25,360 --> 00:35:28,160
bit far away because consumer hardware
1247
00:35:28,160 --> 00:35:28,960
can't really do that.
1248
00:35:28,960 --> 00:35:31,480
So your average hacker or developer
1249
00:35:31,720 --> 00:35:34,030
can just take a $500 billion,
1250
00:35:34,156 --> 00:35:37,240
500 billion parameter model from Reflection
1251
00:35:37,240 --> 00:35:39,200
AI and then fine tune them
1252
00:35:39,200 --> 00:35:39,840
easily.
1253
00:35:40,560 --> 00:35:42,160
But it would be cool to
1254
00:35:42,160 --> 00:35:43,840
see a world where that happens
1255
00:35:44,000 --> 00:35:45,080
and then you can just fine
1256
00:35:45,080 --> 00:35:45,920
tune it to your, uh,
1257
00:35:46,200 --> 00:35:46,960
daily environment.
1258
00:35:47,120 --> 00:35:49,510
Like, hey, help me um,
1259
00:35:50,080 --> 00:35:51,680
figure out my favorite shows on
1260
00:35:51,680 --> 00:35:53,880
TV and like run my home
1261
00:35:53,880 --> 00:35:54,480
for me,
1262
00:35:54,480 --> 00:35:56,510
order my groceries and um,
1263
00:35:56,640 --> 00:35:58,520
answer my emails exactly in the
1264
00:35:58,520 --> 00:35:59,360
tone that I do
1265
00:35:59,990 --> 00:36:01,750
or whatever else your personal use
1266
00:36:01,750 --> 00:36:02,470
case might be.
1267
00:36:04,310 --> 00:36:04,450
Yeah,
1268
00:36:04,530 --> 00:36:07,070
this is an interesting model because
1269
00:36:07,210 --> 00:36:09,310
I actually hadn't heard about Beam
1270
00:36:09,370 --> 00:36:12,110
until you um, mentioned it here.
1271
00:36:12,670 --> 00:36:14,350
Uh, is Beam its own company?
1272
00:36:15,370 --> 00:36:17,070
No, ah, it's from Reflection AI,
1273
00:36:17,170 --> 00:36:21,030
which is popular, um,
1274
00:36:21,190 --> 00:36:22,450
AI research lab, uh,
1275
00:36:22,590 --> 00:36:23,270
in the U.S.
1276
00:36:24,190 --> 00:36:24,610
uh,
1277
00:36:24,730 --> 00:36:25,710
let me pull up some more
1278
00:36:25,770 --> 00:36:26,730
information about them.
1279
00:36:28,150 --> 00:36:30,290
Yeah, that's pretty, ah, neat.
1280
00:36:30,670 --> 00:36:32,390
Um, but I have to,
1281
00:36:32,790 --> 00:36:32,970
I think,
1282
00:36:33,190 --> 00:36:34,090
I don't know if we talked
1283
00:36:34,129 --> 00:36:34,830
about it last time,
1284
00:36:34,910 --> 00:36:36,130
but another model that I want
1285
00:36:36,150 --> 00:36:37,710
to mention while you're pulling up
1286
00:36:37,790 --> 00:36:41,510
about reflection is haiku 5.5 from
1287
00:36:42,010 --> 00:36:46,150
opus. So Anthropic is.
1288
00:36:46,730 --> 00:36:47,650
I want to talk more about
1289
00:36:47,730 --> 00:36:49,570
Anthropic's financials later on in the
1290
00:36:49,650 --> 00:36:50,030
episode,
1291
00:36:50,130 --> 00:36:50,330
but
1292
00:36:51,370 --> 00:36:54,070
Anthropic is really kind of putting
1293
00:36:54,090 --> 00:36:55,330
their best foot forward
1294
00:36:55,890 --> 00:36:57,130
and they are
1295
00:36:58,270 --> 00:36:59,650
really coming to compete.
1296
00:36:59,930 --> 00:37:02,110
So for reference, um,
1297
00:37:02,210 --> 00:37:02,410
I know
1298
00:37:03,230 --> 00:37:04,650
I wouldn't normally talk about these
1299
00:37:04,770 --> 00:37:06,610
small little models, um,
1300
00:37:06,950 --> 00:37:08,830
but if you look at the
1301
00:37:10,550 --> 00:37:14,350
benchmarks, uh, for 5.5, uh,
1302
00:37:14,910 --> 00:37:17,130
you'll see that the
1303
00:37:17,530 --> 00:37:18,790
terms of raw capability.
1304
00:37:19,270 --> 00:37:19,670
Claude,
1305
00:37:20,170 --> 00:37:23,210
Haiku 5.5 is very close
1306
00:37:23,750 --> 00:37:25,690
in terms of raw capability to
1307
00:37:25,810 --> 00:37:30,250
Kimik3 so remember Kimik3
1308
00:37:30,870 --> 00:37:31,511
as of um,
1309
00:37:31,830 --> 00:37:33,590
what like two months ago was
1310
00:37:33,650 --> 00:37:36,170
the top of the line open
1311
00:37:36,290 --> 00:37:37,650
source model.
1312
00:37:38,050 --> 00:37:38,290
That
1313
00:37:38,970 --> 00:37:40,590
or sorry open weight model that
1314
00:37:40,630 --> 00:37:41,390
required
1315
00:37:42,030 --> 00:37:43,310
you to run it on
1316
00:37:43,850 --> 00:37:45,650
very high end hardware that was
1317
00:37:45,710 --> 00:37:47,750
basically impossible for people to run
1318
00:37:47,850 --> 00:37:49,870
on their own devices.
1319
00:37:50,490 --> 00:37:52,570
People were singing the praises saying
1320
00:37:52,650 --> 00:37:53,650
China has fully caught up.
1321
00:37:53,790 --> 00:37:53,850
I,
1322
00:37:53,930 --> 00:37:55,050
I'm sure I probably even said
1323
00:37:55,090 --> 00:37:55,510
that too.
1324
00:37:56,290 --> 00:37:59,750
And kimik3 was lauded for.
1325
00:37:59,910 --> 00:38:00,650
And that was,
1326
00:38:00,970 --> 00:38:02,590
that was a 3 trillion parameter
1327
00:38:02,650 --> 00:38:02,910
model.
1328
00:38:03,010 --> 00:38:03,730
Like 2 point uh,
1329
00:38:04,050 --> 00:38:06,090
78 trillion parameter model. Yes.
1330
00:38:06,370 --> 00:38:09,270
And very big it's caught uh,
1331
00:38:09,330 --> 00:38:12,210
so kimik3 costs around $3
1332
00:38:12,710 --> 00:38:14,090
per million in and $15
1333
00:38:15,130 --> 00:38:15,890
from being out.
1334
00:38:16,190 --> 00:38:20,050
So again it's cheaper than
1335
00:38:20,430 --> 00:38:21,690
a lot of the other top
1336
00:38:21,730 --> 00:38:22,650
of the models at the time.
1337
00:38:22,710 --> 00:38:23,410
So people were saying wow,
1338
00:38:23,510 --> 00:38:26,870
Kimmy K3 is so inexpensive
1339
00:38:28,090 --> 00:38:29,330
everybody should be worried.
1340
00:38:29,930 --> 00:38:30,090
Um,
1341
00:38:30,530 --> 00:38:34,290
but Anthropic said hold my beer,
1342
00:38:35,090 --> 00:38:36,390
check out Haiku.
1343
00:38:36,930 --> 00:38:37,530
So Claude,
1344
00:38:37,590 --> 00:38:40,150
Haiku is
1345
00:38:41,470 --> 00:38:43,470
with like almost the same in
1346
00:38:43,530 --> 00:38:45,430
terms of raw intelligence as K3
1347
00:38:45,870 --> 00:38:46,150
but
1348
00:38:46,750 --> 00:38:49,190
the cost on the API
1349
00:38:50,270 --> 00:38:51,430
a tenth of the price per
1350
00:38:51,490 --> 00:38:51,970
task.
1351
00:38:52,590 --> 00:38:54,310
Yes it is
1352
00:38:54,850 --> 00:38:56,650
10 cents per million in and
1353
00:38:57,270 --> 00:38:59,850
50 cents per million out um,
1354
00:39:00,270 --> 00:39:01,770
for under 100,000 tokens.
1355
00:39:01,790 --> 00:39:02,890
And then after that uh,
1356
00:39:03,070 --> 00:39:04,450
they 5x the price to
1357
00:39:04,850 --> 00:39:06,090
like 50 cents per million and
1358
00:39:06,110 --> 00:39:07,590
$2.50 per million million out.
1359
00:39:08,090 --> 00:39:11,050
Um, but basically it's like ah,
1360
00:39:11,350 --> 00:39:11,770
what is that?
1361
00:39:11,850 --> 00:39:12,650
I can't even do the math.
1362
00:39:12,690 --> 00:39:14,810
It's like a 30x
1363
00:39:15,370 --> 00:39:18,530
reduction in price from Kimmy K3.
1364
00:39:18,690 --> 00:39:19,230
Is that, is that right?
1365
00:39:19,750 --> 00:39:20,410
If 15.
1366
00:39:20,910 --> 00:39:22,370
Uh, uh, let me,
1367
00:39:22,450 --> 00:39:23,110
let me do this math.
1368
00:39:23,310 --> 00:39:24,830
15 divided by 0.5.
1369
00:39:25,590 --> 00:39:26,210
Yeah, yeah it's ah,
1370
00:39:26,270 --> 00:39:27,950
it's a 30, 30.
1371
00:39:28,230 --> 00:39:29,410
It's 30 times cheaper
1372
00:39:30,210 --> 00:39:32,230
um than Kimmy K3 and it
1373
00:39:32,290 --> 00:39:33,810
only took like a couple months
1374
00:39:33,930 --> 00:39:35,970
and it is also
1375
00:39:36,370 --> 00:39:37,970
way faster too.
1376
00:39:38,110 --> 00:39:38,870
So um,
1377
00:39:39,630 --> 00:39:41,090
you're able to get around
1378
00:39:41,570 --> 00:39:43,830
200 plus tokens per second
1379
00:39:44,470 --> 00:39:45,950
on Haiku and
1380
00:39:46,630 --> 00:39:48,190
it has still a million token
1381
00:39:48,250 --> 00:39:50,690
context window and it's
1382
00:39:51,270 --> 00:39:53,490
just as smart as Kimike 3.
1383
00:39:53,750 --> 00:39:53,930
Um,
1384
00:39:54,210 --> 00:39:56,230
like anthropic right now is kind
1385
00:39:56,270 --> 00:39:56,310
of
1386
00:39:56,870 --> 00:39:58,710
wiping the floor with all these
1387
00:39:58,770 --> 00:40:00,990
other model companies by a lot.
1388
00:40:01,290 --> 00:40:02,350
And um,
1389
00:40:03,310 --> 00:40:04,370
in my own testing
1390
00:40:04,890 --> 00:40:06,490
even though it shows that like
1391
00:40:06,550 --> 00:40:08,690
deep seq 4.1 flash and like
1392
00:40:08,990 --> 00:40:10,250
Xiaomi or uh,
1393
00:40:10,370 --> 00:40:12,550
whatever have slightly higher
1394
00:40:13,130 --> 00:40:14,650
scores than Haiku.
1395
00:40:14,790 --> 00:40:16,610
In my testing for my tasks
1396
00:40:16,750 --> 00:40:19,070
I have found that Haiku um,
1397
00:40:19,370 --> 00:40:20,590
makes less mistakes
1398
00:40:21,010 --> 00:40:24,190
and is overall smarter uh
1399
00:40:24,790 --> 00:40:25,130
than
1400
00:40:25,610 --> 00:40:26,790
some of like the Deep Sea
1401
00:40:26,850 --> 00:40:27,630
models or
1402
00:40:28,090 --> 00:40:30,130
some of the Xiaomi like V2.6
1403
00:40:30,170 --> 00:40:31,170
Pro, um,
1404
00:40:31,430 --> 00:40:34,570
that had better benchmark scores
1405
00:40:35,649 --> 00:40:36,330
at least for,
1406
00:40:36,510 --> 00:40:37,570
for my use cases.
1407
00:40:38,010 --> 00:40:38,150
Um,
1408
00:40:38,410 --> 00:40:40,770
Haiku has been just knocking my
1409
00:40:40,830 --> 00:40:41,810
socks off and
1410
00:40:42,270 --> 00:40:44,390
completely blowing me away with how
1411
00:40:44,470 --> 00:40:46,310
good it is for the price.
1412
00:40:46,390 --> 00:40:47,290
I mean, um,
1413
00:40:47,790 --> 00:40:49,370
we are getting incredibly
1414
00:40:49,890 --> 00:40:51,930
close to intelligence that is just
1415
00:40:51,990 --> 00:40:52,870
too cheap to monitor.
1416
00:40:52,910 --> 00:40:54,030
And I think that for most
1417
00:40:54,090 --> 00:40:55,150
people, uh,
1418
00:40:55,550 --> 00:40:59,450
Haiku is more than good enough.
1419
00:40:59,550 --> 00:41:01,090
And unless you're working on something
1420
00:41:01,810 --> 00:41:02,010
uh,
1421
00:41:02,510 --> 00:41:02,650
like
1422
00:41:03,510 --> 00:41:05,870
scientific or really that'll stress the
1423
00:41:05,930 --> 00:41:06,270
models.
1424
00:41:06,350 --> 00:41:06,590
I mean,
1425
00:41:06,790 --> 00:41:08,189
Haiku is just
1426
00:41:08,930 --> 00:41:10,890
an absolutely amazing model.
1427
00:41:10,910 --> 00:41:11,650
And I think that
1428
00:41:12,570 --> 00:41:13,850
I would argue that even though
1429
00:41:13,890 --> 00:41:15,910
this isn't really true on all
1430
00:41:15,950 --> 00:41:16,610
the benchmarks,
1431
00:41:17,190 --> 00:41:19,630
in my opinion Haiku is really
1432
00:41:19,690 --> 00:41:21,610
kind of pushing forward the Pareto
1433
00:41:22,090 --> 00:41:26,030
frontier on terms of like price
1434
00:41:26,110 --> 00:41:28,590
and speed and intelligence.
1435
00:41:28,910 --> 00:41:29,130
Um,
1436
00:41:29,990 --> 00:41:30,210
and
1437
00:41:30,650 --> 00:41:31,610
maybe not everybody would grab the
1438
00:41:31,650 --> 00:41:33,250
benchmark, but it's right there.
1439
00:41:33,450 --> 00:41:34,710
And I think that for most
1440
00:41:34,770 --> 00:41:35,690
people, uh,
1441
00:41:36,010 --> 00:41:37,410
Haiku may be the model that
1442
00:41:37,490 --> 00:41:38,630
you want to use
1443
00:41:39,010 --> 00:41:40,750
if you're okay with Anthropic.
1444
00:41:40,790 --> 00:41:40,990
And now
1445
00:41:42,510 --> 00:41:44,330
Anthropic has a lot of issues,
1446
00:41:44,970 --> 00:41:46,790
but right now it's like the
1447
00:41:46,850 --> 00:41:47,910
best model for the price in
1448
00:41:47,950 --> 00:41:53,170
my opinion. Um, yeah, it's amazing.
1449
00:41:53,250 --> 00:41:54,850
And uh, speaking of, uh,
1450
00:41:55,470 --> 00:41:57,350
fast and cheap, um,
1451
00:41:57,470 --> 00:41:58,270
I'm going to switch gears a
1452
00:41:58,310 --> 00:42:00,030
little bit and talk about another
1453
00:42:00,090 --> 00:42:00,390
model
1454
00:42:00,910 --> 00:42:02,690
that is not quite a transformer
1455
00:42:02,730 --> 00:42:03,050
model.
1456
00:42:03,270 --> 00:42:05,670
It is a text diffusion model.
1457
00:42:06,330 --> 00:42:07,490
And this is a company that
1458
00:42:07,510 --> 00:42:08,430
I've been following for a while
1459
00:42:08,490 --> 00:42:09,090
because uh,
1460
00:42:09,630 --> 00:42:10,790
I was interested in text diffusion
1461
00:42:10,830 --> 00:42:11,910
models when I was at Google
1462
00:42:11,970 --> 00:42:13,350
and working on uh,
1463
00:42:13,490 --> 00:42:14,050
some of the,
1464
00:42:14,470 --> 00:42:16,630
or using some of the DeepMind
1465
00:42:16,690 --> 00:42:17,690
text diffusion models.
1466
00:42:18,230 --> 00:42:18,810
And it was just
1467
00:42:19,770 --> 00:42:20,910
so fast
1468
00:42:21,410 --> 00:42:22,050
to be able to see it
1469
00:42:22,090 --> 00:42:23,790
just spit out some text and
1470
00:42:23,810 --> 00:42:25,290
then boom, it's ready.
1471
00:42:25,810 --> 00:42:27,430
That was mind, um,
1472
00:42:27,490 --> 00:42:28,530
blowing to see at that time.
1473
00:42:29,030 --> 00:42:31,470
So this, unlike the transformer models,
1474
00:42:31,690 --> 00:42:32,690
doesn't need to generate
1475
00:42:33,210 --> 00:42:34,110
one token at a time.
1476
00:42:34,610 --> 00:42:36,070
It just spits out a bunch
1477
00:42:36,130 --> 00:42:36,490
of uh,
1478
00:42:36,610 --> 00:42:38,690
characters on text and then changes
1479
00:42:38,790 --> 00:42:40,150
it to look like
1480
00:42:41,270 --> 00:42:42,550
what it needs the answer to
1481
00:42:42,590 --> 00:42:42,830
be.
1482
00:42:43,450 --> 00:42:43,590
Um,
1483
00:42:43,690 --> 00:42:44,590
so if you've used any of
1484
00:42:44,630 --> 00:42:46,230
these, uh, image diffusion models,
1485
00:42:46,890 --> 00:42:47,870
that's kind of what they do.
1486
00:42:48,050 --> 00:42:48,950
And if you've seen one of
1487
00:42:49,010 --> 00:42:52,530
them, um, in progress,
1488
00:42:52,650 --> 00:42:53,450
if you use like some of
1489
00:42:53,490 --> 00:42:54,830
the open source versions like stable
1490
00:42:54,870 --> 00:42:56,030
diffusion, um,
1491
00:42:56,150 --> 00:42:57,170
or seen a couple demos of
1492
00:42:57,210 --> 00:42:57,830
how they do it,
1493
00:42:58,150 --> 00:42:58,890
they start with a bunch of
1494
00:42:58,950 --> 00:43:00,710
noise and then keep changing the
1495
00:43:00,770 --> 00:43:03,030
pixels to get closer to what
1496
00:43:04,450 --> 00:43:06,350
the model thinks the end result
1497
00:43:06,410 --> 00:43:07,450
of that image should look like.
1498
00:43:08,130 --> 00:43:08,490
So
1499
00:43:09,770 --> 00:43:11,110
let's see how this kind of
1500
00:43:11,150 --> 00:43:11,530
compares.
1501
00:43:11,590 --> 00:43:13,070
So this compares
1502
00:43:13,630 --> 00:43:16,830
to the haiku 5.5 GPT, uh,
1503
00:43:17,350 --> 00:43:19,810
6 Luna, and it is really,
1504
00:43:19,890 --> 00:43:20,450
really fast.
1505
00:43:20,510 --> 00:43:21,830
It's like a thousand plus tokens
1506
00:43:21,870 --> 00:43:22,390
per second.
1507
00:43:23,050 --> 00:43:23,310
Um,
1508
00:43:23,490 --> 00:43:23,990
that is
1509
00:43:24,710 --> 00:43:26,410
four times as fast as the
1510
00:43:26,410 --> 00:43:26,411
uh,
1511
00:43:26,430 --> 00:43:26,790
Haiku
1512
00:43:27,250 --> 00:43:28,390
model that Mark was talking about,
1513
00:43:28,450 --> 00:43:31,210
which is already much faster, uh,
1514
00:43:31,370 --> 00:43:32,330
than Kimike 3
1515
00:43:32,950 --> 00:43:34,130
and it's at the same price.
1516
00:43:34,350 --> 00:43:34,970
They're pricing it
1517
00:43:35,830 --> 00:43:38,170
the same as haiku 5.5 maybe
1518
00:43:38,250 --> 00:43:38,530
to
1519
00:43:39,430 --> 00:43:41,210
give people an incentive to use
1520
00:43:41,290 --> 00:43:41,650
this.
1521
00:43:42,290 --> 00:43:45,850
Um, there's some, uh, uh,
1522
00:43:46,050 --> 00:43:46,950
staggered pricing.
1523
00:43:47,010 --> 00:43:48,130
It's going to be the same
1524
00:43:48,170 --> 00:43:49,270
and then go up, uh,
1525
00:43:49,410 --> 00:43:49,630
to
1526
00:43:50,410 --> 00:43:52,470
$0.20 per million output tokens
1527
00:43:53,050 --> 00:43:54,030
per million input tokens,
1528
00:43:54,070 --> 00:43:55,890
and then $0.7 per million output,
1529
00:43:56,390 --> 00:43:57,890
which is in the same ballpark.
1530
00:43:58,570 --> 00:43:58,750
Uh,
1531
00:43:58,930 --> 00:43:59,390
but this is
1532
00:43:59,950 --> 00:44:01,570
really cool in a world where
1533
00:44:02,110 --> 00:44:03,310
everything is a transformer.
1534
00:44:03,710 --> 00:44:03,950
Uh,
1535
00:44:03,950 --> 00:44:05,890
it's cool to see this alternative
1536
00:44:05,890 --> 00:44:07,890
approach because in my mind
1537
00:44:08,370 --> 00:44:10,250
I think this would be amazing
1538
00:44:10,250 --> 00:44:12,210
for use cases that need edits
1539
00:44:12,450 --> 00:44:13,970
as opposed to whole
1540
00:44:14,450 --> 00:44:15,890
generations from scratch.
1541
00:44:16,370 --> 00:44:18,930
If you have an existing document,
1542
00:44:19,410 --> 00:44:20,850
this would be able to edit
1543
00:44:21,570 --> 00:44:23,330
just that one specific thing.
1544
00:44:23,770 --> 00:44:23,900
Um,
1545
00:44:24,530 --> 00:44:26,370
it would be equivalent to
1546
00:44:27,130 --> 00:44:29,610
if your coding ide
1547
00:44:30,330 --> 00:44:32,770
is doing inline edits as opposed
1548
00:44:32,770 --> 00:44:34,370
to generating the entire code from
1549
00:44:34,370 --> 00:44:34,810
scratch.
1550
00:44:35,430 --> 00:44:35,770
Um,
1551
00:44:36,010 --> 00:44:37,610
I don't know what models do
1552
00:44:37,690 --> 00:44:39,250
nowadays because a lot of it
1553
00:44:39,250 --> 00:44:40,050
is, um,
1554
00:44:40,490 --> 00:44:42,849
hidden and obscured behind the coding
1555
00:44:42,849 --> 00:44:43,530
agent harness.
1556
00:44:43,930 --> 00:44:45,850
But this would be great for
1557
00:44:45,850 --> 00:44:46,890
inline edits.
1558
00:44:47,290 --> 00:44:49,460
Some small tweaks, um,
1559
00:44:49,770 --> 00:44:51,770
maybe giving an answer where
1560
00:44:52,420 --> 00:44:54,580
the beginning of the answer references
1561
00:44:54,580 --> 00:44:55,700
something at the end.
1562
00:44:56,270 --> 00:44:56,750
Um,
1563
00:44:56,750 --> 00:44:57,460
where it could
1564
00:44:57,940 --> 00:44:59,460
change it as it is
1565
00:45:00,100 --> 00:45:02,030
editing all the tokens, uh,
1566
00:45:02,030 --> 00:45:02,740
on the page.
1567
00:45:02,740 --> 00:45:05,300
So another cool model from a
1568
00:45:05,780 --> 00:45:08,459
new AI lab or an existing
1569
00:45:08,459 --> 00:45:09,840
AI lab, uh,
1570
00:45:09,940 --> 00:45:12,100
but flying under the radar.
1571
00:45:12,900 --> 00:45:15,620
So I'm not super familiar with
1572
00:45:16,180 --> 00:45:18,180
what all of what you said
1573
00:45:18,180 --> 00:45:18,580
means.
1574
00:45:18,930 --> 00:45:19,210
Um,
1575
00:45:19,550 --> 00:45:20,750
can you maybe dumb it down
1576
00:45:20,910 --> 00:45:22,350
for the rest of us and
1577
00:45:22,430 --> 00:45:25,430
explain a little bit like, uh,
1578
00:45:25,430 --> 00:45:27,430
where you'd want to use this
1579
00:45:27,430 --> 00:45:28,270
type of model?
1580
00:45:28,750 --> 00:45:29,870
I know you said edits,
1581
00:45:29,870 --> 00:45:31,630
but I don't think I quite
1582
00:45:31,870 --> 00:45:33,150
understand what that means.
1583
00:45:33,670 --> 00:45:33,830
Um,
1584
00:45:33,830 --> 00:45:35,549
because I know that LLMs can
1585
00:45:35,549 --> 00:45:35,950
also
1586
00:45:36,670 --> 00:45:39,310
do edits as well to
1587
00:45:39,870 --> 00:45:40,750
code bases.
1588
00:45:41,080 --> 00:45:42,080
Um,
1589
00:45:42,080 --> 00:45:42,830
but maybe
1590
00:45:43,470 --> 00:45:44,750
dig into what you mean by
1591
00:45:44,830 --> 00:45:45,550
edits here,
1592
00:45:45,550 --> 00:45:46,350
because I think you're trying to
1593
00:45:46,350 --> 00:45:46,990
make a distinction.
1594
00:45:46,990 --> 00:45:48,190
I just don't understand what you're
1595
00:45:48,190 --> 00:45:48,670
trying to say.
1596
00:45:49,820 --> 00:45:53,310
Um, uh, to clarify, um,
1597
00:45:54,130 --> 00:45:56,730
traditional LLMs transformer base can't do
1598
00:45:56,730 --> 00:45:57,250
edits.
1599
00:45:57,650 --> 00:45:59,170
They use tools
1600
00:45:59,810 --> 00:46:02,210
and output text that says, hey,
1601
00:46:02,210 --> 00:46:04,450
call this tool and edit
1602
00:46:05,010 --> 00:46:06,690
and change this sentence
1603
00:46:07,090 --> 00:46:08,090
that says blah,
1604
00:46:08,090 --> 00:46:09,890
blah blah into this other sentence
1605
00:46:09,890 --> 00:46:11,090
that says blah blah blah.
1606
00:46:11,650 --> 00:46:15,110
And this can take, uh,
1607
00:46:15,260 --> 00:46:16,540
an existing text,
1608
00:46:17,420 --> 00:46:19,020
spit it out immediately,
1609
00:46:19,500 --> 00:46:21,540
and then change one specific part
1610
00:46:21,540 --> 00:46:22,980
of it as opposed to a
1611
00:46:22,980 --> 00:46:24,300
traditional transformer.
1612
00:46:24,620 --> 00:46:25,900
If it wants to give you
1613
00:46:25,900 --> 00:46:28,300
a new, um, page,
1614
00:46:28,380 --> 00:46:29,980
it will generate one token at
1615
00:46:29,980 --> 00:46:30,380
a time.
1616
00:46:30,860 --> 00:46:32,860
So tldr to dumb it down.
1617
00:46:32,940 --> 00:46:34,540
This is really, really fast.
1618
00:46:34,860 --> 00:46:36,780
1,000 tokens per second as opposed
1619
00:46:36,780 --> 00:46:39,020
to Haiku's 200 tokens per second.
1620
00:46:39,420 --> 00:46:40,540
Haiku generate
1621
00:46:41,010 --> 00:46:42,450
Things one token at a time.
1622
00:46:42,930 --> 00:46:44,970
This spits everything out on the
1623
00:46:44,970 --> 00:46:47,360
screen and then change uh,
1624
00:46:47,490 --> 00:46:48,210
it spits
1625
00:46:48,770 --> 00:46:50,050
a bunch of random words on
1626
00:46:50,050 --> 00:46:51,810
the screen and then changes those
1627
00:46:51,810 --> 00:46:53,170
words to match what it
1628
00:46:53,570 --> 00:46:54,850
needs the output to be.
1629
00:46:56,609 --> 00:46:57,090
Okay,
1630
00:46:57,090 --> 00:46:58,570
so I think I get what
1631
00:46:58,570 --> 00:47:00,690
you're saying where the LLM is
1632
00:47:00,690 --> 00:47:03,570
just going and making one additional
1633
00:47:03,570 --> 00:47:04,130
token.
1634
00:47:04,530 --> 00:47:05,730
It's just one token at a
1635
00:47:05,730 --> 00:47:05,890
time.
1636
00:47:05,890 --> 00:47:06,890
So if it wants to edit
1637
00:47:06,890 --> 00:47:07,130
something,
1638
00:47:07,130 --> 00:47:08,370
it needs to edit a token
1639
00:47:08,810 --> 00:47:09,930
where it will need to read
1640
00:47:09,930 --> 00:47:11,970
the entire document and then it'll
1641
00:47:11,970 --> 00:47:13,530
say I have to search the
1642
00:47:13,530 --> 00:47:15,130
document and then make
1643
00:47:16,010 --> 00:47:18,450
a write and an additive. Right.
1644
00:47:18,450 --> 00:47:19,690
Using a tool to go and
1645
00:47:19,690 --> 00:47:21,690
then insert the cursor and then
1646
00:47:21,690 --> 00:47:23,290
type or insert the cursor,
1647
00:47:23,290 --> 00:47:24,330
delete a word and then type
1648
00:47:24,330 --> 00:47:25,690
in a new word essentially.
1649
00:47:25,770 --> 00:47:27,170
So you're saying whereas this can
1650
00:47:27,170 --> 00:47:30,250
just uh, edit it in line.
1651
00:47:30,490 --> 00:47:32,240
But I guess um,
1652
00:47:32,730 --> 00:47:34,810
since the effect is the same,
1653
00:47:35,260 --> 00:47:35,380
um,
1654
00:47:36,060 --> 00:47:38,500
I'm just wondering when like ah,
1655
00:47:38,500 --> 00:47:41,060
the end user would find maybe
1656
00:47:41,060 --> 00:47:41,900
a benefit
1657
00:47:42,380 --> 00:47:44,780
of this model over a transformer
1658
00:47:44,780 --> 00:47:46,940
that calls a tool to do
1659
00:47:46,940 --> 00:47:48,660
the edits directly.
1660
00:47:48,660 --> 00:47:49,420
Because at the end of the
1661
00:47:49,420 --> 00:47:50,620
day If I'm using
1662
00:47:51,179 --> 00:47:53,580
ChatGPT or Claude code or whatever,
1663
00:47:53,950 --> 00:47:54,330
um,
1664
00:47:54,380 --> 00:47:56,180
I could see it writing a
1665
00:47:56,180 --> 00:47:57,580
document on my behalf.
1666
00:47:57,900 --> 00:47:58,300
But
1667
00:47:58,780 --> 00:48:00,940
like I don't really care about
1668
00:48:00,940 --> 00:48:02,860
how exactly is happening under the
1669
00:48:02,860 --> 00:48:03,100
hood.
1670
00:48:03,100 --> 00:48:04,260
All that I would really care
1671
00:48:04,300 --> 00:48:06,320
about is how
1672
00:48:06,800 --> 00:48:08,620
that document gets edited.
1673
00:48:09,320 --> 00:48:10,440
So I don't know if there's
1674
00:48:10,500 --> 00:48:12,060
like a particular use case
1675
00:48:12,720 --> 00:48:14,340
where this
1676
00:48:14,820 --> 00:48:17,280
would make more sense or would
1677
00:48:17,340 --> 00:48:19,680
perform better than other LLMs.
1678
00:48:20,960 --> 00:48:22,060
Yeah, I think uh,
1679
00:48:22,300 --> 00:48:25,400
even generating large documents I think
1680
00:48:25,460 --> 00:48:26,900
this would probably be
1681
00:48:27,540 --> 00:48:28,360
much faster
1682
00:48:29,140 --> 00:48:31,200
with the same performance because to
1683
00:48:31,260 --> 00:48:32,840
generate um,
1684
00:48:33,360 --> 00:48:34,840
like a whole book for example,
1685
00:48:35,140 --> 00:48:35,320
uh,
1686
00:48:35,659 --> 00:48:36,620
maybe a whole book that is
1687
00:48:36,680 --> 00:48:38,680
within the context limit uh,
1688
00:48:38,680 --> 00:48:39,560
of 1 million tokens.
1689
00:48:40,040 --> 00:48:42,320
It would take a regular transformer
1690
00:48:42,380 --> 00:48:44,100
based LLM a while to like
1691
00:48:44,720 --> 00:48:45,600
generate line by line,
1692
00:48:45,660 --> 00:48:46,760
line by line as opposed to
1693
00:48:46,820 --> 00:48:47,140
this.
1694
00:48:47,780 --> 00:48:48,400
I feel like it would be
1695
00:48:48,460 --> 00:48:49,480
able to like spit out the
1696
00:48:49,540 --> 00:48:51,060
whole million tokens
1697
00:48:51,540 --> 00:48:52,640
and be able to edit it
1698
00:48:52,700 --> 00:48:54,720
in place much faster than a
1699
00:48:54,760 --> 00:48:55,560
traditional LLM.
1700
00:48:56,080 --> 00:48:57,180
So to generate a lot of
1701
00:48:57,260 --> 00:49:00,500
stuff and be fast at it.
1702
00:49:01,740 --> 00:49:02,580
Okay, that makes sense.
1703
00:49:02,660 --> 00:49:02,880
So
1704
00:49:03,840 --> 00:49:05,000
you'd say that probably that the
1705
00:49:05,220 --> 00:49:06,920
main kind of thing that this
1706
00:49:06,960 --> 00:49:07,540
benefits is
1707
00:49:08,160 --> 00:49:09,200
almost like you would say this
1708
00:49:09,260 --> 00:49:10,760
is kind of like an algorithmic
1709
00:49:10,840 --> 00:49:12,500
improvement to just make
1710
00:49:13,100 --> 00:49:14,320
text generation
1711
00:49:14,980 --> 00:49:16,060
a lot lot faster
1712
00:49:16,880 --> 00:49:17,820
compared to LLMs.
1713
00:49:18,840 --> 00:49:19,020
You know
1714
00:49:20,360 --> 00:49:21,260
it's tough to say if there's
1715
00:49:21,300 --> 00:49:23,280
an algorithmic improvement because uh,
1716
00:49:23,980 --> 00:49:27,960
the autoregressive uh, uh,
1717
00:49:28,000 --> 00:49:30,560
transformer based sequential generation does have
1718
00:49:30,600 --> 00:49:32,160
some benefits because a lot of
1719
00:49:32,180 --> 00:49:32,260
the
1720
00:49:32,680 --> 00:49:35,140
thinking traces are optimized for these
1721
00:49:35,220 --> 00:49:37,220
transformer based LLMs which, you know,
1722
00:49:37,600 --> 00:49:38,740
think in a sequence
1723
00:49:40,240 --> 00:49:41,000
step by step.
1724
00:49:41,440 --> 00:49:41,900
You know, we,
1725
00:49:41,940 --> 00:49:42,100
we
1726
00:49:42,520 --> 00:49:43,960
coined terms for that chain of
1727
00:49:44,000 --> 00:49:44,700
thought reasoning.
1728
00:49:45,280 --> 00:49:45,540
And
1729
00:49:46,460 --> 00:49:47,760
I don't know how well this
1730
00:49:47,800 --> 00:49:50,680
diffusion model would do in complex
1731
00:49:51,060 --> 00:49:51,540
thinking.
1732
00:49:52,200 --> 00:49:52,540
Um,
1733
00:49:52,860 --> 00:49:53,520
maybe you could like
1734
00:49:55,320 --> 00:49:58,680
think in steps by generating text,
1735
00:49:58,980 --> 00:49:59,040
uh,
1736
00:49:59,140 --> 00:50:00,560
with the diffusion model in steps
1737
00:50:00,600 --> 00:50:02,540
and then keep refining its output
1738
00:50:02,600 --> 00:50:03,740
in multiple outputs.
1739
00:50:04,440 --> 00:50:04,720
Um,
1740
00:50:05,380 --> 00:50:06,480
but this is like a different
1741
00:50:06,540 --> 00:50:07,200
kind of model
1742
00:50:07,860 --> 00:50:09,260
and the thing about optimizations is
1743
00:50:09,320 --> 00:50:09,680
that, you know,
1744
00:50:09,740 --> 00:50:11,620
like the whole world has converged
1745
00:50:11,660 --> 00:50:12,400
to this one
1746
00:50:13,259 --> 00:50:15,000
architecture of transformers which is just
1747
00:50:15,040 --> 00:50:15,920
like taken off.
1748
00:50:15,980 --> 00:50:16,660
And we,
1749
00:50:16,661 --> 00:50:18,340
we're optimizing GPUs for it,
1750
00:50:18,400 --> 00:50:20,000
we're optimizing, ah,
1751
00:50:20,100 --> 00:50:21,400
our cloud compute for it,
1752
00:50:21,540 --> 00:50:23,180
even the agent harness and architectures
1753
00:50:23,220 --> 00:50:24,280
are being optimized for it.
1754
00:50:24,360 --> 00:50:25,520
So um,
1755
00:50:25,960 --> 00:50:26,760
it's hard to say
1756
00:50:27,360 --> 00:50:29,000
how well this would fit into
1757
00:50:29,060 --> 00:50:31,120
that entire ecosystem as a whole.
1758
00:50:31,520 --> 00:50:33,220
But for this one particular use
1759
00:50:33,280 --> 00:50:35,000
case of spitting text out really
1760
00:50:35,060 --> 00:50:36,820
fast, it's really, really good.
1761
00:50:37,240 --> 00:50:38,700
So I would say try it
1762
00:50:38,800 --> 00:50:40,860
in cases where you need to
1763
00:50:40,920 --> 00:50:42,080
output a lot of text
1764
00:50:42,540 --> 00:50:43,660
and you kind of have a
1765
00:50:43,700 --> 00:50:45,640
rough idea of exactly what it
1766
00:50:45,680 --> 00:50:46,500
needs to output.
1767
00:50:47,000 --> 00:50:49,740
If you have the information
1768
00:50:50,820 --> 00:50:51,800
and you just need to
1769
00:50:53,040 --> 00:50:54,120
generate a lot of text
1770
00:50:54,540 --> 00:50:55,340
from that information,
1771
00:50:55,420 --> 00:50:56,040
I feel like this would be
1772
00:50:56,100 --> 00:50:56,860
a good use case.
1773
00:50:57,760 --> 00:50:58,860
Okay, so maybe like a, uh,
1774
00:50:59,000 --> 00:51:00,520
translation task or something like that.
1775
00:51:00,560 --> 00:51:01,300
It could be pretty good.
1776
00:51:02,040 --> 00:51:02,260
Um,
1777
00:51:03,380 --> 00:51:05,440
or maybe something fleshing out an
1778
00:51:05,500 --> 00:51:07,540
email with some bullet points maybe.
1779
00:51:07,980 --> 00:51:08,260
Yeah,
1780
00:51:09,180 --> 00:51:10,780
or maybe like making a summary
1781
00:51:10,840 --> 00:51:11,380
document
1782
00:51:11,840 --> 00:51:14,340
from something that might be good
1783
00:51:14,400 --> 00:51:15,760
where it doesn't need to do
1784
00:51:15,800 --> 00:51:17,880
a lot of thinking but you'd
1785
00:51:17,940 --> 00:51:19,380
prefer it to be fast.
1786
00:51:20,060 --> 00:51:21,640
Um, yeah, exactly. Okay.
1787
00:51:22,000 --> 00:51:23,280
All right, I like it.
1788
00:51:23,860 --> 00:51:25,040
That's a good,
1789
00:51:25,700 --> 00:51:27,220
exciting model that I think that
1790
00:51:27,260 --> 00:51:28,100
a lot of people will be
1791
00:51:28,180 --> 00:51:31,040
excited to try. Um, moving on.
1792
00:51:31,080 --> 00:51:32,620
I think that's the bulk of
1793
00:51:32,660 --> 00:51:33,340
the models.
1794
00:51:33,640 --> 00:51:33,820
Um,
1795
00:51:34,280 --> 00:51:36,240
not as many model releases this
1796
00:51:36,300 --> 00:51:36,600
week.
1797
00:51:37,140 --> 00:51:37,340
Um,
1798
00:51:37,780 --> 00:51:37,960
but
1799
00:51:38,360 --> 00:51:39,320
one thing I wanted to talk
1800
00:51:39,400 --> 00:51:42,840
about is this new Tesla Roadster,
1801
00:51:43,420 --> 00:51:44,020
which is really cool.
1802
00:51:44,021 --> 00:51:44,760
Different kind of model.
1803
00:51:45,740 --> 00:51:46,620
Yeah, I don't know,
1804
00:51:46,720 --> 00:51:47,980
it's not really AI related,
1805
00:51:48,080 --> 00:51:49,720
but man, is it cool.
1806
00:51:50,180 --> 00:51:52,180
Um, I don't know. Shit.
1807
00:51:52,380 --> 00:51:53,540
Have you seen this thing?
1808
00:51:53,680 --> 00:51:54,180
This thing is.
1809
00:51:54,420 --> 00:51:55,300
I saw the whole thing as
1810
00:51:55,320 --> 00:51:56,180
soon as it came out.
1811
00:51:56,800 --> 00:51:57,000
Yeah,
1812
00:51:57,200 --> 00:51:57,540
I saw
1813
00:51:58,120 --> 00:51:59,700
a reel on Instagram and I
1814
00:51:59,920 --> 00:52:00,800
thought it was a joke.
1815
00:52:01,160 --> 00:52:02,120
And then I was like, there's,
1816
00:52:02,180 --> 00:52:04,080
there's a Tesla announcement and this
1817
00:52:04,120 --> 00:52:05,260
might actually be real.
1818
00:52:05,920 --> 00:52:06,360
The car
1819
00:52:06,880 --> 00:52:08,220
has driven up a wall,
1820
00:52:08,900 --> 00:52:12,300
gone upside down and um, uh,
1821
00:52:12,800 --> 00:52:13,980
landed like a SpaceX
1822
00:52:14,440 --> 00:52:16,240
rocket and flipped over a bunch
1823
00:52:16,300 --> 00:52:18,080
of cars. Insane.
1824
00:52:19,420 --> 00:52:19,520
Yeah,
1825
00:52:19,600 --> 00:52:20,720
I mean this is the coolest
1826
00:52:20,800 --> 00:52:21,960
car demo that I've ever seen.
1827
00:52:22,000 --> 00:52:22,500
I mean, One,
1828
00:52:22,580 --> 00:52:23,980
the car looks like, ah,
1829
00:52:24,020 --> 00:52:26,760
a Lamborghini, which is cool looking.
1830
00:52:27,180 --> 00:52:29,540
Two, it could drive upside down,
1831
00:52:30,500 --> 00:52:32,420
it could do barrel rolls and
1832
00:52:33,100 --> 00:52:33,920
I don't know, like, uh,
1833
00:52:34,460 --> 00:52:35,660
you could take this thing to
1834
00:52:35,720 --> 00:52:36,820
go over the Grand Canyon or
1835
00:52:36,860 --> 00:52:42,040
something. Ridiculous. Absolutely insane.
1836
00:52:42,080 --> 00:52:43,380
And now I am
1837
00:52:43,900 --> 00:52:46,180
slightly skeptical because I, uh,
1838
00:52:46,400 --> 00:52:48,780
know that people only slightly, well,
1839
00:52:49,460 --> 00:52:52,660
slightly skeptical because, uh,
1840
00:52:53,120 --> 00:52:54,560
Tesla does not have a good
1841
00:52:54,640 --> 00:52:56,700
track record of actually shipping their
1842
00:52:56,760 --> 00:52:59,600
Roadsters. Um, because they.
1843
00:52:59,820 --> 00:53:00,280
What was it,
1844
00:53:00,360 --> 00:53:01,580
I think it was like 2017
1845
00:53:01,620 --> 00:53:02,560
that they were requiring a
1846
00:53:02,980 --> 00:53:06,940
$50,000 deposit, uh, for their, uh.
1847
00:53:07,240 --> 00:53:08,100
About 10 years ago.
1848
00:53:08,160 --> 00:53:08,900
Yeah, yeah,
1849
00:53:09,820 --> 00:53:11,720
and that's 10 years ago where.
1850
00:53:11,740 --> 00:53:12,340
And I don't think they were
1851
00:53:12,380 --> 00:53:13,180
giving refunds.
1852
00:53:13,240 --> 00:53:14,260
So it's like, yeah,
1853
00:53:14,400 --> 00:53:15,960
almost 10 years ago that they
1854
00:53:16,780 --> 00:53:17,161
introduced uh,
1855
00:53:17,180 --> 00:53:18,620
like Roadster V1
1856
00:53:19,360 --> 00:53:21,200
and there are no Roadster V
1857
00:53:21,300 --> 00:53:22,360
ones on the road.
1858
00:53:22,720 --> 00:53:23,560
Um, and now they're just like,
1859
00:53:23,680 --> 00:53:25,040
okay, forget that.
1860
00:53:25,100 --> 00:53:27,060
We have this new Roadster V2
1861
00:53:27,160 --> 00:53:28,580
and it can fly.
1862
00:53:29,320 --> 00:53:30,320
Uh, so that's cool.
1863
00:53:30,580 --> 00:53:31,620
But this is,
1864
00:53:31,840 --> 00:53:33,620
this is the first Tesla, uh,
1865
00:53:33,621 --> 00:53:35,760
Roadster that I'm screen sharing in
1866
00:53:35,940 --> 00:53:36,680
2019.
1867
00:53:37,360 --> 00:53:37,481
This, uh,
1868
00:53:37,500 --> 00:53:38,380
is when I was working Mountain
1869
00:53:38,440 --> 00:53:39,800
View and uh, Tesla had a,
1870
00:53:40,600 --> 00:53:42,381
um, some office, uh,
1871
00:53:42,400 --> 00:53:42,880
there or something.
1872
00:53:43,000 --> 00:53:44,040
I drove by and I saw
1873
00:53:44,080 --> 00:53:45,660
the Tesla Roadster there.
1874
00:53:46,020 --> 00:53:46,520
2019,
1875
00:53:46,580 --> 00:53:48,700
proof of concept that never shipped.
1876
00:53:49,820 --> 00:53:50,040
Yeah.
1877
00:53:50,320 --> 00:53:50,540
So
1878
00:53:50,920 --> 00:53:52,600
that is why if you're listening
1879
00:53:52,640 --> 00:53:54,120
to this and you can afford
1880
00:53:54,200 --> 00:53:54,300
it.
1881
00:53:54,340 --> 00:53:54,400
Well,
1882
00:53:54,500 --> 00:53:54,720
I mean
1883
00:53:55,240 --> 00:53:55,900
if you can afford it,
1884
00:53:55,960 --> 00:53:56,600
it's your own money,
1885
00:53:56,860 --> 00:53:57,600
do what you want.
1886
00:53:58,020 --> 00:53:59,441
But maybe m,
1887
00:53:59,540 --> 00:54:01,380
wait a little bit of time
1888
00:54:01,440 --> 00:54:02,820
before you put in the pre
1889
00:54:02,900 --> 00:54:04,760
order because you might be kissing
1890
00:54:04,820 --> 00:54:05,940
that money, uh,
1891
00:54:06,120 --> 00:54:07,360
goodbye if
1892
00:54:07,840 --> 00:54:09,300
the last roadster happens because
1893
00:54:09,780 --> 00:54:12,260
we might be talking in 2036
1894
00:54:12,660 --> 00:54:13,680
and then there'll be some sort
1895
00:54:13,800 --> 00:54:16,740
of roadster V3 that is announced
1896
00:54:17,280 --> 00:54:19,620
that never actually gets shipped and
1897
00:54:19,820 --> 00:54:21,060
we're just giving more money to
1898
00:54:21,120 --> 00:54:21,680
Elon Musk.
1899
00:54:21,740 --> 00:54:24,760
But regardless if this is real,
1900
00:54:25,680 --> 00:54:26,500
and I mean,
1901
00:54:26,620 --> 00:54:27,560
I think that there probably is
1902
00:54:27,600 --> 00:54:28,280
a proof of concept.
1903
00:54:28,340 --> 00:54:28,660
This is
1904
00:54:29,020 --> 00:54:30,460
the coolest car that I've probably
1905
00:54:30,520 --> 00:54:31,780
ever seen in my entire life.
1906
00:54:32,020 --> 00:54:35,460
It is insane. Insane.
1907
00:54:35,520 --> 00:54:35,740
So,
1908
00:54:36,180 --> 00:54:37,100
talking about some of the specs,
1909
00:54:37,340 --> 00:54:38,100
this goes uh,
1910
00:54:38,380 --> 00:54:39,640
0 to 60 in under a
1911
00:54:39,700 --> 00:54:41,320
second, which is um,
1912
00:54:42,440 --> 00:54:43,640
close to half a second or
1913
00:54:43,680 --> 00:54:45,460
something, which is just freaking phenomenal.
1914
00:54:46,000 --> 00:54:46,240
Um,
1915
00:54:46,440 --> 00:54:47,560
it does that through a combination
1916
00:54:47,600 --> 00:54:48,320
of a lot of things.
1917
00:54:48,900 --> 00:54:49,300
Um,
1918
00:54:49,540 --> 00:54:51,740
so one is it has um,
1919
00:54:52,280 --> 00:54:55,260
high pressure air, compressed uh,
1920
00:54:55,620 --> 00:54:56,640
air in large
1921
00:54:57,160 --> 00:54:59,840
freaking tanks that are um,
1922
00:55:01,580 --> 00:55:02,880
just massive.
1923
00:55:02,880 --> 00:55:02,881
Uh,
1924
00:55:02,920 --> 00:55:03,240
it's like
1925
00:55:04,040 --> 00:55:05,240
spread your arms around and try
1926
00:55:05,300 --> 00:55:05,920
to hug.
1927
00:55:06,360 --> 00:55:08,060
And it's as big as that.
1928
00:55:08,700 --> 00:55:10,100
And it has
1929
00:55:10,980 --> 00:55:14,000
that in the trunk which
1930
00:55:14,440 --> 00:55:15,260
fires off
1931
00:55:15,780 --> 00:55:17,120
in order to give it uh,
1932
00:55:17,420 --> 00:55:18,640
thrust and
1933
00:55:19,040 --> 00:55:20,680
push it to reach that
1934
00:55:21,720 --> 00:55:24,700
close to half a second, ah,
1935
00:55:24,840 --> 00:55:27,020
or sub 1 second 0 to
1936
00:55:27,220 --> 00:55:27,400
60.
1937
00:55:28,100 --> 00:55:28,260
Um,
1938
00:55:28,500 --> 00:55:30,380
and it has a large fan
1939
00:55:30,440 --> 00:55:31,740
that is sucking it
1940
00:55:32,640 --> 00:55:33,940
to the ground to give it
1941
00:55:34,000 --> 00:55:35,860
that downforce so that the tires
1942
00:55:35,900 --> 00:55:37,520
can stick to the ground,
1943
00:55:38,040 --> 00:55:39,700
get enough traction and go fast
1944
00:55:39,760 --> 00:55:40,120
enough.
1945
00:55:40,700 --> 00:55:40,940
Um,
1946
00:55:41,180 --> 00:55:42,480
because there's like a limit to
1947
00:55:42,540 --> 00:55:43,600
how much traction tires,
1948
00:55:44,020 --> 00:55:44,680
tires can get.
1949
00:55:44,740 --> 00:55:45,460
I think it's called like the
1950
00:55:46,000 --> 00:55:47,640
mu factor or something.
1951
00:55:48,200 --> 00:55:50,200
Um, and short of uh,
1952
00:55:50,680 --> 00:55:52,600
material science and finding like a
1953
00:55:52,640 --> 00:55:54,060
new kind of substance and putting
1954
00:55:54,120 --> 00:55:55,640
super glue on the tire, there's,
1955
00:55:55,680 --> 00:55:56,500
there's only so much
1956
00:55:57,500 --> 00:55:58,740
traction you can get because of
1957
00:55:58,800 --> 00:55:58,980
how,
1958
00:55:59,320 --> 00:56:00,680
how much downforce there is until
1959
00:56:00,720 --> 00:56:02,320
it starts uh, skidding.
1960
00:56:02,440 --> 00:56:04,010
So it sucks the air
1961
00:56:04,570 --> 00:56:06,730
from underground to press the car
1962
00:56:08,410 --> 00:56:08,450
uh,
1963
00:56:08,490 --> 00:56:10,930
more towards the ground to get
1964
00:56:10,930 --> 00:56:12,650
more traction on the tires and
1965
00:56:12,650 --> 00:56:14,200
then shoots rockets uh,
1966
00:56:14,410 --> 00:56:16,930
with cold air thrusters uh,
1967
00:56:16,970 --> 00:56:19,010
out the back to shoot it
1968
00:56:19,010 --> 00:56:19,250
as,
1969
00:56:19,250 --> 00:56:20,810
as fast as possible to get
1970
00:56:20,970 --> 00:56:22,820
three GS of force, uh,
1971
00:56:22,820 --> 00:56:24,190
which is I think what uh,
1972
00:56:24,410 --> 00:56:28,660
people in a rocket experience. Crazy.
1973
00:56:29,060 --> 00:56:31,220
Yeah, absolutely insane.
1974
00:56:31,220 --> 00:56:32,260
And I really,
1975
00:56:32,260 --> 00:56:33,860
really want this car to be
1976
00:56:33,860 --> 00:56:34,460
on the road.
1977
00:56:34,460 --> 00:56:35,660
I mean how cool would it
1978
00:56:35,660 --> 00:56:37,320
be to just be uh,
1979
00:56:37,320 --> 00:56:38,780
driving around and seeing this thing
1980
00:56:38,780 --> 00:56:39,260
next to you.
1981
00:56:39,260 --> 00:56:40,780
I mean that would be just
1982
00:56:40,780 --> 00:56:41,380
insane.
1983
00:56:41,460 --> 00:56:42,700
So Elon,
1984
00:56:42,700 --> 00:56:43,860
if you're listening to this,
1985
00:56:43,860 --> 00:56:44,820
and I know you're probably a
1986
00:56:44,820 --> 00:56:45,780
big fan of the pod,
1987
00:56:47,220 --> 00:56:48,940
make sure this actually gets shipped
1988
00:56:48,940 --> 00:56:50,260
because I want to see this
1989
00:56:50,260 --> 00:56:53,140
on the road. It's really exciting.
1990
00:56:54,440 --> 00:56:55,000
This seems like ah,
1991
00:56:55,080 --> 00:56:56,120
the culmination of uh,
1992
00:56:56,280 --> 00:56:58,440
Elon's Mega corporation,
1993
00:56:58,440 --> 00:57:00,360
Tesla and SpaceX coming together.
1994
00:57:01,040 --> 00:57:02,570
Um, a lot of the uh,
1995
00:57:02,570 --> 00:57:03,320
obviously the,
1996
00:57:03,560 --> 00:57:05,760
the rocket thrusters that push the
1997
00:57:05,760 --> 00:57:07,560
car forward and push the car
1998
00:57:07,560 --> 00:57:09,820
up and make it flip, um,
1999
00:57:09,820 --> 00:57:10,840
which I didn't get,
2000
00:57:11,000 --> 00:57:11,740
I didn't uh,
2001
00:57:11,740 --> 00:57:13,080
cover much but like it's able
2002
00:57:13,080 --> 00:57:14,560
to um,
2003
00:57:14,560 --> 00:57:17,040
shoot up with the same um,
2004
00:57:17,040 --> 00:57:19,000
cold air thrusters except you need
2005
00:57:19,000 --> 00:57:19,320
another
2006
00:57:20,230 --> 00:57:21,910
tank that takes the passenger seat
2007
00:57:21,910 --> 00:57:23,590
which turns it into like a
2008
00:57:23,590 --> 00:57:24,470
one passenger car.
2009
00:57:24,470 --> 00:57:26,710
If you add the um,
2010
00:57:27,270 --> 00:57:27,420
rocket uh,
2011
00:57:27,790 --> 00:57:29,310
thrusters that allow it to shoot
2012
00:57:29,310 --> 00:57:29,990
up in the air.
2013
00:57:30,230 --> 00:57:31,910
It has six of those rockets,
2014
00:57:31,990 --> 00:57:33,750
four at each corner and then
2015
00:57:33,750 --> 00:57:35,070
two on the front and two
2016
00:57:35,070 --> 00:57:36,710
in the back to steer.
2017
00:57:37,110 --> 00:57:38,590
And then it's able to shoot
2018
00:57:38,590 --> 00:57:39,630
up straight in the air with
2019
00:57:39,630 --> 00:57:42,470
these rockets, flip upside down,
2020
00:57:44,070 --> 00:57:45,830
thrust it forward and the land,
2021
00:57:46,260 --> 00:57:46,430
uh,
2022
00:57:46,700 --> 00:57:48,380
and it's able to go over
2023
00:57:49,060 --> 00:57:49,460
uh,
2024
00:57:49,460 --> 00:57:51,060
the entire lineup of Tesla uh,
2025
00:57:51,460 --> 00:57:53,300
vehicles, uh, without a ramp.
2026
00:57:53,300 --> 00:57:53,860
It's just, you know,
2027
00:57:53,860 --> 00:57:56,460
it's like a VTOL vertical takeoff
2028
00:57:56,460 --> 00:57:57,120
and liftoff.
2029
00:57:57,120 --> 00:58:00,860
Uh, vehicle, um, just freaking phenomenal.
2030
00:58:01,030 --> 00:58:01,370
Um,
2031
00:58:01,370 --> 00:58:02,860
and they have some actuators that
2032
00:58:02,860 --> 00:58:05,340
they used from the Optimus robot,
2033
00:58:05,450 --> 00:58:05,460
uh,
2034
00:58:05,740 --> 00:58:07,380
to Control the spoiler to make
2035
00:58:07,380 --> 00:58:08,620
it go up and down,
2036
00:58:08,620 --> 00:58:10,660
which has more degrees of freedom
2037
00:58:10,660 --> 00:58:12,820
than any other spoiler in any
2038
00:58:12,820 --> 00:58:14,090
other car today.
2039
00:58:14,630 --> 00:58:14,890
Um,
2040
00:58:14,890 --> 00:58:16,530
I think the only other car
2041
00:58:16,530 --> 00:58:18,170
that has done anything even close
2042
00:58:18,170 --> 00:58:19,850
to this has been,
2043
00:58:20,090 --> 00:58:21,410
I forget how to say it
2044
00:58:21,410 --> 00:58:22,640
but it's some um,
2045
00:58:24,090 --> 00:58:26,260
supercar company that uh,
2046
00:58:26,650 --> 00:58:28,330
added large fans in the back
2047
00:58:28,330 --> 00:58:28,730
to
2048
00:58:29,210 --> 00:58:31,370
suction the car towards the ground.
2049
00:58:31,850 --> 00:58:32,270
Um,
2050
00:58:32,270 --> 00:58:34,210
add some plastic flaps on the
2051
00:58:34,210 --> 00:58:36,170
side so it creates a vacuum
2052
00:58:36,170 --> 00:58:37,330
and is able to hug the
2053
00:58:37,330 --> 00:58:39,730
ground and go really tight around
2054
00:58:39,730 --> 00:58:40,400
corners. Corners.
2055
00:58:40,400 --> 00:58:42,560
It's able to corner faster than
2056
00:58:42,560 --> 00:58:43,680
a Porsche 911.
2057
00:58:44,240 --> 00:58:44,640
Um,
2058
00:58:44,640 --> 00:58:45,720
and this car can do the
2059
00:58:45,720 --> 00:58:46,560
same while being
2060
00:58:47,040 --> 00:58:48,400
twice the weight or something,
2061
00:58:48,400 --> 00:58:49,600
which is insane.
2062
00:58:50,140 --> 00:58:50,540
Um,
2063
00:58:50,540 --> 00:58:53,280
if anyone has ever cornered really
2064
00:58:53,280 --> 00:58:55,330
quickly and uh, felt uh,
2065
00:58:55,440 --> 00:58:56,880
yourself losing traction,
2066
00:58:56,960 --> 00:58:58,080
this will never do that.
2067
00:58:58,240 --> 00:59:00,120
This can do this faster than
2068
00:59:00,120 --> 00:59:01,840
any car that you have while
2069
00:59:01,840 --> 00:59:03,040
being twice as heavy.
2070
00:59:03,280 --> 00:59:05,440
Just phenomenal engineering.
2071
00:59:06,100 --> 00:59:07,170
Um, and we,
2072
00:59:07,170 --> 00:59:08,040
we're seeing a lot of uh,
2073
00:59:08,290 --> 00:59:10,620
memes and AI generated videos on
2074
00:59:10,620 --> 00:59:11,050
um,
2075
00:59:11,050 --> 00:59:12,930
TikTok and Instagram that is showing
2076
00:59:13,170 --> 00:59:14,360
how people uh,
2077
00:59:15,730 --> 00:59:18,330
this made up future early buyer
2078
00:59:18,330 --> 00:59:18,620
of uh,
2079
00:59:18,620 --> 00:59:20,250
this Tesla Roadster being stuck in
2080
00:59:20,250 --> 00:59:22,050
traffic and just say to hell
2081
00:59:22,050 --> 00:59:22,290
with this,
2082
00:59:22,290 --> 00:59:23,770
I'm going to flip over and
2083
00:59:23,770 --> 00:59:24,290
then uh,
2084
00:59:24,290 --> 00:59:25,690
overtake all the other cars and
2085
00:59:25,690 --> 00:59:26,470
then uh,
2086
00:59:26,470 --> 00:59:28,730
escape the cops or take this
2087
00:59:28,730 --> 00:59:29,750
parking spot that uh,
2088
00:59:29,750 --> 00:59:30,610
no one else can access.
2089
00:59:30,690 --> 00:59:33,470
So amazing, uh,
2090
00:59:33,470 --> 00:59:37,340
physics defying car here. That's true.
2091
00:59:37,340 --> 00:59:38,660
I wonder if you could go
2092
00:59:38,660 --> 00:59:39,780
if you own the car,
2093
00:59:39,780 --> 00:59:40,660
if you had like a big
2094
00:59:40,660 --> 00:59:41,620
house you could, I don't know,
2095
00:59:41,620 --> 00:59:42,660
like park it on your roof
2096
00:59:42,660 --> 00:59:43,100
or something.
2097
00:59:43,770 --> 00:59:45,060
Um, it's like, oh,
2098
00:59:45,060 --> 00:59:45,580
let's just
2099
00:59:46,140 --> 00:59:47,180
go put it up here.
2100
00:59:47,180 --> 00:59:48,180
It's like come and get it
2101
00:59:48,180 --> 00:59:48,460
now.
2102
00:59:49,980 --> 00:59:50,510
Well uh,
2103
00:59:50,510 --> 00:59:52,060
I think we mentioned Yann Lecun
2104
00:59:52,060 --> 00:59:52,790
and uh,
2105
00:59:52,790 --> 00:59:54,140
Elon Musk's feud a couple times.
2106
00:59:54,140 --> 00:59:55,700
Yann Lecun was the um,
2107
00:59:57,110 --> 00:59:58,530
head of AI at Facebook, uh,
2108
00:59:59,550 --> 00:59:59,910
uh,
2109
00:59:59,910 --> 01:00:02,850
slash meta and he's been uh,
2110
01:00:02,870 --> 01:00:05,030
complaining about all of Elon's uh,
2111
01:00:05,030 --> 01:00:06,750
products even though he owns two
2112
01:00:06,750 --> 01:00:07,190
Teslas.
2113
01:00:07,190 --> 01:00:07,910
He's just not a big fan
2114
01:00:07,910 --> 01:00:09,030
of Elon's politics.
2115
01:00:09,360 --> 01:00:09,900
Um,
2116
01:00:09,990 --> 01:00:11,470
but he did point out that
2117
01:00:11,470 --> 01:00:13,070
this can only last for about
2118
01:00:13,070 --> 01:00:15,430
10 seconds with these thrusters because
2119
01:00:15,510 --> 01:00:19,030
you need these uh, air chambers,
2120
01:00:19,030 --> 01:00:19,390
these,
2121
01:00:19,390 --> 01:00:21,070
this compressed air chamber to be
2122
01:00:21,070 --> 01:00:22,310
at maximum capacity.
2123
01:00:22,840 --> 01:00:23,220
Um,
2124
01:00:23,270 --> 01:00:24,630
it's not like a
2125
01:00:25,110 --> 01:00:26,670
battery which you can drain it
2126
01:00:26,670 --> 01:00:28,430
down to 0% or actually even
2127
01:00:28,430 --> 01:00:29,870
batteries don't quite work like that
2128
01:00:29,870 --> 01:00:31,590
in the material science there's always
2129
01:00:31,590 --> 01:00:32,710
a little bit of charge left.
2130
01:00:32,710 --> 01:00:33,580
But um,
2131
01:00:33,750 --> 01:00:36,110
you need these thrusters at maximum
2132
01:00:36,110 --> 01:00:36,630
capacity.
2133
01:00:36,630 --> 01:00:37,750
If you think about it like
2134
01:00:38,230 --> 01:00:38,440
a ah,
2135
01:00:38,550 --> 01:00:39,990
balloon that is really filled.
2136
01:00:40,230 --> 01:00:41,790
If you want to blow a
2137
01:00:41,790 --> 01:00:42,670
strong gust of air,
2138
01:00:42,670 --> 01:00:43,830
you need the balloon to be
2139
01:00:43,830 --> 01:00:44,950
really, really filled.
2140
01:00:45,190 --> 01:00:46,790
Once it starts deflating,
2141
01:00:46,790 --> 01:00:48,150
it doesn't have as much compressed
2142
01:00:48,150 --> 01:00:48,580
air or
2143
01:00:48,970 --> 01:00:50,250
as much force with the air
2144
01:00:50,250 --> 01:00:51,370
coming out of it anymore.
2145
01:00:51,530 --> 01:00:54,010
And so it can realistically only
2146
01:00:54,090 --> 01:00:56,170
last for about 10 seconds before
2147
01:00:56,170 --> 01:00:58,440
it runs out of its, uh,
2148
01:00:58,440 --> 01:01:03,530
flying, uh, juice. Okay. Freaking phenomenal.
2149
01:01:03,850 --> 01:01:04,929
Yeah. All right.
2150
01:01:04,929 --> 01:01:06,050
There's always going to be the
2151
01:01:06,050 --> 01:01:06,650
haters,
2152
01:01:06,650 --> 01:01:07,050
but
2153
01:01:07,770 --> 01:01:09,130
I don't see you making any
2154
01:01:09,130 --> 01:01:09,930
flying cars,
2155
01:01:11,850 --> 01:01:13,130
so why don't you shut your
2156
01:01:13,130 --> 01:01:14,700
mouth and, uh,
2157
01:01:15,240 --> 01:01:16,600
you had a flying car
2158
01:01:17,000 --> 01:01:19,800
that would last like an hour.
2159
01:01:20,440 --> 01:01:20,840
Great,
2160
01:01:21,080 --> 01:01:22,640
then I would agree with your
2161
01:01:22,640 --> 01:01:23,160
opinion.
2162
01:01:23,240 --> 01:01:25,360
But if you're not making flying
2163
01:01:25,360 --> 01:01:26,920
cars and just commenting on people,
2164
01:01:26,920 --> 01:01:28,360
I have no respect for you.
2165
01:01:28,740 --> 01:01:29,440
Uh, I mean,
2166
01:01:29,440 --> 01:01:30,600
I have a lot of respect
2167
01:01:30,600 --> 01:01:31,520
for John McCune,
2168
01:01:31,520 --> 01:01:33,400
but for that comment. Come on.
2169
01:01:34,120 --> 01:01:35,480
Uh, be kind.
2170
01:01:35,720 --> 01:01:36,720
If people are trying to do
2171
01:01:36,720 --> 01:01:37,160
new things,
2172
01:01:37,160 --> 01:01:38,360
I think that we should show
2173
01:01:38,360 --> 01:01:40,480
them some grace. Yeah.
2174
01:01:40,480 --> 01:01:41,080
This is just
2175
01:01:41,850 --> 01:01:43,050
freaking Elon at his best.
2176
01:01:43,240 --> 01:01:43,460
Um,
2177
01:01:43,770 --> 01:01:44,770
and it seems like he just
2178
01:01:44,770 --> 01:01:47,600
did this to answer, uh,
2179
01:01:48,050 --> 01:01:49,570
back to Peter Thiel over a
2180
01:01:49,570 --> 01:01:50,610
feud where he was like,
2181
01:01:50,610 --> 01:01:51,760
you know, m.
2182
01:01:51,850 --> 01:01:53,330
People dreamt about the future and
2183
01:01:53,330 --> 01:01:54,570
thought we'd have flying cars,
2184
01:01:54,570 --> 01:01:55,930
but we ended up with 140
2185
01:01:55,930 --> 01:01:56,570
characters.
2186
01:01:56,730 --> 01:01:58,210
And the Elon didn't want to
2187
01:01:58,210 --> 01:01:58,970
just let that slide.
2188
01:01:58,970 --> 01:02:00,810
And he's like, you know, the.
2189
01:02:01,850 --> 01:02:02,890
I'm gonna shut you up and
2190
01:02:02,890 --> 01:02:03,970
just make a flying car just
2191
01:02:03,970 --> 01:02:04,970
to prove you wrong.
2192
01:02:06,250 --> 01:02:06,770
I mean,
2193
01:02:06,770 --> 01:02:08,130
it really does feel like the
2194
01:02:08,130 --> 01:02:08,810
future. Right.
2195
01:02:09,080 --> 01:02:09,570
Um, actually,
2196
01:02:09,570 --> 01:02:11,260
I read a book about like,
2197
01:02:11,260 --> 01:02:12,540
you know, where's my flying car?
2198
01:02:12,540 --> 01:02:13,500
And I think that there are
2199
01:02:13,500 --> 01:02:14,860
a lot of issues with like,
2200
01:02:14,860 --> 01:02:15,820
real flying cars.
2201
01:02:15,820 --> 01:02:17,220
Because just as like,
2202
01:02:17,300 --> 01:02:20,480
I don't necessarily want people, um,
2203
01:02:20,480 --> 01:02:21,600
driving, um,
2204
01:02:21,600 --> 01:02:22,580
all the time because there's a
2205
01:02:22,580 --> 01:02:23,300
lot of bad drivers.
2206
01:02:23,460 --> 01:02:25,380
Imagine if everybody was a pilot.
2207
01:02:25,540 --> 01:02:25,760
Um,
2208
01:02:25,760 --> 01:02:26,820
I think that would be
2209
01:02:27,220 --> 01:02:28,260
potentially a very,
2210
01:02:28,420 --> 01:02:29,220
A little bit scary.
2211
01:02:29,220 --> 01:02:29,500
I mean,
2212
01:02:29,500 --> 01:02:30,220
you have to worry about like,
2213
01:02:30,220 --> 01:02:32,480
air traffic control and, um,
2214
01:02:32,480 --> 01:02:33,540
you don't have to worry about
2215
01:02:33,620 --> 01:02:35,380
thousand ton machines flying,
2216
01:02:35,380 --> 01:02:36,305
like falling out of the sky.
2217
01:02:36,305 --> 01:02:37,500
The sky and falling on you.
2218
01:02:38,140 --> 01:02:39,660
Yeah, exactly.
2219
01:02:39,660 --> 01:02:41,380
So I, I don't think,
2220
01:02:41,380 --> 01:02:42,300
like practically
2221
01:02:43,100 --> 01:02:44,340
we should have a lot of
2222
01:02:44,340 --> 01:02:45,260
flying cars.
2223
01:02:45,380 --> 01:02:45,640
Um,
2224
01:02:45,660 --> 01:02:46,940
maybe if they were self driving,
2225
01:02:46,940 --> 01:02:48,540
that'd be more okay.
2226
01:02:49,100 --> 01:02:50,220
But again,
2227
01:02:50,620 --> 01:02:52,659
the fact that there is actually
2228
01:02:52,659 --> 01:02:55,500
like a legitimate automotive company
2229
01:02:56,700 --> 01:02:59,500
announcing a true flying car that
2230
01:02:59,500 --> 01:03:01,180
isn't some sort of like a
2231
01:03:01,180 --> 01:03:02,060
random startup,
2232
01:03:02,500 --> 01:03:03,340
that is exciting to me.
2233
01:03:03,340 --> 01:03:03,940
I mean, like,
2234
01:03:04,740 --> 01:03:06,740
if it was Honda or Toyota
2235
01:03:06,740 --> 01:03:08,340
or like Ford that announced the
2236
01:03:08,340 --> 01:03:08,820
flying car,
2237
01:03:08,820 --> 01:03:10,220
I'd be a little bit more
2238
01:03:10,220 --> 01:03:11,740
confident that would actually release than
2239
01:03:11,740 --> 01:03:12,140
Tesla.
2240
01:03:12,140 --> 01:03:14,260
But Tesla actually is a real
2241
01:03:14,260 --> 01:03:14,660
company.
2242
01:03:15,140 --> 01:03:15,940
That is public,
2243
01:03:16,260 --> 01:03:18,860
that is shipping real cars to
2244
01:03:18,860 --> 01:03:20,260
real people that are driving on
2245
01:03:20,260 --> 01:03:20,980
the roads.
2246
01:03:20,980 --> 01:03:22,300
And the fact that they said
2247
01:03:22,300 --> 01:03:23,740
it gives it a lot more
2248
01:03:23,740 --> 01:03:25,700
credibility than if it's some sort
2249
01:03:25,700 --> 01:03:27,620
of concept car that a startup
2250
01:03:27,620 --> 01:03:28,970
that never had real least anything
2251
01:03:29,210 --> 01:03:30,410
would uh, have come out with.
2252
01:03:31,040 --> 01:03:34,170
Um, anyway, yeah, uh, moving on.
2253
01:03:34,720 --> 01:03:36,770
Um, tangentially related to AI,
2254
01:03:36,770 --> 01:03:38,090
not really, but it's, it's,
2255
01:03:38,090 --> 01:03:39,650
it was too exciting to pass
2256
01:03:39,650 --> 01:03:40,010
up.
2257
01:03:40,660 --> 01:03:40,790
Uh,
2258
01:03:40,970 --> 01:03:42,370
so I wanted to talk a
2259
01:03:42,370 --> 01:03:44,400
little bit about the anthropic um,
2260
01:03:45,170 --> 01:03:47,530
IPO and a little bit of
2261
01:03:49,610 --> 01:03:49,660
kind ah,
2262
01:03:49,890 --> 01:03:52,610
of controversy around whether or not
2263
01:03:52,610 --> 01:03:53,710
they are profitable.
2264
01:03:54,020 --> 01:03:56,020
So um, again I'm not.
2265
01:03:57,460 --> 01:03:59,060
Yeah, and they are,
2266
01:03:59,060 --> 01:04:00,820
they're claiming to be profitable.
2267
01:04:00,980 --> 01:04:03,600
And apparently that is on an
2268
01:04:03,680 --> 01:04:05,240
adjusted basis.
2269
01:04:05,660 --> 01:04:07,380
So I guess the word adjusted
2270
01:04:07,460 --> 01:04:08,980
there is doing a lot of
2271
01:04:09,720 --> 01:04:10,960
heavy listing lifting.
2272
01:04:11,400 --> 01:04:12,500
So um,
2273
01:04:12,780 --> 01:04:13,760
now I want to talk a
2274
01:04:13,840 --> 01:04:16,940
little bit about uh, GAAP accounting.
2275
01:04:17,800 --> 01:04:19,200
So GAAP
2276
01:04:19,700 --> 01:04:20,860
or in the United States the
2277
01:04:21,200 --> 01:04:23,760
Generally Accepted Accounting Principles
2278
01:04:24,420 --> 01:04:26,840
is a set of rules that
2279
01:04:26,960 --> 01:04:29,520
all companies uh, have to follow,
2280
01:04:30,000 --> 01:04:32,800
um, for how to uh,
2281
01:04:33,160 --> 01:04:34,720
present their financials.
2282
01:04:34,820 --> 01:04:36,480
Because while I think that
2283
01:04:36,900 --> 01:04:37,580
for a lot of us we
2284
01:04:37,640 --> 01:04:38,820
know that if a company is
2285
01:04:38,860 --> 01:04:39,320
profitable,
2286
01:04:39,380 --> 01:04:39,960
you know that they
2287
01:04:40,560 --> 01:04:41,920
make more money than they spend,
2288
01:04:42,440 --> 01:04:44,360
but it's oftentimes not always that
2289
01:04:44,460 --> 01:04:45,900
easy to say like if a
2290
01:04:45,960 --> 01:04:47,520
company is profitable, profitable.
2291
01:04:47,580 --> 01:04:49,220
So I'll give one example before
2292
01:04:49,240 --> 01:04:50,680
we talk about anthropics numbers.
2293
01:04:51,220 --> 01:04:55,120
So one concept that exists uh,
2294
01:04:55,320 --> 01:04:56,500
for a lot of companies is
2295
01:04:56,560 --> 01:04:59,620
the concept of depreciation where um,
2296
01:05:00,420 --> 01:05:01,940
if you spend money today,
2297
01:05:02,700 --> 01:05:03,860
the thing that you spend the
2298
01:05:03,900 --> 01:05:05,920
money on will have a useful
2299
01:05:06,080 --> 01:05:07,300
life over time.
2300
01:05:07,820 --> 01:05:09,640
So even though you might be
2301
01:05:10,100 --> 01:05:12,080
down a certain amount of money
2302
01:05:12,140 --> 01:05:12,580
today,
2303
01:05:13,500 --> 01:05:14,680
if you're able to use that
2304
01:05:14,800 --> 01:05:15,140
over,
2305
01:05:15,220 --> 01:05:16,780
let's say a 10 year useful
2306
01:05:16,860 --> 01:05:17,760
life, um,
2307
01:05:18,160 --> 01:05:19,540
you may not say it as
2308
01:05:19,620 --> 01:05:20,440
a complete loss.
2309
01:05:20,720 --> 01:05:22,540
For example, if I, um,
2310
01:05:22,660 --> 01:05:24,060
let's say I am a taxi
2311
01:05:24,100 --> 01:05:25,300
company, um,
2312
01:05:25,440 --> 01:05:26,520
and I know that I'm going
2313
01:05:26,560 --> 01:05:28,060
to buy a car and that
2314
01:05:28,140 --> 01:05:29,140
car is going to cost,
2315
01:05:29,200 --> 01:05:30,780
let's say to make numbers easy,
2316
01:05:30,860 --> 01:05:31,840
$10,000.
2317
01:05:32,800 --> 01:05:34,320
But I'm able to drive that
2318
01:05:34,380 --> 01:05:35,880
car for 10 years.
2319
01:05:36,580 --> 01:05:37,520
So um,
2320
01:05:37,860 --> 01:05:39,360
I might be able to depreciate
2321
01:05:39,420 --> 01:05:43,480
that car $1,000 every 10 years.
2322
01:05:43,700 --> 01:05:45,100
So if I'm going to record
2323
01:05:45,140 --> 01:05:45,920
my profitability,
2324
01:05:46,400 --> 01:05:47,580
I might not say that, oh,
2325
01:05:47,680 --> 01:05:50,540
I lost $10,000 on year one.
2326
01:05:51,540 --> 01:05:53,340
I would say that I only
2327
01:05:53,400 --> 01:05:55,860
lost $1,000 on year one.
2328
01:05:55,880 --> 01:05:57,560
And I would say that um,
2329
01:05:58,060 --> 01:06:00,520
I am spending $1,000
2330
01:06:01,680 --> 01:06:02,480
over 10 years.
2331
01:06:02,560 --> 01:06:05,200
So minus $1,000 every single year.
2332
01:06:05,800 --> 01:06:06,220
And
2333
01:06:06,700 --> 01:06:08,580
if we do that then um,
2334
01:06:08,840 --> 01:06:10,920
I've only actually spent um,
2335
01:06:11,080 --> 01:06:12,060
a thousand dollars this year.
2336
01:06:12,260 --> 01:06:13,640
Even Though in my bank account
2337
01:06:13,700 --> 01:06:17,060
balance, um, I'm actually minus $10,000.
2338
01:06:17,220 --> 01:06:17,400
But
2339
01:06:18,180 --> 01:06:19,020
since I'm spreading this out over
2340
01:06:19,060 --> 01:06:22,560
10 years, we're actually maybe profitable.
2341
01:06:23,040 --> 01:06:24,360
So let's say the
2342
01:06:24,760 --> 01:06:27,480
hypothetical taxi company, uh,
2343
01:06:28,620 --> 01:06:31,260
made $2,000 this year.
2344
01:06:31,920 --> 01:06:32,160
Um,
2345
01:06:32,500 --> 01:06:33,660
like from all the rides it
2346
01:06:33,920 --> 01:06:35,020
made $2,000.
2347
01:06:35,280 --> 01:06:36,740
But actually in its bank account
2348
01:06:38,480 --> 01:06:39,740
it's minus $10,000.
2349
01:06:39,800 --> 01:06:40,420
Maybe they even had to get
2350
01:06:40,480 --> 01:06:41,740
a loan or whatever to spend
2351
01:06:41,760 --> 01:06:43,240
that $10,000 to buy that taxi.
2352
01:06:43,800 --> 01:06:46,000
But for investors,
2353
01:06:46,280 --> 01:06:47,580
they would say that actually
2354
01:06:48,080 --> 01:06:50,760
we're profitable $1000 because really
2355
01:06:51,400 --> 01:06:52,540
that car has a 10 year
2356
01:06:52,600 --> 01:06:53,240
useful life.
2357
01:06:53,740 --> 01:06:55,840
And since we made $2000
2358
01:06:56,400 --> 01:06:56,520
but
2359
01:06:56,900 --> 01:07:00,480
only spent $1000 on the car,
2360
01:07:00,840 --> 01:07:03,200
we're profiting 1000.
2361
01:07:03,680 --> 01:07:04,601
Even though our, uh,
2362
01:07:04,640 --> 01:07:08,460
bank account balance is minus, uh,
2363
01:07:08,700 --> 01:07:10,680
$8,000. Does that make sense?
2364
01:07:12,380 --> 01:07:13,260
Okay, so,
2365
01:07:13,320 --> 01:07:13,860
so that's like
2366
01:07:14,380 --> 01:07:16,060
one of the things where judgment
2367
01:07:16,220 --> 01:07:17,320
is being used.
2368
01:07:17,480 --> 01:07:18,360
So there's a lot of things
2369
01:07:18,520 --> 01:07:20,880
like this where a company can
2370
01:07:20,960 --> 01:07:23,120
claim to be profitable even though
2371
01:07:23,200 --> 01:07:24,920
they're losing money on a free
2372
01:07:25,000 --> 01:07:25,720
cash flow basis.
2373
01:07:25,760 --> 01:07:27,200
So a free cash flow basis
2374
01:07:27,580 --> 01:07:29,320
is loosely like how much money
2375
01:07:29,380 --> 01:07:31,801
does the um, business, uh,
2376
01:07:31,840 --> 01:07:32,540
have in the bank?
2377
01:07:33,060 --> 01:07:34,840
So if, uh,
2378
01:07:34,960 --> 01:07:36,540
or like have available to spend,
2379
01:07:36,600 --> 01:07:38,160
it's kind of like excess profits.
2380
01:07:38,760 --> 01:07:41,160
Um, so you can, uh,
2381
01:07:41,340 --> 01:07:42,400
so some people call this like
2382
01:07:42,500 --> 01:07:45,440
ebitda, like earnings before interest, taxes,
2383
01:07:45,500 --> 01:07:48,560
depreciation, depreciation and amortization.
2384
01:07:48,620 --> 01:07:49,560
Just figure out like how much
2385
01:07:49,600 --> 01:07:50,980
money they actually have.
2386
01:07:51,440 --> 01:07:52,340
Um, but like,
2387
01:07:52,740 --> 01:07:55,820
if you have things like depreciation
2388
01:07:55,920 --> 01:07:57,260
or interest or whatever,
2389
01:07:57,460 --> 01:07:58,800
you can make it look that
2390
01:07:58,840 --> 01:07:59,280
you're like,
2391
01:07:59,360 --> 01:08:00,760
you're more profitable than you actually
2392
01:08:00,840 --> 01:08:01,060
are.
2393
01:08:01,660 --> 01:08:03,260
So with all that being said,
2394
01:08:03,270 --> 01:08:05,350
there's a lot of judgments involved.
2395
01:08:05,350 --> 01:08:06,230
So, you know,
2396
01:08:06,230 --> 01:08:07,750
in the example I said that
2397
01:08:07,750 --> 01:08:10,220
a car might have a, uh,
2398
01:08:10,230 --> 01:08:12,510
depreciation schedule of 10 years where
2399
01:08:12,510 --> 01:08:13,350
you're depreciating,
2400
01:08:13,670 --> 01:08:15,150
writing off a thousand dollars every
2401
01:08:15,150 --> 01:08:15,510
year.
2402
01:08:15,750 --> 01:08:16,890
But, uh,
2403
01:08:16,890 --> 01:08:18,550
your profitability would look a lot
2404
01:08:18,550 --> 01:08:18,870
different
2405
01:08:19,290 --> 01:08:20,490
if that was all written off
2406
01:08:20,490 --> 01:08:21,610
on year one to say, oh,
2407
01:08:21,610 --> 01:08:21,730
well,
2408
01:08:21,730 --> 01:08:23,010
you can't depreciate the use of
2409
01:08:23,010 --> 01:08:23,930
life of a car at all.
2410
01:08:24,090 --> 01:08:24,250
If, uh,
2411
01:08:24,370 --> 01:08:25,170
it's just like one year,
2412
01:08:25,170 --> 01:08:26,410
you have to write off all
2413
01:08:26,490 --> 01:08:28,090
$10,000 all at once.
2414
01:08:28,090 --> 01:08:29,330
And then it would actually look
2415
01:08:29,330 --> 01:08:31,450
like you're losing $8,000 in our
2416
01:08:31,690 --> 01:08:35,090
hypothetical $2,000 of revenue example for
2417
01:08:35,090 --> 01:08:36,650
this hypothetical taxi company.
2418
01:08:37,050 --> 01:08:37,850
And uh,
2419
01:08:37,850 --> 01:08:38,810
it would look different if it's
2420
01:08:38,810 --> 01:08:40,330
five years, three years, six years,
2421
01:08:40,330 --> 01:08:42,490
whatever. Hope, um, that makes sense.
2422
01:08:42,490 --> 01:08:43,410
I know that was a lot
2423
01:08:43,410 --> 01:08:45,050
of numbers, but regardless,
2424
01:08:45,229 --> 01:08:45,669
the point is,
2425
01:08:45,669 --> 01:08:47,629
is you can mess with profitability
2426
01:08:47,709 --> 01:08:48,989
in different ways.
2427
01:08:49,389 --> 01:08:52,220
So anthropic is uh,
2428
01:08:52,509 --> 01:08:55,229
claiming to be Profitable to around
2429
01:08:55,309 --> 01:08:57,709
the tune of their total revenue
2430
01:08:57,709 --> 01:08:59,480
being uh,
2431
01:08:59,480 --> 01:09:02,669
about $16.8 billion
2432
01:09:03,109 --> 01:09:04,751
as they mentioned on their October
2433
01:09:04,867 --> 01:09:06,829
14th investment.
2434
01:09:07,980 --> 01:09:10,400
Uh uh, and uh,
2435
01:09:10,920 --> 01:09:11,880
at their October
2436
01:09:12,400 --> 01:09:15,040
uh October 14th investor meeting and
2437
01:09:15,040 --> 01:09:16,480
they mentioned that they were free
2438
01:09:16,480 --> 01:09:19,240
cash flow positive with an 11%
2439
01:09:19,560 --> 01:09:20,680
operating margin.
2440
01:09:20,840 --> 01:09:23,480
So I believe 11% of 16.8
2441
01:09:23,480 --> 01:09:24,360
billion is
2442
01:09:24,840 --> 01:09:28,120
around like 1.6 billion ish would
2443
01:09:28,120 --> 01:09:28,680
be their
2444
01:09:29,040 --> 01:09:31,520
uh actual like profit uh if
2445
01:09:31,520 --> 01:09:34,080
you look at that from their
2446
01:09:34,080 --> 01:09:35,640
gross uh revenue.
2447
01:09:37,170 --> 01:09:38,180
But um,
2448
01:09:38,180 --> 01:09:39,130
a lot of the things that
2449
01:09:39,130 --> 01:09:40,930
people are pointing to is that
2450
01:09:41,330 --> 01:09:43,970
this is excluding a lot of
2451
01:09:43,970 --> 01:09:44,290
things.
2452
01:09:44,610 --> 01:09:45,240
So uh,
2453
01:09:45,240 --> 01:09:46,770
people are mentioning that this is
2454
01:09:47,170 --> 01:09:50,770
excluding some stock based compensation where
2455
01:09:51,260 --> 01:09:51,920
um,
2456
01:09:52,610 --> 01:09:55,810
even though Anthropic is giving stock
2457
01:09:56,130 --> 01:09:58,370
options um to people
2458
01:09:58,820 --> 01:10:01,890
uh oftentimes like that stock isn't
2459
01:10:01,890 --> 01:10:02,530
actual
2460
01:10:03,780 --> 01:10:05,380
money out but it's more of
2461
01:10:05,380 --> 01:10:06,260
like a paper
2462
01:10:06,720 --> 01:10:07,900
um thing because it's like if
2463
01:10:07,900 --> 01:10:08,860
I own Anthropic,
2464
01:10:08,860 --> 01:10:11,220
well they are Anthropic so they
2465
01:10:11,220 --> 01:10:12,540
can essentially give away pieces of
2466
01:10:12,540 --> 01:10:12,980
themselves.
2467
01:10:12,980 --> 01:10:13,600
So uh,
2468
01:10:13,600 --> 01:10:15,500
people aren't including that even though
2469
01:10:15,500 --> 01:10:17,460
that will be a future hit
2470
01:10:17,540 --> 01:10:17,940
on
2471
01:10:18,520 --> 01:10:21,140
um total profitability later because once
2472
01:10:21,140 --> 01:10:22,100
they do go public
2473
01:10:22,600 --> 01:10:23,540
um the
2474
01:10:24,260 --> 01:10:26,740
employee compensation will be an actual
2475
01:10:26,740 --> 01:10:28,900
outflow that they could have kept.
2476
01:10:29,260 --> 01:10:31,440
Um also uh,
2477
01:10:31,450 --> 01:10:33,530
a lot of the revenue booked
2478
01:10:33,900 --> 01:10:36,570
uh is gross not net meaning
2479
01:10:36,570 --> 01:10:37,930
like they're talking about the top
2480
01:10:37,930 --> 01:10:39,450
line revenue not the bottom line.
2481
01:10:39,610 --> 01:10:41,250
So I guess Claude is sold
2482
01:10:41,250 --> 01:10:42,410
through a lot of cloud partners
2483
01:10:42,410 --> 01:10:44,650
and the partner share is recorded
2484
01:10:45,370 --> 01:10:47,050
as ah sales and marketing expenses
2485
01:10:47,050 --> 01:10:48,810
not netted out of revenue.
2486
01:10:48,890 --> 01:10:50,610
And the analysts argue that this
2487
01:10:50,610 --> 01:10:52,170
makes the top line look larger
2488
01:10:52,170 --> 01:10:54,570
than our competitors which report the
2489
01:10:54,790 --> 01:10:56,630
net revenue from the cloud partners.
2490
01:10:56,710 --> 01:10:57,760
So uh,
2491
01:10:57,760 --> 01:11:00,300
what that means is um,
2492
01:11:00,950 --> 01:11:01,350
if
2493
01:11:01,940 --> 01:11:05,070
uh Anthropic is being sold through
2494
01:11:05,070 --> 01:11:06,790
their partners such as
2495
01:11:07,750 --> 01:11:11,250
AWS like Amazon's Bedrock or Microsoft,
2496
01:11:11,250 --> 01:11:13,170
uh Azure, um,
2497
01:11:13,170 --> 01:11:16,950
whatever uh both Amazon and
2498
01:11:18,470 --> 01:11:18,530
Microsoft uh
2499
01:11:19,430 --> 01:11:20,470
want to make a cut
2500
01:11:20,990 --> 01:11:22,270
from that because they are doing
2501
01:11:22,270 --> 01:11:23,230
a lot of the hosting.
2502
01:11:23,470 --> 01:11:25,480
But apparently uh
2503
01:11:26,110 --> 01:11:28,150
Anthropic in some of their numbers
2504
01:11:28,150 --> 01:11:31,880
aren't necessarily ah always um
2505
01:11:32,950 --> 01:11:33,870
uh making that
2506
01:11:34,430 --> 01:11:36,950
cut uh that Amazon or Amazon
2507
01:11:36,950 --> 01:11:39,390
or Microsoft is taking
2508
01:11:39,790 --> 01:11:41,990
super clear and they're reporting the
2509
01:11:41,990 --> 01:11:43,710
top line and not the actual
2510
01:11:43,710 --> 01:11:44,990
profit they're making from that.
2511
01:11:45,070 --> 01:11:46,950
I don't know exactly but this
2512
01:11:46,950 --> 01:11:48,310
is what some people have been
2513
01:11:48,830 --> 01:11:50,030
um reporting.
2514
01:11:50,030 --> 01:11:51,240
And then uh,
2515
01:11:51,240 --> 01:11:52,390
last but not least
2516
01:11:53,190 --> 01:11:54,950
a lot of people are saying
2517
01:11:54,950 --> 01:11:58,310
that the uh compute costs
2518
01:11:58,710 --> 01:12:01,990
are uh potentially being understated
2519
01:12:02,710 --> 01:12:03,710
going forward.
2520
01:12:03,790 --> 01:12:06,750
So Anthropic has a lot of
2521
01:12:07,170 --> 01:12:09,610
compute that they are spending.
2522
01:12:09,690 --> 01:12:10,590
They're doing some
2523
01:12:11,550 --> 01:12:14,950
recent deals with uh Xai and
2524
01:12:15,190 --> 01:12:17,350
they plan on giving
2525
01:12:18,170 --> 01:12:19,550
uh the new SpaceX
2526
01:12:20,210 --> 01:12:21,550
uh a lot of money to
2527
01:12:21,750 --> 01:12:24,370
use some of their servers
2528
01:12:25,190 --> 01:12:27,770
and they are uh,
2529
01:12:28,030 --> 01:12:32,150
doing some interesting depreciation schedules around
2530
01:12:32,230 --> 01:12:32,570
that.
2531
01:12:33,090 --> 01:12:33,330
Um,
2532
01:12:33,670 --> 01:12:34,230
because
2533
01:12:34,829 --> 01:12:36,490
a lot of this money isn't
2534
01:12:36,550 --> 01:12:39,310
necessarily money that is being spent
2535
01:12:39,370 --> 01:12:40,050
right now,
2536
01:12:40,550 --> 01:12:42,630
but this is money that Anthropic
2537
01:12:42,710 --> 01:12:43,990
plans on spending later.
2538
01:12:44,330 --> 01:12:46,171
So basically uh,
2539
01:12:46,190 --> 01:12:47,310
what that all is to say
2540
01:12:47,390 --> 01:12:49,790
is that Anthropic is,
2541
01:12:50,590 --> 01:12:53,270
it seems to be potentially um,
2542
01:12:53,730 --> 01:12:56,130
doing some adjustments to.
2543
01:12:56,550 --> 01:12:57,210
Some would say,
2544
01:12:57,790 --> 01:13:00,450
I don't know exactly, but uh,
2545
01:13:01,090 --> 01:13:05,430
they're claiming a $16.8 billion uh,
2546
01:13:05,890 --> 01:13:07,750
like net revenue with that is
2547
01:13:07,790 --> 01:13:08,310
profitable.
2548
01:13:08,790 --> 01:13:11,750
But under some regular general accounting
2549
01:13:11,790 --> 01:13:13,671
principles you uh,
2550
01:13:13,710 --> 01:13:14,370
might be able to make the
2551
01:13:14,450 --> 01:13:16,270
argument that they're actually having a
2552
01:13:16,470 --> 01:13:18,130
42 billion dollar loss.
2553
01:13:18,790 --> 01:13:18,910
Um,
2554
01:13:19,030 --> 01:13:20,710
so lots of different ways to
2555
01:13:20,750 --> 01:13:21,150
say that.
2556
01:13:21,310 --> 01:13:22,530
But regardless
2557
01:13:24,210 --> 01:13:25,910
there is a slight
2558
01:13:26,730 --> 01:13:30,130
difference between $16.8 billion profit and
2559
01:13:30,230 --> 01:13:31,930
ah, $42 billion loss.
2560
01:13:32,190 --> 01:13:32,470
Um,
2561
01:13:33,110 --> 01:13:35,330
I think those numbers aren't the
2562
01:13:35,390 --> 01:13:35,710
same.
2563
01:13:36,170 --> 01:13:36,390
Um,
2564
01:13:36,650 --> 01:13:36,910
so
2565
01:13:37,410 --> 01:13:39,750
whether or not Anthropic is profitable,
2566
01:13:40,330 --> 01:13:41,170
there's a lot of ways they
2567
01:13:41,210 --> 01:13:42,990
can try to show profitability.
2568
01:13:43,430 --> 01:13:43,570
Um,
2569
01:13:43,930 --> 01:13:46,170
but the fact that they are,
2570
01:13:46,590 --> 01:13:48,110
my understanding is they're still raising
2571
01:13:48,150 --> 01:13:48,370
money
2572
01:13:48,850 --> 01:13:49,790
and the fact that they are
2573
01:13:49,850 --> 01:13:51,170
looking to the public markets to
2574
01:13:51,230 --> 01:13:53,290
raise even more money, um,
2575
01:13:54,730 --> 01:13:56,110
makes me think that they may
2576
01:13:56,210 --> 01:13:57,850
actually not be quite as profitable
2577
01:13:58,030 --> 01:13:58,790
as they claim.
2578
01:13:59,250 --> 01:14:00,610
However, uh,
2579
01:14:02,850 --> 01:14:04,770
Anthropic is still making the best
2580
01:14:04,830 --> 01:14:05,170
model.
2581
01:14:05,450 --> 01:14:08,170
But whether or not this
2582
01:14:08,650 --> 01:14:09,650
actually is going to be a
2583
01:14:09,670 --> 01:14:11,590
$2 trillion IPO
2584
01:14:12,129 --> 01:14:13,230
and if they will,
2585
01:14:13,530 --> 01:14:13,930
if the
2586
01:14:14,710 --> 01:14:16,510
public investors will actually give them
2587
01:14:16,550 --> 01:14:16,910
much money,
2588
01:14:17,050 --> 01:14:19,270
I think is very much uh,
2589
01:14:19,590 --> 01:14:21,030
a wait and see type thing.
2590
01:14:21,290 --> 01:14:24,370
And I hope Anthropic does well,
2591
01:14:24,710 --> 01:14:24,950
um,
2592
01:14:25,190 --> 01:14:26,950
because I don't want to see
2593
01:14:27,010 --> 01:14:28,030
companies fail.
2594
01:14:28,230 --> 01:14:28,930
And no matter
2595
01:14:29,370 --> 01:14:31,270
how much I have issues with
2596
01:14:31,350 --> 01:14:32,310
some of the uh,
2597
01:14:33,110 --> 01:14:34,250
things that Anthropic is doing,
2598
01:14:34,350 --> 01:14:35,890
I think that they are
2599
01:14:36,970 --> 01:14:38,650
truly pushing AI forward and I
2600
01:14:38,710 --> 01:14:39,990
don't want to see a market
2601
01:14:40,050 --> 01:14:40,490
crash.
2602
01:14:40,950 --> 01:14:42,710
Um, although this uh,
2603
01:14:43,090 --> 01:14:45,570
gives me a cause for concern
2604
01:14:45,670 --> 01:14:48,510
that once Anthropic
2605
01:14:49,150 --> 01:14:51,090
does publish their S1
2606
01:14:52,030 --> 01:14:55,030
and they do have true audited
2607
01:14:55,110 --> 01:14:56,510
GAAP earnings, um,
2608
01:14:56,650 --> 01:14:58,850
right now we don't really know
2609
01:14:58,950 --> 01:15:00,670
if they're truly profitable or not,
2610
01:15:00,850 --> 01:15:02,890
but it seems like they most
2611
01:15:02,930 --> 01:15:04,190
likely are not profitable.
2612
01:15:04,310 --> 01:15:05,590
But we'll have to find out.
2613
01:15:08,430 --> 01:15:09,950
Um, I, I always like uh,
2614
01:15:10,110 --> 01:15:11,930
Mark's skepticism because it brings me
2615
01:15:11,970 --> 01:15:14,250
down from my uh, euphoric, uh,
2616
01:15:14,370 --> 01:15:16,530
excitement for AI and its future.
2617
01:15:17,030 --> 01:15:17,230
Um,
2618
01:15:17,590 --> 01:15:19,410
I agree with
2619
01:15:19,850 --> 01:15:20,510
most of what you said,
2620
01:15:20,590 --> 01:15:22,410
but I was googling, uh,
2621
01:15:22,430 --> 01:15:23,870
and checking um,
2622
01:15:24,070 --> 01:15:26,350
what standard GAAP accounting principles
2623
01:15:26,850 --> 01:15:28,250
mean and how
2624
01:15:28,750 --> 01:15:31,370
stock options and asset depreciations are
2625
01:15:31,430 --> 01:15:31,810
handled.
2626
01:15:31,850 --> 01:15:31,950
And
2627
01:15:32,510 --> 01:15:34,770
it doesn't say that what Anthropic
2628
01:15:34,810 --> 01:15:37,050
is doing is necessarily, um,
2629
01:15:37,770 --> 01:15:39,510
completely out of the ordinary.
2630
01:15:39,550 --> 01:15:40,330
Like uh,
2631
01:15:40,410 --> 01:15:41,790
a lot of asset breakdowns are
2632
01:15:41,850 --> 01:15:43,870
handled like uh, ah,
2633
01:15:43,950 --> 01:15:44,810
one year at a time.
2634
01:15:44,850 --> 01:15:46,810
And then um, uh,
2635
01:15:47,130 --> 01:15:47,410
a lot of
2636
01:15:48,030 --> 01:15:50,010
standard gap accounting principles don't include
2637
01:15:50,050 --> 01:15:50,690
stock options.
2638
01:15:50,750 --> 01:15:51,830
So you know, I,
2639
01:15:52,410 --> 01:15:53,490
I don't think they're trying to
2640
01:15:53,550 --> 01:15:54,770
be, uh, deceptive.
2641
01:15:55,190 --> 01:15:56,410
They might be trying to put
2642
01:15:56,450 --> 01:15:57,770
their best foot forward and
2643
01:15:58,450 --> 01:15:59,810
show uh,
2644
01:16:00,190 --> 01:16:01,470
that they are doing better.
2645
01:16:01,690 --> 01:16:02,150
Which is,
2646
01:16:02,210 --> 01:16:03,270
which is a good sign.
2647
01:16:03,810 --> 01:16:04,930
I mean I think if anyone
2648
01:16:05,850 --> 01:16:06,430
can do,
2649
01:16:07,170 --> 01:16:08,830
can be profitable in this AI
2650
01:16:08,890 --> 01:16:09,050
race,
2651
01:16:09,110 --> 01:16:11,230
it's Anthropic with the ridiculous amounts
2652
01:16:11,590 --> 01:16:13,610
that they're charging for their uh,
2653
01:16:13,770 --> 01:16:14,470
API pricing.
2654
01:16:14,710 --> 01:16:16,810
And it's because they're at the,
2655
01:16:17,370 --> 01:16:18,349
at the forefront.
2656
01:16:18,970 --> 01:16:19,030
I,
2657
01:16:19,090 --> 01:16:19,250
I know
2658
01:16:19,790 --> 01:16:20,490
you and I have had this
2659
01:16:20,530 --> 01:16:22,350
discussion a couple times and uh,
2660
01:16:22,450 --> 01:16:23,430
talked about whether AI is a
2661
01:16:23,470 --> 01:16:24,390
commodity or not.
2662
01:16:24,470 --> 01:16:25,411
And we uh,
2663
01:16:25,430 --> 01:16:26,370
talked about how a lot of
2664
01:16:26,430 --> 01:16:28,590
these cheaper AI, uh, uh,
2665
01:16:28,730 --> 01:16:30,310
models from China are coming out
2666
01:16:30,350 --> 01:16:31,350
and how you can host them
2667
01:16:31,390 --> 01:16:31,930
for free.
2668
01:16:32,470 --> 01:16:32,810
Um,
2669
01:16:32,970 --> 01:16:33,550
you don't need to pay the
2670
01:16:34,070 --> 01:16:35,830
subscription cost or API pressing if
2671
01:16:35,870 --> 01:16:37,310
you have your own local, uh,
2672
01:16:37,311 --> 01:16:37,730
hardware.
2673
01:16:37,750 --> 01:16:38,390
But hey,
2674
01:16:39,490 --> 01:16:40,930
turns out you have two cloud
2675
01:16:41,750 --> 01:16:44,870
Pro subscription or cloud, uh, max,
2676
01:16:44,950 --> 01:16:45,370
whatever.
2677
01:16:46,010 --> 01:16:46,330
Uh,
2678
01:16:46,630 --> 01:16:48,590
and so you have converted to
2679
01:16:49,270 --> 01:16:50,470
the House of Anthropic
2680
01:16:51,270 --> 01:16:51,850
because it is,
2681
01:16:52,270 --> 01:16:53,050
it is the best.
2682
01:16:53,210 --> 01:16:54,150
They are the leader
2683
01:16:54,810 --> 01:16:56,170
and they have the best models.
2684
01:16:56,710 --> 01:16:57,050
And so,
2685
01:16:57,210 --> 01:17:00,150
I mean I am pretty bullish.
2686
01:17:00,790 --> 01:17:01,330
Um,
2687
01:17:02,370 --> 01:17:04,850
I will try to probably participate
2688
01:17:04,910 --> 01:17:06,110
in their IPO if and when
2689
01:17:06,170 --> 01:17:07,870
they decide to ipo.
2690
01:17:08,670 --> 01:17:11,250
Um, and you know, I'm,
2691
01:17:11,550 --> 01:17:12,870
I'm bullish. Anthropic.
2692
01:17:14,290 --> 01:17:16,310
Yeah, so you're, you're right.
2693
01:17:16,570 --> 01:17:17,470
And I know I've talked a
2694
01:17:17,490 --> 01:17:18,710
lot of shit about Anthropic in
2695
01:17:18,750 --> 01:17:20,970
the past, but as of today,
2696
01:17:21,590 --> 01:17:24,790
um, October 20th, I have, uh,
2697
01:17:24,910 --> 01:17:26,350
not one, but two, uh,
2698
01:17:26,570 --> 01:17:27,770
max subscriptions.
2699
01:17:28,810 --> 01:17:29,610
I am paying for two
2700
01:17:29,990 --> 01:17:31,290
$100 subscriptions,
2701
01:17:31,370 --> 01:17:33,990
which I think is uh,
2702
01:17:34,210 --> 01:17:36,690
more cost efficient than maybe someone
2703
01:17:36,730 --> 01:17:38,211
can, uh, comment uh,
2704
01:17:38,270 --> 01:17:39,930
in down below if they have
2705
01:17:39,950 --> 01:17:40,570
any thoughts on this.
2706
01:17:40,630 --> 01:17:41,830
But I feel like a regular
2707
01:17:41,950 --> 01:17:43,370
200 subscription would be better
2708
01:17:43,870 --> 01:17:45,470
than 200 subscriptions.
2709
01:17:45,510 --> 01:17:48,090
But um, so anyways, Black box.
2710
01:17:48,470 --> 01:17:48,730
Yeah.
2711
01:17:49,170 --> 01:17:49,570
The thing is
2712
01:17:49,990 --> 01:17:51,330
there are some people saying
2713
01:17:51,910 --> 01:17:56,231
on the interwebs that you uh,
2714
01:17:56,270 --> 01:17:58,570
don't actually get a 20. So.
2715
01:17:59,190 --> 01:17:59,470
All right,
2716
01:17:59,530 --> 01:18:00,270
let me take a step back.
2717
01:18:00,930 --> 01:18:04,070
Anthropic has three different plans
2718
01:18:04,750 --> 01:18:06,190
for their pricing subscriptions.
2719
01:18:06,490 --> 01:18:07,790
They have the basic,
2720
01:18:08,250 --> 01:18:08,990
I think it's the pro and
2721
01:18:09,010 --> 01:18:09,510
the max.
2722
01:18:09,810 --> 01:18:11,730
The basic is 1x usage.
2723
01:18:12,170 --> 01:18:12,210
Uh,
2724
01:18:12,330 --> 01:18:13,350
this is like their regular base
2725
01:18:13,430 --> 01:18:13,790
usage.
2726
01:18:14,270 --> 01:18:14,430
Um,
2727
01:18:14,590 --> 01:18:15,810
and then they have their
2728
01:18:16,250 --> 01:18:17,910
20 plan, uh,
2729
01:18:18,070 --> 01:18:19,430
or so like the, it's the,
2730
01:18:19,730 --> 01:18:20,790
that's a $20 plan.
2731
01:18:21,290 --> 01:18:22,690
And then they have $100 plan
2732
01:18:22,730 --> 01:18:23,830
which is 5x usage.
2733
01:18:23,910 --> 01:18:24,250
Which is,
2734
01:18:24,470 --> 01:18:25,470
which appears to be
2735
01:18:26,130 --> 01:18:27,850
five times the amount of usage
2736
01:18:28,310 --> 01:18:29,930
that the regular $20 plan.
2737
01:18:29,990 --> 01:18:31,770
So the 5x or the
2738
01:18:32,170 --> 01:18:33,550
$20 plan, um,
2739
01:18:33,830 --> 01:18:36,550
would be arguably just 5x more.
2740
01:18:37,010 --> 01:18:39,290
But where things get confusing is
2741
01:18:39,310 --> 01:18:40,970
then they have their $200 plan
2742
01:18:41,030 --> 01:18:42,030
which is claimed to be the
2743
01:18:42,290 --> 01:18:42,790
20x
2744
01:18:43,150 --> 01:18:44,670
which in theory you think like
2745
01:18:44,730 --> 01:18:46,190
okay, 20x should,
2746
01:18:46,230 --> 01:18:48,130
should be 20 times the amount
2747
01:18:48,190 --> 01:18:50,350
of usage as the initial uh,
2748
01:18:50,490 --> 01:18:51,090
base plan.
2749
01:18:51,910 --> 01:18:52,090
But
2750
01:18:52,570 --> 01:18:54,170
it seems like that may not
2751
01:18:54,230 --> 01:18:54,930
be the case.
2752
01:18:55,250 --> 01:18:56,250
And uh,
2753
01:18:56,490 --> 01:18:57,710
it's hard to say the exact
2754
01:18:57,750 --> 01:18:58,150
multiple,
2755
01:18:58,510 --> 01:19:00,090
but it seems like the multiple
2756
01:19:00,170 --> 01:19:02,330
of usage is actually not 20x
2757
01:19:02,390 --> 01:19:05,130
but it's something less than that,
2758
01:19:05,270 --> 01:19:07,750
probably closer to 5x, maybe um,
2759
01:19:07,830 --> 01:19:08,270
10x.
2760
01:19:08,470 --> 01:19:12,410
But it is um, uh,
2761
01:19:13,230 --> 01:19:16,310
at 20x the usage on a
2762
01:19:16,390 --> 01:19:18,370
five hour window meaning
2763
01:19:19,010 --> 01:19:19,790
my read of it,
2764
01:19:20,270 --> 01:19:21,570
it means that you can use
2765
01:19:21,610 --> 01:19:23,850
your model usage really really fast.
2766
01:19:23,990 --> 01:19:24,710
So if you want to use
2767
01:19:24,750 --> 01:19:25,550
it all at once,
2768
01:19:25,630 --> 01:19:27,410
do burst usage, um,
2769
01:19:27,870 --> 01:19:29,650
that is where the $200 plan
2770
01:19:30,130 --> 01:19:31,690
uh, would really shine.
2771
01:19:31,770 --> 01:19:33,310
But if you don't care about
2772
01:19:33,350 --> 01:19:35,350
it being really fast, um,
2773
01:19:35,610 --> 01:19:38,010
but you are okay with spreading
2774
01:19:38,050 --> 01:19:39,330
the task out over time you
2775
01:19:39,370 --> 01:19:41,230
can actually get more total usage,
2776
01:19:41,290 --> 01:19:42,910
especially if you have longer running
2777
01:19:42,970 --> 01:19:43,450
tasks.
2778
01:19:43,510 --> 01:19:44,650
And I think that actually kind
2779
01:19:44,690 --> 01:19:45,290
of makes sense.
2780
01:19:45,930 --> 01:19:46,050
And
2781
01:19:46,510 --> 01:19:47,590
let me make the argument for
2782
01:19:47,670 --> 01:19:48,190
why that
2783
01:19:48,770 --> 01:19:50,850
maybe makes sense on Anthropic's end.
2784
01:19:51,190 --> 01:19:52,150
And it's because
2785
01:19:52,770 --> 01:19:54,270
if they can spread out your
2786
01:19:54,390 --> 01:19:57,470
usage over time, you're not, they,
2787
01:19:57,670 --> 01:19:59,770
they can serve more concurrent requests
2788
01:19:59,870 --> 01:20:01,330
and use it users, right?
2789
01:20:01,410 --> 01:20:03,330
Because if they have let's say
2790
01:20:03,450 --> 01:20:05,530
one Nvidia, uh,
2791
01:20:05,630 --> 01:20:07,450
what is it like their B300
2792
01:20:07,870 --> 01:20:10,850
um, like server, uh,
2793
01:20:11,230 --> 01:20:13,290
they can maybe and let's say
2794
01:20:13,490 --> 01:20:14,510
that uh,
2795
01:20:14,930 --> 01:20:16,730
one Nvidia rack is able to
2796
01:20:16,810 --> 01:20:18,130
generate um,
2797
01:20:19,930 --> 01:20:20,830
let's say it's like
2798
01:20:21,250 --> 01:20:21,510
I don't know,
2799
01:20:21,830 --> 01:20:24,430
a thousand tokens per second.
2800
01:20:24,890 --> 01:20:25,110
Um,
2801
01:20:25,270 --> 01:20:26,090
and then you can decide to
2802
01:20:26,190 --> 01:20:26,950
spread that out any,
2803
01:20:27,290 --> 01:20:28,050
any way you want, right?
2804
01:20:28,150 --> 01:20:29,590
So it's like if one Nvidia
2805
01:20:29,670 --> 01:20:31,670
rack can do 1,000 tokens per
2806
01:20:31,710 --> 01:20:34,170
second, well you could give um,
2807
01:20:34,790 --> 01:20:36,770
one person, uh,
2808
01:20:37,070 --> 01:20:39,130
all 1,000 tokens all at once,
2809
01:20:39,330 --> 01:20:40,330
right? That, that would work.
2810
01:20:40,750 --> 01:20:42,510
Or you could give um,
2811
01:20:43,410 --> 01:20:45,530
10 people 100 tokens per second
2812
01:20:45,610 --> 01:20:47,170
at once and serve 10 concurrent
2813
01:20:47,250 --> 01:20:48,290
users all at the same time.
2814
01:20:48,670 --> 01:20:50,550
So if you are doing
2815
01:20:51,290 --> 01:20:53,890
burst usage you would be um,
2816
01:20:55,050 --> 01:20:56,990
you could take all 1,000 tokens
2817
01:20:57,350 --> 01:20:58,450
per second all at the same
2818
01:20:58,510 --> 01:20:58,910
time.
2819
01:20:59,190 --> 01:21:00,170
But then that would be at
2820
01:21:00,230 --> 01:21:02,150
the expense of
2821
01:21:02,750 --> 01:21:04,510
the nine other people who now
2822
01:21:04,570 --> 01:21:05,790
have to use a different server
2823
01:21:05,830 --> 01:21:07,549
and they'd have to scale up
2824
01:21:07,590 --> 01:21:07,730
right?
2825
01:21:07,810 --> 01:21:09,590
So it's actually very expensive uh,
2826
01:21:09,830 --> 01:21:11,670
for these companies to scale up
2827
01:21:12,110 --> 01:21:12,210
uh,
2828
01:21:12,330 --> 01:21:14,610
the token usage in that way.
2829
01:21:15,010 --> 01:21:15,170
Um,
2830
01:21:15,750 --> 01:21:15,970
so
2831
01:21:16,830 --> 01:21:18,710
my read on this is they
2832
01:21:18,830 --> 01:21:21,690
are uh, charging you more,
2833
01:21:21,750 --> 01:21:23,550
more because you are a token
2834
01:21:23,730 --> 01:21:24,970
hog, um,
2835
01:21:25,110 --> 01:21:27,090
and using the thousand tokens
2836
01:21:27,730 --> 01:21:28,470
all at once.
2837
01:21:29,010 --> 01:21:30,350
Whereas if you said like, okay,
2838
01:21:30,510 --> 01:21:31,850
I'm okay with
2839
01:21:32,610 --> 01:21:34,450
not doing this task uh,
2840
01:21:34,850 --> 01:21:36,670
in 10 minutes,
2841
01:21:36,710 --> 01:21:37,990
but I'm okay doing it in
2842
01:21:38,030 --> 01:21:38,650
an hour
2843
01:21:39,170 --> 01:21:40,110
and I'll have it a little
2844
01:21:40,130 --> 01:21:40,970
bit slower.
2845
01:21:41,410 --> 01:21:41,550
Uh,
2846
01:21:41,990 --> 01:21:44,030
but I'm actually able to increase
2847
01:21:44,090 --> 01:21:46,210
the total amount of usage.
2848
01:21:46,470 --> 01:21:47,310
Uh, so it's like okay,
2849
01:21:47,450 --> 01:21:48,650
I'm not getting uh,
2850
01:21:48,850 --> 01:21:50,030
a thousand tokens per second,
2851
01:21:50,430 --> 01:21:51,710
but I'm actually going to get
2852
01:21:51,810 --> 01:21:53,290
100 tokens per second spread out
2853
01:21:53,350 --> 01:21:53,910
over an hour.
2854
01:21:54,350 --> 01:21:54,570
Um,
2855
01:21:54,990 --> 01:21:55,830
as long as you have enough
2856
01:21:55,910 --> 01:21:57,871
tasks to give it, you uh,
2857
01:21:57,910 --> 01:22:00,230
can actually get more usage.
2858
01:22:00,290 --> 01:22:01,730
And I think that is where
2859
01:22:01,810 --> 01:22:02,970
Anthropic is coming from.
2860
01:22:03,130 --> 01:22:04,530
But they make it very opaque
2861
01:22:04,590 --> 01:22:05,790
on their website that that's what
2862
01:22:05,810 --> 01:22:06,190
they're doing.
2863
01:22:06,230 --> 01:22:07,190
But that's my read of it.
2864
01:22:08,430 --> 01:22:08,490
Yeah,
2865
01:22:08,510 --> 01:22:09,390
it turns out they have a
2866
01:22:09,450 --> 01:22:10,530
lawsuit over this too.
2867
01:22:10,709 --> 01:22:12,709
The 20x claim, um,
2868
01:22:13,270 --> 01:22:14,470
is only the uh,
2869
01:22:14,650 --> 01:22:15,410
on the five hour
2870
01:22:16,010 --> 01:22:17,250
limit and not the weekly limit.
2871
01:22:17,270 --> 01:22:18,990
And there's, there's an ongoing lawsuit,
2872
01:22:19,890 --> 01:22:20,170
class uh,
2873
01:22:20,171 --> 01:22:21,930
action lawsuit filed against Anthropic in
2874
01:22:21,950 --> 01:22:24,610
the US District Court in California.
2875
01:22:25,670 --> 01:22:26,110
But. Right.
2876
01:22:26,190 --> 01:22:28,550
So it's heavily implied that you
2877
01:22:28,590 --> 01:22:29,730
get 20 times the usage,
2878
01:22:29,850 --> 01:22:31,990
but you really don't um,
2879
01:22:32,810 --> 01:22:34,310
like the total amount of tokens
2880
01:22:34,410 --> 01:22:35,710
is actually not 20x.
2881
01:22:36,110 --> 01:22:36,150
Uh,
2882
01:22:36,250 --> 01:22:38,490
so that is why I subscribe
2883
01:22:38,550 --> 01:22:39,230
to not one,
2884
01:22:39,390 --> 01:22:41,070
but two $100 plans.
2885
01:22:41,610 --> 01:22:42,510
Um, but, but uh,
2886
01:22:42,990 --> 01:22:43,950
if you can, this uh,
2887
01:22:43,970 --> 01:22:44,910
would be like a fun experiment
2888
01:22:44,950 --> 01:22:45,550
for our readers
2889
01:22:45,970 --> 01:22:47,030
if you can track your token
2890
01:22:47,090 --> 01:22:47,450
usage.
2891
01:22:47,530 --> 01:22:48,290
I know, um,
2892
01:22:48,510 --> 01:22:49,890
the subscriptions make it hard.
2893
01:22:50,430 --> 01:22:50,710
Um,
2894
01:22:51,090 --> 01:22:53,450
but I'm Pretty sure the 200amonth
2895
01:22:53,490 --> 01:22:53,810
plan
2896
01:22:54,330 --> 01:22:56,570
is twice the weekly usage limit
2897
01:22:56,650 --> 01:22:57,870
as the $100 a month plan.
2898
01:22:58,370 --> 01:22:58,570
Um,
2899
01:22:58,690 --> 01:23:00,430
a couple Reddit threads have confirmed
2900
01:23:00,490 --> 01:23:00,630
it.
2901
01:23:01,270 --> 01:23:02,010
I know this is all just
2902
01:23:02,090 --> 01:23:03,310
anecdotal, but I think uh,
2903
01:23:04,250 --> 01:23:05,410
it might make your life easier
2904
01:23:05,470 --> 01:23:05,650
to
2905
01:23:06,110 --> 01:23:07,670
manage one plan as opposed to
2906
01:23:07,730 --> 01:23:07,930
two.
2907
01:23:07,970 --> 01:23:09,090
But uh,
2908
01:23:09,770 --> 01:23:12,371
and you would get higher um,
2909
01:23:12,390 --> 01:23:13,190
five hour limits.
2910
01:23:13,870 --> 01:23:15,950
Yeah, that's probably true.
2911
01:23:16,050 --> 01:23:16,270
Um,
2912
01:23:16,630 --> 01:23:17,830
but one other reason why I
2913
01:23:17,890 --> 01:23:18,710
kind of like having the two
2914
01:23:18,810 --> 01:23:20,990
100 plans is in case
2915
01:23:21,690 --> 01:23:22,610
I mess something up.
2916
01:23:22,950 --> 01:23:24,270
And I have.
2917
01:23:24,330 --> 01:23:26,430
So one thing I noticed,
2918
01:23:26,470 --> 01:23:26,790
and this is,
2919
01:23:26,830 --> 01:23:27,390
this is a little bit
2920
01:23:27,750 --> 01:23:29,090
more inside baseball for people,
2921
01:23:29,150 --> 01:23:29,350
but
2922
01:23:30,890 --> 01:23:33,530
I ran into an issue where
2923
01:23:34,310 --> 01:23:35,950
I had um,
2924
01:23:37,210 --> 01:23:38,570
a bunch of sub agents queued
2925
01:23:38,630 --> 01:23:38,770
up.
2926
01:23:39,010 --> 01:23:41,050
And Anthropic seems to be very
2927
01:23:41,110 --> 01:23:42,370
verbose when it comes to sub
2928
01:23:42,430 --> 01:23:44,170
agents because uh,
2929
01:23:44,410 --> 01:23:45,250
when you have
2930
01:23:45,790 --> 01:23:46,930
multiple sub agents running,
2931
01:23:47,130 --> 01:23:48,710
there's a lot of tokens spent
2932
01:23:48,810 --> 01:23:50,630
on coordinating these sub agents.
2933
01:23:50,850 --> 01:23:52,870
And oftentimes I would find that
2934
01:23:52,950 --> 01:23:54,270
If I had a few sub
2935
01:23:54,310 --> 01:23:56,050
agents running, yeah,
2936
01:23:56,170 --> 01:23:57,410
you could get things done a
2937
01:23:57,450 --> 01:23:58,150
little bit faster.
2938
01:23:58,270 --> 01:24:00,110
But it just eats tokens like
2939
01:24:00,170 --> 01:24:02,410
ridiculously quickly because there's just so
2940
01:24:02,450 --> 01:24:04,350
many tokens spent trying to
2941
01:24:04,750 --> 01:24:06,630
manage what each individual sub agent
2942
01:24:06,690 --> 01:24:07,110
is doing.
2943
01:24:07,230 --> 01:24:09,431
And oftentimes you um,
2944
01:24:09,470 --> 01:24:12,150
can go and just have one
2945
01:24:12,230 --> 01:24:13,370
thread do the single task.
2946
01:24:13,390 --> 01:24:13,690
And yeah,
2947
01:24:13,770 --> 01:24:14,170
it'll take
2948
01:24:14,810 --> 01:24:17,030
longer but the overall results will
2949
01:24:17,050 --> 01:24:18,289
be better and you won't have
2950
01:24:18,309 --> 01:24:18,990
as many tokens.
2951
01:24:19,090 --> 01:24:20,090
And for a lot of things,
2952
01:24:20,170 --> 01:24:20,550
if you're
2953
01:24:21,450 --> 01:24:23,410
not waiting for the work to
2954
01:24:23,450 --> 01:24:24,870
be done, it's fine to wait.
2955
01:24:24,930 --> 01:24:26,250
So oftentimes I'll set a long
2956
01:24:26,290 --> 01:24:27,850
running task overnight and I'll wake
2957
01:24:27,930 --> 01:24:28,850
up and it'll be done.
2958
01:24:29,090 --> 01:24:30,250
And whether it took like you
2959
01:24:30,290 --> 01:24:32,490
know, one hour or five,
2960
01:24:32,850 --> 01:24:33,670
I sleep for eight so I
2961
01:24:33,690 --> 01:24:35,930
don't really care, you know.
2962
01:24:36,290 --> 01:24:36,470
Uh,
2963
01:24:36,550 --> 01:24:37,250
so for a lot of these
2964
01:24:37,290 --> 01:24:37,550
things,
2965
01:24:37,670 --> 01:24:39,490
like it's fine to not use
2966
01:24:39,570 --> 01:24:41,410
subagents and have it uh,
2967
01:24:41,670 --> 01:24:42,750
go for a little bit longer.
2968
01:24:42,790 --> 01:24:44,490
But I was finding that like
2969
01:24:45,110 --> 01:24:46,190
for the five hour limit if
2970
01:24:46,250 --> 01:24:47,710
I use multiple subagents I can
2971
01:24:47,750 --> 01:24:49,030
burn through the entire five hour
2972
01:24:49,070 --> 01:24:51,730
limit, uh, in like you know,
2973
01:24:52,090 --> 01:24:53,270
30 minutes or something like that
2974
01:24:53,650 --> 01:24:55,110
when I use multiple sub agents
2975
01:24:55,170 --> 01:24:55,990
or even less.
2976
01:24:56,410 --> 01:24:56,530
Uh,
2977
01:24:56,730 --> 01:24:57,550
and a lot of that is
2978
01:24:57,590 --> 01:24:58,970
just coordination, um,
2979
01:24:59,070 --> 01:25:00,510
where you just completely run out
2980
01:25:00,570 --> 01:25:01,970
of your five hour limit and
2981
01:25:02,010 --> 01:25:03,350
you have to wait another four
2982
01:25:03,450 --> 01:25:04,590
hours for it to reset.
2983
01:25:04,950 --> 01:25:06,070
And if I mess that up
2984
01:25:06,150 --> 01:25:07,070
on one of the accounts,
2985
01:25:07,609 --> 01:25:08,950
at uh, least the other one,
2986
01:25:09,030 --> 01:25:10,110
I probably didn't have all the
2987
01:25:10,170 --> 01:25:11,010
sub agents running.
2988
01:25:11,070 --> 01:25:12,270
So therefore I can just switch
2989
01:25:12,310 --> 01:25:13,230
to that and I won't have
2990
01:25:13,290 --> 01:25:14,270
to stop my work.
2991
01:25:14,670 --> 01:25:14,790
Um,
2992
01:25:14,890 --> 01:25:16,250
so therefore that's another reason why
2993
01:25:16,270 --> 01:25:17,670
I actually kind of don't mind
2994
01:25:17,730 --> 01:25:19,090
the two 100 plans.
2995
01:25:19,570 --> 01:25:20,290
Uh, it's not bad.
2996
01:25:20,370 --> 01:25:21,210
I still use the same credit
2997
01:25:21,250 --> 01:25:21,410
card,
2998
01:25:21,470 --> 01:25:22,990
it's just two separate emails and
2999
01:25:23,050 --> 01:25:24,170
I can also have different settings
3000
01:25:24,210 --> 01:25:24,790
for each one.
3001
01:25:24,990 --> 01:25:26,530
So for me it's,
3002
01:25:26,990 --> 01:25:27,970
it's not so bad.
3003
01:25:28,030 --> 01:25:28,350
Um, yeah,
3004
01:25:28,410 --> 01:25:30,010
it probably would be slightly easier
3005
01:25:30,050 --> 01:25:31,330
to have a 1, 200 plan,
3006
01:25:31,550 --> 01:25:33,710
but I'm okay with this.
3007
01:25:34,030 --> 01:25:34,190
Um,
3008
01:25:34,390 --> 01:25:36,010
this has been working for me
3009
01:25:36,090 --> 01:25:36,370
and I'm,
3010
01:25:37,030 --> 01:25:39,170
I'm like pretty happy with it.
3011
01:25:39,430 --> 01:25:39,850
And uh,
3012
01:25:40,150 --> 01:25:41,890
I have not actually
3013
01:25:42,890 --> 01:25:43,890
this has been the first time
3014
01:25:43,930 --> 01:25:44,350
in a while
3015
01:25:44,730 --> 01:25:46,050
that I haven't just burned through
3016
01:25:46,170 --> 01:25:47,210
all of my usage.
3017
01:25:47,290 --> 01:25:48,550
So it seems to be like
3018
01:25:48,630 --> 01:25:49,590
pretty generous.
3019
01:25:49,630 --> 01:25:50,850
Like last week I had like
3020
01:25:50,990 --> 01:25:52,010
20% less on one of the
3021
01:25:52,470 --> 01:25:53,490
left on one of the plans.
3022
01:25:54,170 --> 01:25:57,890
So it's been very generous.
3023
01:25:58,150 --> 01:25:59,830
Uh, from anthropic, uh,
3024
01:26:00,170 --> 01:26:01,790
at least right now I
3025
01:26:02,310 --> 01:26:03,950
really am happy with clock code.
3026
01:26:05,630 --> 01:26:06,290
I know we dove into this
3027
01:26:06,350 --> 01:26:08,130
topic, uh, a lot already.
3028
01:26:08,190 --> 01:26:09,970
But, uh, I'm curious why you,
3029
01:26:10,110 --> 01:26:10,350
um,
3030
01:26:10,690 --> 01:26:12,429
didn't keep one of your ChatGPT
3031
01:26:12,470 --> 01:26:13,010
subscriptions.
3032
01:26:13,050 --> 01:26:13,870
Because, you know,
3033
01:26:13,950 --> 01:26:14,350
you don't want
3034
01:26:15,670 --> 01:26:17,670
two models from the same distribution,
3035
01:26:17,950 --> 01:26:19,350
you know, working with each other.
3036
01:26:19,430 --> 01:26:20,910
It might be helpful to have
3037
01:26:20,970 --> 01:26:22,730
a different model, um,
3038
01:26:22,970 --> 01:26:24,910
kind of be adversarial and work
3039
01:26:24,970 --> 01:26:25,370
together
3040
01:26:25,870 --> 01:26:27,530
and bring in a different set
3041
01:26:27,570 --> 01:26:28,790
of training data to
3042
01:26:29,390 --> 01:26:30,850
fill in the gaps where OPUS
3043
01:26:30,910 --> 01:26:31,610
kind of fails.
3044
01:26:32,230 --> 01:26:34,010
Um, yeah, you know,
3045
01:26:34,070 --> 01:26:34,850
that's a good question.
3046
01:26:34,890 --> 01:26:36,110
And I actually did do that
3047
01:26:36,190 --> 01:26:37,570
for a little while.
3048
01:26:37,870 --> 01:26:38,090
Um,
3049
01:26:38,290 --> 01:26:40,510
so for
3050
01:26:41,570 --> 01:26:43,650
maybe two weeks or so, two,
3051
01:26:43,690 --> 01:26:45,550
three weeks, I had ah,
3052
01:26:45,750 --> 01:26:47,830
a 200 Codex subscription and I
3053
01:26:47,870 --> 01:26:50,731
had a $100 cloud, ah,
3054
01:26:50,770 --> 01:26:51,570
code subscription.
3055
01:26:52,070 --> 01:26:53,370
And you should still have your
3056
01:26:53,550 --> 01:26:54,210
credits, right?
3057
01:26:54,610 --> 01:26:55,890
You can continue using the Open
3058
01:26:55,930 --> 01:26:56,750
Air because they gave you a
3059
01:26:56,770 --> 01:26:57,470
bunch of credits.
3060
01:26:58,130 --> 01:26:58,910
Yeah, I have like,
3061
01:26:59,750 --> 01:27:00,130
I don't know,
3062
01:27:00,510 --> 01:27:01,810
60 something thousand credits,
3063
01:27:02,050 --> 01:27:02,730
which I haven't.
3064
01:27:02,770 --> 01:27:05,450
That's like 2500 worth of credits.
3065
01:27:05,510 --> 01:27:06,090
You should use it.
3066
01:27:06,570 --> 01:27:07,530
Yeah, I should use it.
3067
01:27:08,150 --> 01:27:08,790
So I, I will,
3068
01:27:08,910 --> 01:27:09,850
I will use it.
3069
01:27:09,970 --> 01:27:11,830
Um, and I,
3070
01:27:11,930 --> 01:27:12,870
I hope they don't expire.
3071
01:27:13,230 --> 01:27:13,970
Uh, but yeah,
3072
01:27:14,070 --> 01:27:15,590
they do expire at the end
3073
01:27:15,610 --> 01:27:15,950
of the year.
3074
01:27:16,650 --> 01:27:17,330
All right, well,
3075
01:27:17,410 --> 01:27:19,650
then I will definitely find something
3076
01:27:19,710 --> 01:27:20,530
to do with them.
3077
01:27:20,750 --> 01:27:22,950
Um, because that is worth some,
3078
01:27:23,230 --> 01:27:23,790
some real money.
3079
01:27:24,330 --> 01:27:27,130
Um, so anyways, I have that.
3080
01:27:27,390 --> 01:27:29,490
I have like 60 something thousand
3081
01:27:29,550 --> 01:27:30,670
credits, which I haven't touched.
3082
01:27:30,850 --> 01:27:31,830
But the thing is,
3083
01:27:32,110 --> 01:27:32,570
I was actually
3084
01:27:33,430 --> 01:27:34,870
doing exactly what you said where
3085
01:27:34,990 --> 01:27:35,610
I would, uh,
3086
01:27:35,710 --> 01:27:37,610
try the same task, um,
3087
01:27:38,970 --> 01:27:40,450
on both models and
3088
01:27:40,810 --> 01:27:42,390
see how both, uh, did.
3089
01:27:42,450 --> 01:27:43,210
And I'd have a bit of
3090
01:27:43,230 --> 01:27:45,730
an adversarial relationship, um, between them.
3091
01:27:46,490 --> 01:27:48,090
And what I noticed is on
3092
01:27:48,170 --> 01:27:51,090
Codex, um, one,
3093
01:27:51,990 --> 01:27:53,390
when I was using Astra,
3094
01:27:53,490 --> 01:27:53,650
which
3095
01:27:54,090 --> 01:27:55,070
you kind of need to be
3096
01:27:55,110 --> 01:27:56,290
using Astra if you want to
3097
01:27:56,830 --> 01:27:58,810
compete with Opus 5.5, one,
3098
01:27:59,070 --> 01:28:00,590
it was a lot slower.
3099
01:28:01,330 --> 01:28:02,310
Um, two,
3100
01:28:02,630 --> 01:28:05,360
it was much more token hungry.
3101
01:28:05,500 --> 01:28:06,860
So I would ask like one
3102
01:28:07,240 --> 01:28:09,320
seemingly simple request and it would
3103
01:28:09,380 --> 01:28:11,720
burn through around 1% of my,
3104
01:28:12,280 --> 01:28:12,400
uh,
3105
01:28:12,540 --> 01:28:13,500
weekly limit
3106
01:28:14,080 --> 01:28:14,920
or a couple percent of my
3107
01:28:14,960 --> 01:28:15,600
weekly limit.
3108
01:28:15,780 --> 01:28:17,300
Whereas I could do
3109
01:28:17,980 --> 01:28:20,540
multiple requests with Opus 5.5 and
3110
01:28:20,620 --> 01:28:21,100
it wouldn't
3111
01:28:21,680 --> 01:28:23,320
burn through credits that quickly.
3112
01:28:23,760 --> 01:28:25,360
So OPUS was
3113
01:28:25,940 --> 01:28:27,840
much more sparse about the credit
3114
01:28:27,920 --> 01:28:28,280
usage.
3115
01:28:28,880 --> 01:28:31,160
Um, and uh,
3116
01:28:31,840 --> 01:28:34,420
oftentimes the results from Opus were
3117
01:28:34,460 --> 01:28:37,300
just far superior to Astra.
3118
01:28:37,380 --> 01:28:40,520
So, um, I would have, uh,
3119
01:28:40,760 --> 01:28:41,160
Opus
3120
01:28:41,640 --> 01:28:42,700
maybe do a design or a
3121
01:28:42,760 --> 01:28:45,280
code change and I would
3122
01:28:45,920 --> 01:28:47,340
have Astra review it and I
3123
01:28:47,360 --> 01:28:48,320
would say like, hey, Astra,
3124
01:28:48,400 --> 01:28:49,560
did you find any issues with
3125
01:28:49,620 --> 01:28:49,820
this?
3126
01:28:50,240 --> 01:28:51,720
And it would say, no,
3127
01:28:52,300 --> 01:28:53,660
I found no issues with that
3128
01:28:53,720 --> 01:28:55,340
code. It seems perfect.
3129
01:28:55,800 --> 01:28:56,301
Um, I'm m.
3130
01:28:56,320 --> 01:28:56,980
Like, all right, well,
3131
01:28:57,060 --> 01:28:58,640
that is useless.
3132
01:28:59,080 --> 01:28:59,980
Um, I mean, not useless.
3133
01:29:00,040 --> 01:29:00,780
To know that it's like,
3134
01:29:01,060 --> 01:29:01,620
you know, really good,
3135
01:29:01,680 --> 01:29:02,600
but it wouldn't be able to
3136
01:29:02,640 --> 01:29:03,380
find any issues.
3137
01:29:03,640 --> 01:29:05,080
But then I would do the
3138
01:29:05,220 --> 01:29:06,240
opposite where I would have, like,
3139
01:29:06,299 --> 01:29:09,020
you know, Claude review asterisk code,
3140
01:29:09,380 --> 01:29:10,040
and it would find a bunch
3141
01:29:10,080 --> 01:29:11,140
of stuff, um,
3142
01:29:11,300 --> 01:29:12,560
that was wrong with the asterisk
3143
01:29:12,620 --> 01:29:12,920
code.
3144
01:29:13,140 --> 01:29:14,580
And then when I had
3145
01:29:15,140 --> 01:29:17,600
Astronomy review the Claude code,
3146
01:29:17,800 --> 01:29:19,500
the Astro wouldn't find anything.
3147
01:29:19,940 --> 01:29:21,120
But if I had,
3148
01:29:21,860 --> 01:29:22,980
what I found actually worked even
3149
01:29:23,020 --> 01:29:24,700
better was I would sometimes have
3150
01:29:24,800 --> 01:29:26,120
Claude review Claude.
3151
01:29:26,800 --> 01:29:27,000
Um,
3152
01:29:27,120 --> 01:29:28,100
and then that seems to be
3153
01:29:28,120 --> 01:29:29,440
working okay, where, like, you know,
3154
01:29:29,600 --> 01:29:30,720
Claude, being a better model,
3155
01:29:30,940 --> 01:29:32,380
would be able to find things
3156
01:29:32,440 --> 01:29:33,660
that Astro wouldn't be able to
3157
01:29:33,720 --> 01:29:34,040
find.
3158
01:29:34,120 --> 01:29:35,160
So just running, like,
3159
01:29:35,220 --> 01:29:36,740
another instance of Claude saying, like,
3160
01:29:36,820 --> 01:29:38,720
hey, review this work as, like,
3161
01:29:38,760 --> 01:29:39,420
a black box,
3162
01:29:39,520 --> 01:29:41,500
it would find things that, uh,
3163
01:29:42,080 --> 01:29:42,860
Astra couldn't.
3164
01:29:43,120 --> 01:29:45,020
Um, so, like, to me,
3165
01:29:45,480 --> 01:29:46,900
it feels like, uh,
3166
01:29:47,180 --> 01:29:50,240
anthropic opus 5.5 is, like,
3167
01:29:51,320 --> 01:29:53,820
a step change from, uh,
3168
01:29:54,680 --> 01:29:54,880
what
3169
01:29:55,440 --> 01:29:57,340
OpenAI would claim is true AGI.
3170
01:29:57,800 --> 01:29:59,600
We probably hyped as AGI too,
3171
01:29:59,960 --> 01:30:00,140
but.
3172
01:30:00,180 --> 01:30:01,740
Yeah, so if, uh,
3173
01:30:02,120 --> 01:30:04,280
if Astra is AGI,
3174
01:30:04,440 --> 01:30:06,700
then opus 5.5 is like AGI
3175
01:30:06,780 --> 01:30:08,740
plus, because it's, like,
3176
01:30:08,840 --> 01:30:11,880
even better AGI. Nice.
3177
01:30:12,840 --> 01:30:14,000
Well, um, yeah,
3178
01:30:14,080 --> 01:30:15,580
Anthropic has been doing an amazing
3179
01:30:15,620 --> 01:30:15,880
job,
3180
01:30:15,960 --> 01:30:17,820
and they're supposedly going to Release,
3181
01:30:17,900 --> 01:30:19,720
uh, Fable 5.5 tomorrow, too.
3182
01:30:19,780 --> 01:30:21,420
So, uh,
3183
01:30:21,540 --> 01:30:22,420
if you want to burn some,
3184
01:30:23,180 --> 01:30:25,080
um, money,
3185
01:30:25,200 --> 01:30:26,220
because I don't think they'll include
3186
01:30:26,260 --> 01:30:27,420
it in the subscription usage,
3187
01:30:27,640 --> 01:30:28,420
you can try that out.
3188
01:30:29,520 --> 01:30:29,661
We'll, uh,
3189
01:30:29,680 --> 01:30:30,960
get back to our users about,
3190
01:30:31,060 --> 01:30:31,180
uh,
3191
01:30:31,400 --> 01:30:32,780
how the frontier has changed yet
3192
01:30:32,820 --> 01:30:33,140
again.
3193
01:30:34,000 --> 01:30:34,120
Yeah,
3194
01:30:34,220 --> 01:30:35,300
probably next week would be like,
3195
01:30:35,380 --> 01:30:35,800
oh, man,
3196
01:30:35,900 --> 01:30:37,560
Opus 5.5 is so dumb.
3197
01:30:38,200 --> 01:30:41,580
Uh, Fable 5.5, step change.
3198
01:30:43,080 --> 01:30:44,240
It did my taxes for me
3199
01:30:44,900 --> 01:30:46,200
and made me a million dollars.
3200
01:30:46,920 --> 01:30:48,040
Uh, anyways,
3201
01:30:48,480 --> 01:30:50,160
I think that's pretty much all
3202
01:30:50,480 --> 01:30:51,160
I had. I don't know.
3203
01:30:51,240 --> 01:30:51,520
Shashank,
3204
01:30:51,560 --> 01:30:52,360
do you have anything else that
3205
01:30:52,400 --> 01:30:52,960
you want to mention?
3206
01:30:53,800 --> 01:30:54,800
Um, I mean, yeah, there's.
3207
01:30:54,840 --> 01:30:56,940
There's always a couple other things,
3208
01:30:57,020 --> 01:30:57,900
but I think those are the
3209
01:30:57,960 --> 01:30:58,940
major announcements.
3210
01:30:59,040 --> 01:30:59,140
Uh,
3211
01:30:59,220 --> 01:31:00,540
maybe I can mention it briefly.
3212
01:31:01,160 --> 01:31:01,400
Um,
3213
01:31:01,620 --> 01:31:02,760
a lot of the always on
3214
01:31:02,840 --> 01:31:03,300
agents, you know,
3215
01:31:03,360 --> 01:31:06,240
we mentioned Musepark and OpenAI, uh,
3216
01:31:06,280 --> 01:31:06,680
dots.
3217
01:31:07,280 --> 01:31:08,120
Um, I, uh,
3218
01:31:08,140 --> 01:31:09,420
was chatting to Mark about this
3219
01:31:09,480 --> 01:31:12,181
before the podcast, and both, um,
3220
01:31:12,240 --> 01:31:12,580
him and I,
3221
01:31:12,640 --> 01:31:14,040
we don't really use these
3222
01:31:14,740 --> 01:31:16,360
always on agents as much.
3223
01:31:16,500 --> 01:31:17,540
You know, maybe we should be.
3224
01:31:17,580 --> 01:31:20,540
But we have our coding agents,
3225
01:31:20,600 --> 01:31:22,580
our desktops, uh, plugged in,
3226
01:31:22,660 --> 01:31:23,680
always running all the time.
3227
01:31:23,720 --> 01:31:25,600
We have remote control and remote
3228
01:31:25,700 --> 01:31:26,040
access,
3229
01:31:26,120 --> 01:31:27,920
so we can fire off complex,
3230
01:31:28,480 --> 01:31:29,300
you know, um,
3231
01:31:29,460 --> 01:31:30,480
coding tasks that
3232
01:31:31,000 --> 01:31:31,860
do a lot of work.
3233
01:31:32,460 --> 01:31:32,940
Um,
3234
01:31:33,180 --> 01:31:35,580
so it's like we don't uh,
3235
01:31:35,840 --> 01:31:37,080
we're not the target demographic,
3236
01:31:37,240 --> 01:31:37,800
but it is
3237
01:31:38,200 --> 01:31:39,920
cool to see these um,
3238
01:31:40,100 --> 01:31:41,060
agents maturing.
3239
01:31:41,460 --> 01:31:43,340
They're getting their own email addresses,
3240
01:31:43,400 --> 01:31:44,600
their own sandboxes.
3241
01:31:45,160 --> 01:31:45,400
Um,
3242
01:31:45,480 --> 01:31:46,900
they're being treated as first class
3243
01:31:46,940 --> 01:31:47,720
citizens now.
3244
01:31:48,300 --> 01:31:48,520
Um,
3245
01:31:48,640 --> 01:31:49,660
they're getting a lot of pushback
3246
01:31:49,720 --> 01:31:51,600
from some of the infrastructure providers
3247
01:31:51,660 --> 01:31:53,020
because, uh, uh,
3248
01:31:53,540 --> 01:31:54,880
some of the companies are blocking
3249
01:31:54,980 --> 01:31:56,460
access to these agents because they
3250
01:31:56,500 --> 01:31:57,700
don't want, uh,
3251
01:31:57,900 --> 01:32:00,280
bots spamming their websites and taking
3252
01:32:00,320 --> 01:32:01,100
up all the traffic.
3253
01:32:01,200 --> 01:32:02,500
But I think like,
3254
01:32:02,700 --> 01:32:03,820
I want my bots to be
3255
01:32:03,910 --> 01:32:05,270
out there doing work for me
3256
01:32:05,330 --> 01:32:05,590
so like,
3257
01:32:05,750 --> 01:32:07,390
please let this be the new
3258
01:32:07,450 --> 01:32:07,990
normal.
3259
01:32:08,750 --> 01:32:11,390
Um, and yeah, I think, uh,
3260
01:32:11,430 --> 01:32:11,870
I think that's,
3261
01:32:12,070 --> 01:32:13,130
that's all I had on my
3262
01:32:13,210 --> 01:32:13,430
end.
3263
01:32:14,810 --> 01:32:14,950
Yeah,
3264
01:32:15,050 --> 01:32:15,390
these
3265
01:32:16,290 --> 01:32:18,330
always on agents I think are
3266
01:32:18,950 --> 01:32:20,830
kind of bringing the experience that
3267
01:32:20,890 --> 01:32:22,230
you and I have had for
3268
01:32:22,370 --> 01:32:23,810
the last year or so
3269
01:32:24,290 --> 01:32:26,570
and making it easy for everybody
3270
01:32:26,650 --> 01:32:27,030
to use.
3271
01:32:27,090 --> 01:32:29,690
Because I think that even just
3272
01:32:29,770 --> 01:32:32,270
the effort of opening a terminal
3273
01:32:32,390 --> 01:32:33,490
and running a, uh,
3274
01:32:33,670 --> 01:32:34,150
command
3275
01:32:34,590 --> 01:32:36,090
is too much for most people.
3276
01:32:36,150 --> 01:32:36,350
I mean,
3277
01:32:36,430 --> 01:32:37,170
you look at the terminal,
3278
01:32:37,250 --> 01:32:37,870
it's black.
3279
01:32:38,010 --> 01:32:38,170
I mean,
3280
01:32:38,230 --> 01:32:39,950
there's just a blinking cursor there.
3281
01:32:40,230 --> 01:32:40,910
You don't know what to type
3282
01:32:40,990 --> 01:32:41,190
in.
3283
01:32:41,330 --> 01:32:42,430
Like you have to know some
3284
01:32:42,470 --> 01:32:43,730
sort of arcane language
3285
01:32:44,350 --> 01:32:45,130
to run something.
3286
01:32:45,170 --> 01:32:45,530
And yeah,
3287
01:32:45,670 --> 01:32:47,090
it's not that hard once you
3288
01:32:47,130 --> 01:32:47,450
learn it.
3289
01:32:47,570 --> 01:32:48,670
But I think that
3290
01:32:50,050 --> 01:32:52,570
it is too much for a
3291
01:32:52,610 --> 01:32:54,090
lot of people who uh,
3292
01:32:54,250 --> 01:32:55,550
look at the terminal and close
3293
01:32:55,570 --> 01:32:57,730
it immediately because it's too intimidating.
3294
01:32:58,230 --> 01:32:58,430
Um,
3295
01:32:58,730 --> 01:33:00,250
so I think that
3296
01:33:00,970 --> 01:33:01,910
for everybody else,
3297
01:33:02,630 --> 01:33:05,170
these things like dots or Grokbot
3298
01:33:05,310 --> 01:33:05,490
are
3299
01:33:06,510 --> 01:33:08,270
very exciting to be able to
3300
01:33:08,330 --> 01:33:09,750
have it do what work on
3301
01:33:09,810 --> 01:33:11,810
your behalf without having you need
3302
01:33:11,850 --> 01:33:12,990
to prompt it every time and
3303
01:33:13,030 --> 01:33:14,770
making it like very user friendly.
3304
01:33:14,970 --> 01:33:16,050
So I kind of look at
3305
01:33:16,090 --> 01:33:17,810
these things that just like a
3306
01:33:18,050 --> 01:33:20,550
user friendly wrapper around an agent.
3307
01:33:20,610 --> 01:33:22,930
It's basically just like quad code
3308
01:33:23,010 --> 01:33:24,090
or Codex, whatever.
3309
01:33:24,390 --> 01:33:25,310
Plus Cron job
3310
01:33:25,950 --> 01:33:27,470
put on a nice web wrapper
3311
01:33:27,550 --> 01:33:29,270
that you can easily use it,
3312
01:33:29,390 --> 01:33:33,630
which isn't that interesting technically,
3313
01:33:34,030 --> 01:33:35,370
but I'd say it's
3314
01:33:35,850 --> 01:33:37,610
incredibly value for the world and
3315
01:33:37,650 --> 01:33:39,090
it actually makes it so like
3316
01:33:39,150 --> 01:33:40,050
my mom could use it,
3317
01:33:40,270 --> 01:33:41,590
which I think is pretty cool.
3318
01:33:44,190 --> 01:33:48,970
Um, anyways, that is that, um,
3319
01:33:49,530 --> 01:33:50,630
we mentioned at the beginning,
3320
01:33:50,810 --> 01:33:52,550
but thank you for listening.
3321
01:33:53,110 --> 01:33:53,290
Um,
3322
01:33:53,570 --> 01:33:54,890
now it's time for all of
3323
01:33:54,950 --> 01:33:56,310
the ad stuff.
3324
01:33:56,550 --> 01:33:57,310
So one,
3325
01:33:57,470 --> 01:33:58,650
if you're in the Bay Area
3326
01:33:58,770 --> 01:34:00,310
and you want to go to
3327
01:34:00,410 --> 01:34:01,490
a meetup, uh,
3328
01:34:01,650 --> 01:34:01,990
check out
3329
01:34:02,990 --> 01:34:04,870
the meetup that is tied to
3330
01:34:04,910 --> 01:34:05,730
this podcast.
3331
01:34:06,250 --> 01:34:07,110
It's a lot of fun.
3332
01:34:07,190 --> 01:34:07,270
Uh,
3333
01:34:07,350 --> 01:34:09,330
Shashank and I unfortunately aren't in
3334
01:34:09,750 --> 01:34:10,830
the United States anymore,
3335
01:34:10,990 --> 01:34:12,990
so we aren't going ourselves.
3336
01:34:13,310 --> 01:34:14,670
But the meetup is still running.
3337
01:34:14,770 --> 01:34:17,110
Um, we have some awesome uh,
3338
01:34:17,690 --> 01:34:19,490
volunteers who help run it, like,
3339
01:34:19,590 --> 01:34:21,170
uh, Jeff and Suzanne.
3340
01:34:21,690 --> 01:34:23,170
Um, uh,
3341
01:34:23,450 --> 01:34:24,630
a lot of people that are
3342
01:34:24,690 --> 01:34:24,990
helping.
3343
01:34:25,050 --> 01:34:26,510
I'm, um, not help, uh,
3344
01:34:26,630 --> 01:34:28,230
mentioned everybody, but they are helping.
3345
01:34:28,290 --> 01:34:29,730
So, uh, if you're listening,
3346
01:34:29,790 --> 01:34:31,970
thank you so much for helping,
3347
01:34:32,070 --> 01:34:33,750
uh, to run those meetups.
3348
01:34:34,090 --> 01:34:36,490
We definitely ah, really appreciate it.
3349
01:34:36,910 --> 01:34:39,910
Um, also check out novagat AI.
3350
01:34:40,130 --> 01:34:42,491
It is the video, uh,
3351
01:34:42,550 --> 01:34:45,350
editor that I am building.
3352
01:34:45,710 --> 01:34:45,790
Uh,
3353
01:34:46,770 --> 01:34:48,090
a lot of new features have
3354
01:34:48,150 --> 01:34:48,770
dropped for it
3355
01:34:49,410 --> 01:34:50,810
and it's getting a lot more
3356
01:34:50,870 --> 01:34:51,370
powerful.
3357
01:34:51,830 --> 01:34:53,550
Uh, so if you have uh,
3358
01:34:53,630 --> 01:34:54,770
if you're looking for not
3359
01:34:55,670 --> 01:34:56,930
AI generated video,
3360
01:34:57,150 --> 01:34:59,570
but AI assisted editing, um,
3361
01:34:59,690 --> 01:35:01,850
like Premiere Pro, but with uh,
3362
01:35:01,990 --> 01:35:04,150
something like AI could help you,
3363
01:35:04,550 --> 01:35:06,970
this would be a great option
3364
01:35:07,030 --> 01:35:07,450
for you.
3365
01:35:07,950 --> 01:35:08,170
Um,
3366
01:35:08,530 --> 01:35:10,330
so love to get your feedback
3367
01:35:10,390 --> 01:35:10,710
on it.
3368
01:35:11,030 --> 01:35:11,110
Uh,
3369
01:35:11,290 --> 01:35:13,890
it's growing pretty rapidly and
3370
01:35:14,610 --> 01:35:14,830
there
3371
01:35:15,730 --> 01:35:16,930
are a lot of new features
3372
01:35:16,970 --> 01:35:17,190
for it.
3373
01:35:17,230 --> 01:35:17,730
So if you have a lot
3374
01:35:17,770 --> 01:35:18,910
of foot feature,
3375
01:35:19,170 --> 01:35:19,650
if you have a lot of
3376
01:35:19,710 --> 01:35:20,710
footage that you're looking to edit,
3377
01:35:21,010 --> 01:35:23,810
check out novacut AI.
3378
01:35:24,070 --> 01:35:25,330
But other than that,
3379
01:35:26,330 --> 01:35:27,250
thank you for listening.
3380
01:35:27,330 --> 01:35:28,110
We love you all.
3381
01:35:28,350 --> 01:35:30,270
Really appreciate you taking the time
3382
01:35:30,310 --> 01:35:31,290
out of your day to listen
3383
01:35:31,330 --> 01:35:34,870
to our podcast. Thanks. Thanks again.