TRANSCRIPT 🔗
1
00:00:00,399 --> 00:00:02,320
[Speaker 0]
Hello, and welcome to the Geek News Central
2
00:00:02,320 --> 00:00:05,680
[Speaker 0]
podcast, episode eighteen eighty. I'm your host tonight,
3
00:00:05,759 --> 00:00:07,359
[Speaker 0]
Ray Cochran. I have a pretty good episode
4
00:00:07,359 --> 00:00:09,679
[Speaker 0]
lined up for you folks. Leading off a
5
00:00:09,679 --> 00:00:12,639
[Speaker 0]
free tool called REA, short for reverse engineer
6
00:00:12,639 --> 00:00:16,175
[Speaker 0]
anything, blew past two hundred thousand of bookmarks
7
00:00:16,175 --> 00:00:18,494
[Speaker 0]
on GitHub in a matter of days. It
8
00:00:18,494 --> 00:00:20,575
[Speaker 0]
points an AI agent at a finished program
9
00:00:20,575 --> 00:00:22,654
[Speaker 0]
and has it figure out how the program
10
00:00:22,654 --> 00:00:25,295
[Speaker 0]
works with none of the original code in
11
00:00:25,295 --> 00:00:27,215
[Speaker 0]
hand. For a day, it was the single
12
00:00:27,215 --> 00:00:29,920
[Speaker 0]
most popular project on all of GitHub. And
13
00:00:29,920 --> 00:00:31,680
[Speaker 0]
a lot of the flashiest demos flying around
14
00:00:31,680 --> 00:00:33,280
[Speaker 0]
under its name turned out not to use
15
00:00:33,280 --> 00:00:37,120
[Speaker 0]
it at all. In tonight's headlines, Google's newest
16
00:00:37,120 --> 00:00:39,680
[Speaker 0]
frontier model is going out to cyber defenders
17
00:00:39,680 --> 00:00:43,274
[Speaker 0]
first with its cyber guardrails switched off. Google
18
00:00:43,274 --> 00:00:45,195
[Speaker 0]
also put its own AI chips in orbit,
19
00:00:45,195 --> 00:00:47,195
[Speaker 0]
a first test of whether data centers could
20
00:00:47,195 --> 00:00:50,394
[Speaker 0]
someday run-in space. Amazon wants an AI agent
21
00:00:50,394 --> 00:00:53,515
[Speaker 0]
checking your cloud setup every week. Anthropic says
22
00:00:53,515 --> 00:00:55,515
[Speaker 0]
Claude turned up a gene system nobody had
23
00:00:55,515 --> 00:00:58,094
[Speaker 0]
noticed, and nobody yet knows what it does.
24
00:00:58,650 --> 00:01:01,210
[Speaker 0]
Researchers trained software to guess better than chance
25
00:01:01,210 --> 00:01:03,290
[Speaker 0]
what a herd of elephants is up to
26
00:01:03,290 --> 00:01:06,729
[Speaker 0]
from their overlapping rumbles. Meta laid out a
27
00:01:06,729 --> 00:01:09,210
[Speaker 0]
privacy system for AI glasses that isn't on
28
00:01:09,210 --> 00:01:12,705
[Speaker 0]
any glasses yet. Apple is reportedly building a
29
00:01:12,705 --> 00:01:15,045
[Speaker 0]
security camera that never shows you any video.
30
00:01:15,665 --> 00:01:17,985
[Speaker 0]
Apple creator studio picks up new tricks for
31
00:01:17,985 --> 00:01:20,625
[Speaker 0]
iPhone video. And finally, NASA has a free
32
00:01:20,625 --> 00:01:23,845
[Speaker 0]
movie night lined up. Let's get into it.
33
00:01:24,540 --> 00:01:26,140
[Speaker 0]
Tonight's show is brought to you by our
34
00:01:26,140 --> 00:01:29,420
[Speaker 0]
twenty year partner,
Load full transcript
GoDaddy. Exclusive deals and pricing
35
00:01:29,420 --> 00:01:31,660
[Speaker 0]
at geek new central dot com forward slash
36
00:01:31,660 --> 00:01:35,420
[Speaker 0]
GoDaddy. So a little personal update. It's been
37
00:01:35,420 --> 00:01:37,900
[Speaker 0]
a pretty hectic last couple of weeks. My
38
00:01:37,900 --> 00:01:39,660
[Speaker 0]
sister's been going through some drama and actually
39
00:01:39,660 --> 00:01:41,685
[Speaker 0]
ended up totaling her car on the way
40
00:01:41,685 --> 00:01:44,405
[Speaker 0]
to Seattle last month. So, she's in the
41
00:01:44,405 --> 00:01:46,565
[Speaker 0]
market for a new vehicle. Luckily, she had
42
00:01:46,565 --> 00:01:48,645
[Speaker 0]
a back available from family, but it's still
43
00:01:48,645 --> 00:01:50,485
[Speaker 0]
pretty big big loss on her end being
44
00:01:50,485 --> 00:01:52,645
[Speaker 0]
that, you know, she just did a expensive
45
00:01:52,645 --> 00:01:57,220
[Speaker 0]
maintenance on the car's turbocharger. So any recommend...
46
00:01:57,360 --> 00:01:59,200
[Speaker 0]
Recommends on vehicles as she shops around the
47
00:01:59,200 --> 00:02:00,880
[Speaker 0]
market? I think she's been looking at a
48
00:02:00,880 --> 00:02:03,520
[Speaker 0]
few particular brands, but, all the new cars
49
00:02:03,520 --> 00:02:06,479
[Speaker 0]
have some crazy features these days. My aunt's
50
00:02:06,479 --> 00:02:08,720
[Speaker 0]
f one fifty has a auto driving now
51
00:02:08,720 --> 00:02:11,834
[Speaker 0]
too. So, these car... High-tech cars, you know,
52
00:02:11,974 --> 00:02:14,775
[Speaker 0]
they're definitely getting all these new upgrades, and
53
00:02:14,775 --> 00:02:16,694
[Speaker 0]
they're becoming a lot more normalized too in
54
00:02:16,694 --> 00:02:18,614
[Speaker 0]
all the modern vehicles. So I'm sure whatever
55
00:02:18,614 --> 00:02:20,935
[Speaker 0]
she ends up, investing into, she'll be happy
56
00:02:20,935 --> 00:02:24,150
[Speaker 0]
about. But in other news, well, we're gearing
57
00:02:24,150 --> 00:02:25,590
[Speaker 0]
up to make our drive to Michigan here
58
00:02:25,590 --> 00:02:27,430
[Speaker 0]
in a couple days and get settled out
59
00:02:27,430 --> 00:02:29,430
[Speaker 0]
there, getting ready to spend some time in
60
00:02:29,430 --> 00:02:31,590
[Speaker 0]
the studio and really get things locked in.
61
00:02:31,590 --> 00:02:32,950
[Speaker 0]
And a lot of stuff has really been
62
00:02:32,950 --> 00:02:34,870
[Speaker 0]
back and forth, especially trying to schedule episodes
63
00:02:34,870 --> 00:02:37,190
[Speaker 0]
between Chris and myself. And the confusion that
64
00:02:37,190 --> 00:02:39,325
[Speaker 0]
comes with that has just been... It's been
65
00:02:39,325 --> 00:02:41,565
[Speaker 0]
tricky. So we're we're still trying to figure
66
00:02:41,565 --> 00:02:43,245
[Speaker 0]
all that out, but definitely looking to lock
67
00:02:43,245 --> 00:02:44,685
[Speaker 0]
it in once we get access to the
68
00:02:44,685 --> 00:02:50,685
[Speaker 0]
studio again. Otherwise, you know, I've been working
69
00:02:50,685 --> 00:02:52,670
[Speaker 0]
on some side projects, did a little update
70
00:02:52,670 --> 00:02:54,670
[Speaker 0]
to my side and thinking about to pronounce
71
00:02:54,670 --> 00:02:57,410
[Speaker 0]
personal blogs, maybe something for the, the drive.
72
00:02:57,790 --> 00:02:59,470
[Speaker 0]
Do... Hopefully, do a lot of photos while
73
00:02:59,470 --> 00:03:01,390
[Speaker 0]
we're driving out there and, you know, just
74
00:03:01,390 --> 00:03:03,150
[Speaker 0]
be able to post all those on the
75
00:03:03,150 --> 00:03:07,875
[Speaker 0]
site. But enough about me. Let's get in
76
00:03:07,875 --> 00:03:10,835
[Speaker 0]
the, into the stories tonight. The lead story
77
00:03:10,835 --> 00:03:12,515
[Speaker 0]
is a piece of software that took over
78
00:03:12,515 --> 00:03:15,335
[Speaker 0]
GitHub this past week. GitHub is where programmers
79
00:03:15,395 --> 00:03:17,555
[Speaker 0]
store and share their code, and it keeps
80
00:03:17,555 --> 00:03:19,315
[Speaker 0]
a running chart of the project people are
81
00:03:19,315 --> 00:03:21,920
[Speaker 0]
flagging as favorites, which you call stars. The
82
00:03:21,920 --> 00:03:24,560
[Speaker 0]
tool is named REA, short for reverse engineer
83
00:03:24,560 --> 00:03:27,600
[Speaker 0]
anything. On October sixth, it had under ten
84
00:03:27,600 --> 00:03:29,520
[Speaker 0]
thousand of those stars. But by the twelfth,
85
00:03:29,520 --> 00:03:31,440
[Speaker 0]
it cleared a hundred thousand and have to
86
00:03:31,440 --> 00:03:32,960
[Speaker 0]
top the whole chart for a whole day.
87
00:03:32,960 --> 00:03:35,200
[Speaker 0]
And as of today, October fourteenth, it's passed
88
00:03:35,200 --> 00:03:37,815
[Speaker 0]
two hundred forty thousand. That kind of jump
89
00:03:37,815 --> 00:03:39,894
[Speaker 0]
almost never happens. So what does it actually
90
00:03:39,894 --> 00:03:42,614
[Speaker 0]
do? REA aims an AI agent and finish
91
00:03:42,614 --> 00:03:44,534
[Speaker 0]
programs and has it worked out how that
92
00:03:44,534 --> 00:03:46,694
[Speaker 0]
program was built. The agents are tools like
93
00:03:46,694 --> 00:03:48,935
[Speaker 0]
Cloud Code or Cursor, the coding assistance a
94
00:03:48,935 --> 00:03:51,990
[Speaker 0]
lot of developers developers now run. Normally, this
95
00:03:51,990 --> 00:03:53,750
[Speaker 0]
kind of tear down is a specialist job
96
00:03:53,750 --> 00:03:55,830
[Speaker 0]
and a pretty slow one. The tech news
97
00:03:55,830 --> 00:03:57,670
[Speaker 0]
site, Decrypt, summed up why it struck a
98
00:03:57,670 --> 00:03:59,830
[Speaker 0]
nerve, quote, anything that runs on your own
99
00:03:59,830 --> 00:04:01,590
[Speaker 0]
device can now be read by an AI
100
00:04:01,590 --> 00:04:04,395
[Speaker 0]
agent with enough patience and tokens, which erodes
101
00:04:04,395 --> 00:04:07,935
[Speaker 0]
the oldest the oldest copy protection in
software,
102
00:04:08,235 --> 00:04:10,715
[Speaker 0]
the sheer pain of reading machine code, end
103
00:04:10,715 --> 00:04:13,595
[Speaker 0]
quote. And, yeah, that is definitely one of
104
00:04:13,595 --> 00:04:16,200
[Speaker 0]
the biggest reasons a lot of software hasn't
105
00:04:16,600 --> 00:04:19,160
[Speaker 0]
been, reverse engineered today is, yeah, because of
106
00:04:19,160 --> 00:04:21,340
[Speaker 0]
the the pain of reading machine code. But
107
00:04:22,200 --> 00:04:23,800
[Speaker 0]
to kind of see why that's hard, you
108
00:04:23,800 --> 00:04:25,560
[Speaker 0]
can kinda see how... You you start with
109
00:04:25,560 --> 00:04:27,640
[Speaker 0]
how software it gets made. Someone writes a
110
00:04:27,640 --> 00:04:30,460
[Speaker 0]
program in source code, a human readable recipe,
111
00:04:30,735 --> 00:04:32,655
[Speaker 0]
and then a tool called a compiler turns
112
00:04:32,655 --> 00:04:34,895
[Speaker 0]
that recipe into machine code. The long stream
113
00:04:34,895 --> 00:04:38,094
[Speaker 0]
of numbers, the processors actually runs. Download an
114
00:04:38,094 --> 00:04:39,615
[Speaker 0]
app and you get the main... You get
115
00:04:39,615 --> 00:04:42,094
[Speaker 0]
the machine code, not the recipe. So working
116
00:04:42,094 --> 00:04:43,854
[Speaker 0]
backward from the finished dish to the recipe
117
00:04:43,854 --> 00:04:46,450
[Speaker 0]
that made it, that's reverse engineering. And doing
118
00:04:46,450 --> 00:04:48,130
[Speaker 0]
that by hand is pretty brutal. There are
119
00:04:48,130 --> 00:04:50,690
[Speaker 0]
tools that help today. Two big ones are
120
00:04:50,690 --> 00:04:52,450
[Speaker 0]
Ghidra, which is free and was released by
121
00:04:52,450 --> 00:04:54,850
[Speaker 0]
the US National Security Agency back in twenty
122
00:04:54,850 --> 00:04:58,130
[Speaker 0]
nineteen, and IDA Pro, a commercial program from
123
00:04:58,130 --> 00:05:00,290
[Speaker 0]
Belgian from Belgian company that's been around since
124
00:05:00,290 --> 00:05:02,615
[Speaker 0]
the nineteen nineties and can run well over
125
00:05:02,615 --> 00:05:05,014
[Speaker 0]
a thousand dollars a year. Both include a
126
00:05:05,014 --> 00:05:07,495
[Speaker 0]
decompiler, which takes machine code and produces a
127
00:05:07,495 --> 00:05:09,974
[Speaker 0]
rough draft of the original recipe. The catch
128
00:05:09,974 --> 00:05:11,815
[Speaker 0]
is that the draft has the name stripped
129
00:05:11,815 --> 00:05:14,215
[Speaker 0]
out. Instead of a function helpfully called check
130
00:05:14,215 --> 00:05:16,535
[Speaker 0]
password, you get one called sub four zero
131
00:05:16,535 --> 00:05:18,980
[Speaker 0]
one thousand or some... You know, something random.
132
00:05:19,360 --> 00:05:20,880
[Speaker 0]
And there are no notes explaining any of
133
00:05:20,880 --> 00:05:22,560
[Speaker 0]
it. You know, a human expert can burn
134
00:05:22,560 --> 00:05:25,919
[Speaker 0]
hours untangling a single chunk of
that. But
135
00:05:25,919 --> 00:05:28,080
[Speaker 0]
REA itself is easy to misread. It isn't
136
00:05:28,080 --> 00:05:29,680
[Speaker 0]
one of those code reading tools, and it
137
00:05:29,680 --> 00:05:31,680
[Speaker 0]
does no decompiling of its own. It's a
138
00:05:31,680 --> 00:05:34,675
[Speaker 0]
connector running on a standard called MCP, short
139
00:05:34,675 --> 00:05:37,235
[Speaker 0]
for model context protocol. That's the plug that
140
00:05:37,235 --> 00:05:39,875
[Speaker 0]
Anthropic introduced in late twenty twenty four. And
141
00:05:39,875 --> 00:05:41,875
[Speaker 0]
so it basically allows an AI model to
142
00:05:41,875 --> 00:05:44,115
[Speaker 0]
reach out and operate other software. And both
143
00:05:44,115 --> 00:05:46,775
[Speaker 0]
OpenAI and Google adopted it in the following
144
00:05:46,835 --> 00:05:50,409
[Speaker 0]
year. REA uses that plug to write an...
145
00:05:50,409 --> 00:05:52,729
[Speaker 0]
To wire an AI agent into Ghidra or
146
00:05:52,729 --> 00:05:54,490
[Speaker 0]
IDA. So you still need one of those
147
00:05:54,490 --> 00:05:57,210
[Speaker 0]
installed. But the agent drives the tool, reads
148
00:05:57,210 --> 00:05:59,289
[Speaker 0]
back what it finds, and keeps asking follow-up
149
00:05:59,289 --> 00:06:02,905
[Speaker 0]
questions until it understands the program. And the
150
00:06:02,905 --> 00:06:04,905
[Speaker 0]
pitch is what made it spread, boiled down,
151
00:06:04,905 --> 00:06:06,745
[Speaker 0]
point your agent at an app whose features
152
00:06:06,745 --> 00:06:08,505
[Speaker 0]
you envy, and it digs to the program
153
00:06:08,505 --> 00:06:10,345
[Speaker 0]
and hands you back your own working version
154
00:06:10,345 --> 00:06:12,685
[Speaker 0]
of whatever it found. No source code required,
155
00:06:12,745 --> 00:06:14,850
[Speaker 0]
which is pretty cool. You know, online people
156
00:06:14,850 --> 00:06:16,450
[Speaker 0]
ran with it. There were a few clips
157
00:06:16,450 --> 00:06:18,850
[Speaker 0]
of decades old console games booted up on
158
00:06:18,850 --> 00:06:21,250
[Speaker 0]
an iPhone and free look alikes of pricey
159
00:06:21,250 --> 00:06:24,450
[Speaker 0]
creative software. One well known AI commentator, Matthew
160
00:06:24,450 --> 00:06:26,610
[Speaker 0]
Berman, posted a half joking line that just
161
00:06:26,610 --> 00:06:30,215
[Speaker 0]
read, quote, software is dead, end quote. Now
162
00:06:30,215 --> 00:06:32,135
[Speaker 0]
the reality checks, because there are quite a
163
00:06:32,135 --> 00:06:34,055
[Speaker 0]
few and they they do matter. The first
164
00:06:34,055 --> 00:06:36,215
[Speaker 0]
said this is not new technology. It is
165
00:06:36,215 --> 00:06:38,875
[Speaker 0]
a connector bolted onto tools that already existed.
166
00:06:39,335 --> 00:06:41,415
[Speaker 0]
A software blog called Mind Studio put it
167
00:06:41,415 --> 00:06:43,960
[Speaker 0]
plainly, quote, REA is a connector, not a
168
00:06:43,960 --> 00:06:47,879
[Speaker 0]
reverse engineering technology, end quote. People had already
169
00:06:47,879 --> 00:06:49,800
[Speaker 0]
wired AI agents into Ghidra and IDA back
170
00:06:49,800 --> 00:06:51,879
[Speaker 0]
in early twenty twenty five with free projects
171
00:06:51,879 --> 00:06:54,360
[Speaker 0]
like one called Ghidra MCP. And the company
172
00:06:54,360 --> 00:06:56,759
[Speaker 0]
that makes IDA Pro, Hexrays, shipped its own
173
00:06:56,759 --> 00:06:59,135
[Speaker 0]
official version of the same idea on September
174
00:06:59,135 --> 00:07:02,035
[Speaker 0]
twenty fifth, a couple of weeks before REA
175
00:07:02,175 --> 00:07:05,615
[Speaker 0]
ever went viral. And there's also what happens
176
00:07:05,615 --> 00:07:07,375
[Speaker 0]
when people sit down and actually test it.
177
00:07:07,375 --> 00:07:09,455
[Speaker 0]
A French reviewer who goes by Corbin ran
178
00:07:09,455 --> 00:07:11,535
[Speaker 0]
it on four jobs with Claude driving Ghidra
179
00:07:11,535 --> 00:07:13,750
[Speaker 0]
underneath. It worked. Then he did the same
180
00:07:13,750 --> 00:07:15,910
[Speaker 0]
jobs with a plain Claude and no REA
181
00:07:15,910 --> 00:07:17,910
[Speaker 0]
in the middle, and that worked too. His
182
00:07:17,910 --> 00:07:20,710
[Speaker 0]
takeaway quote, REA doesn't do any better than
183
00:07:20,710 --> 00:07:22,949
[Speaker 0]
a Claude you asked to install Ghidra and
184
00:07:22,949 --> 00:07:25,030
[Speaker 0]
use it, end quote. A tester at the
185
00:07:25,030 --> 00:07:27,164
[Speaker 0]
site, Hacker Noon, tried it on a real
186
00:07:27,164 --> 00:07:29,745
[Speaker 0]
app about a hundred megabytes. A quick scan
187
00:07:29,805 --> 00:07:31,965
[Speaker 0]
ran fine with deeper digging through a string
188
00:07:31,965 --> 00:07:35,164
[Speaker 0]
of errors. His overall read read, quote, stars
189
00:07:35,164 --> 00:07:37,004
[Speaker 0]
tell you people are curious, not whether the
190
00:07:37,004 --> 00:07:39,164
[Speaker 0]
tool matches a given job, end quote. So,
191
00:07:39,164 --> 00:07:40,870
[Speaker 0]
yeah, I I guess, you know, just a
192
00:07:40,870 --> 00:07:42,390
[Speaker 0]
caveat with all the bookmarks. You know, it
193
00:07:42,390 --> 00:07:44,150
[Speaker 0]
did... It got popular really quick, but I
194
00:07:44,150 --> 00:07:47,030
[Speaker 0]
think a bit of that is also because
195
00:07:47,030 --> 00:07:51,210
[Speaker 0]
of, you know, the name and, how clickbaity
196
00:07:51,270 --> 00:07:53,110
[Speaker 0]
that is. So the idea that you could
197
00:07:53,110 --> 00:07:55,365
[Speaker 0]
reverse engineer anything is a a a new
198
00:07:55,365 --> 00:07:56,805
[Speaker 0]
idea for a lot of people. So, of
199
00:07:56,805 --> 00:07:58,325
[Speaker 0]
course, you would get excited about it without
200
00:07:58,325 --> 00:08:00,565
[Speaker 0]
doing actual research into tools that already exist
201
00:08:00,565 --> 00:08:04,565
[Speaker 0]
for this. But strangest of all, a lot
202
00:08:04,565 --> 00:08:06,485
[Speaker 0]
of the jaw dropping demos going around under
203
00:08:06,485 --> 00:08:09,180
[Speaker 0]
REA's name were not actually built with REA.
204
00:08:09,500 --> 00:08:11,759
[Speaker 0]
The clearest example is a project called ArtCraft,
205
00:08:11,900 --> 00:08:13,919
[Speaker 0]
a set of free stand ins for Photoshop,
206
00:08:13,979 --> 00:08:15,660
[Speaker 0]
Word, and the rest of that paid software
207
00:08:15,660 --> 00:08:17,419
[Speaker 0]
set. You got held up all over the
208
00:08:17,419 --> 00:08:19,340
[Speaker 0]
place as proof of what REA can do.
209
00:08:19,340 --> 00:08:21,979
[Speaker 0]
Its founder, Brandon Thomas, said flatly that it
210
00:08:21,979 --> 00:08:24,220
[Speaker 0]
does not use REA at all. In his
211
00:08:24,220 --> 00:08:26,935
[Speaker 0]
words, quote, we do everything one hundred percent
212
00:08:26,935 --> 00:08:31,175
[Speaker 0]
clean room, end quote. And, additionally, quote, use
213
00:08:31,175 --> 00:08:35,015
[Speaker 0]
of Ghidra is disallowed, end quote. And that
214
00:08:35,015 --> 00:08:36,615
[Speaker 0]
phrase clean room is kind of the heart
215
00:08:36,615 --> 00:08:38,455
[Speaker 0]
of the legal question here, so it's kinda
216
00:08:38,455 --> 00:08:40,590
[Speaker 0]
worth a second. A clean room rebuild is
217
00:08:40,590 --> 00:08:42,510
[Speaker 0]
an old and perfectly legal way to copy
218
00:08:42,510 --> 00:08:44,750
[Speaker 0]
what a program does without copying how it
219
00:08:44,750 --> 00:08:47,310
[Speaker 0]
was written. One team studies the original and
220
00:08:47,310 --> 00:08:49,230
[Speaker 0]
writes a plain description of what it does.
221
00:08:49,230 --> 00:08:51,250
[Speaker 0]
A second team who never sees the original
222
00:08:51,310 --> 00:08:53,550
[Speaker 0]
builds a fresh version working only from that
223
00:08:53,550 --> 00:08:56,095
[Speaker 0]
description. The classic case is from nineteen eighty
224
00:08:56,095 --> 00:08:58,575
[Speaker 0]
four when Compaq cloned the startup software inside
225
00:08:58,575 --> 00:09:00,735
[Speaker 0]
the IBM PC this way and kicked off
226
00:09:00,735 --> 00:09:03,535
[Speaker 0]
the entire market of PC clones. Why does
227
00:09:03,535 --> 00:09:05,935
[Speaker 0]
that matter right now? Copyright protects a specific
228
00:09:05,935 --> 00:09:08,255
[Speaker 0]
code a person wrote, not the general idea
229
00:09:08,255 --> 00:09:10,000
[Speaker 0]
of what the program does. That is what
230
00:09:10,000 --> 00:09:12,720
[Speaker 0]
makes a clean copy... A clean room copy
231
00:09:12,720 --> 00:09:15,279
[Speaker 0]
legal. The hard part was always keeping the
232
00:09:15,279 --> 00:09:17,360
[Speaker 0]
two teams truly walled off, and that was
233
00:09:17,360 --> 00:09:19,440
[Speaker 0]
slow and expensive. A lawyer at the firm
234
00:09:19,440 --> 00:09:22,160
[Speaker 0]
Reed Smith, Gerard Donovan, wrote that AI collapses
235
00:09:22,160 --> 00:09:24,315
[Speaker 0]
that
cost. Quote, a process that might have
236
00:09:24,315 --> 00:09:26,315
[Speaker 0]
required a team of engineers and many months
237
00:09:26,315 --> 00:09:28,555
[Speaker 0]
now take one developer and AI agents less
238
00:09:28,555 --> 00:09:30,155
[Speaker 0]
than a week, end quote. And that's true
239
00:09:30,155 --> 00:09:33,755
[Speaker 0]
because to some degree, if you make it
240
00:09:33,755 --> 00:09:36,810
[Speaker 0]
so that this agent can't look at any
241
00:09:36,810 --> 00:09:38,650
[Speaker 0]
sort of references and it can only create
242
00:09:38,650 --> 00:09:40,570
[Speaker 0]
a tool based on what you're describing, then
243
00:09:40,570 --> 00:09:42,910
[Speaker 0]
you are technically creating a clean room copy.
244
00:09:44,570 --> 00:09:46,090
[Speaker 0]
But does raise kind of a question no
245
00:09:46,090 --> 00:09:48,335
[Speaker 0]
one's answered yet. With ArtCraft, the rule that
246
00:09:48,335 --> 00:09:50,255
[Speaker 0]
AI must never look at the original lives
247
00:09:50,255 --> 00:09:52,654
[Speaker 0]
in a text file the model simply trusted
248
00:09:52,654 --> 00:09:55,855
[Speaker 0]
to obey. And trust me, they disobey that
249
00:09:55,855 --> 00:09:58,335
[Speaker 0]
all the time. And back in the compact
250
00:09:58,335 --> 00:09:59,774
[Speaker 0]
days, the work of doing the rebuild had
251
00:09:59,774 --> 00:10:01,710
[Speaker 0]
genuinely never seen the original code. Now it's
252
00:10:01,710 --> 00:10:03,550
[Speaker 0]
an AI told in a config file what
253
00:10:03,550 --> 00:10:05,550
[Speaker 0]
to steer clear of and taken at its
254
00:10:05,550 --> 00:10:07,550
[Speaker 0]
word for it. Whether it's as good as
255
00:10:07,550 --> 00:10:09,630
[Speaker 0]
a real wall between two teams, nobody can
256
00:10:09,630 --> 00:10:12,430
[Speaker 0]
really say yet. And copyright isn't the only
257
00:10:12,430 --> 00:10:14,905
[Speaker 0]
hook. Donovan Poetens out that trade secret and
258
00:10:14,905 --> 00:10:17,465
[Speaker 0]
patent claims can land land on you even
259
00:10:17,465 --> 00:10:20,525
[Speaker 0]
when the copying itself comes out clean. So
260
00:10:21,225 --> 00:10:25,709
[Speaker 0]
definitely a really, really prickly... You know, I
261
00:10:25,709 --> 00:10:26,910
[Speaker 0]
don't know. How do you put it? Like,
262
00:10:26,910 --> 00:10:30,829
[Speaker 0]
a prickly field to kinda navigate through. But
263
00:10:30,829 --> 00:10:32,930
[Speaker 0]
the law underneath all this is a patchwork.
264
00:10:33,389 --> 00:10:35,470
[Speaker 0]
In the US, reverse engineering a program so
265
00:10:35,470 --> 00:10:36,910
[Speaker 0]
your own thing can work with it has
266
00:10:36,910 --> 00:10:39,245
[Speaker 0]
been ruled legal, but the fine print of
267
00:10:39,245 --> 00:10:41,885
[Speaker 0]
a software license can still forbid it. And
268
00:10:41,885 --> 00:10:44,605
[Speaker 0]
courts have enforced those bans. Europe runs roughly
269
00:10:44,605 --> 00:10:46,765
[Speaker 0]
the
other way. License there can't stop you
270
00:10:46,765 --> 00:10:49,085
[Speaker 0]
studying a program for compatibility, but you are
271
00:10:49,085 --> 00:10:50,845
[Speaker 0]
not allowed to turn around and use what
272
00:10:50,845 --> 00:10:53,959
[Speaker 0]
you learned to ship a copycat competitor. Enforcement
273
00:10:54,019 --> 00:10:56,920
[Speaker 0]
is already showing up. Decrypt reports that Activision
274
00:10:57,060 --> 00:10:59,060
[Speaker 0]
pulled footage of an AI rebuild Call of
275
00:10:59,060 --> 00:11:02,420
[Speaker 0]
Duty offline, and Bethesda sent a legal warning
276
00:11:02,420 --> 00:11:04,420
[Speaker 0]
over an AI made browser version of one
277
00:11:04,420 --> 00:11:07,915
[Speaker 0]
of its games. Step back from the copyright
278
00:11:07,915 --> 00:11:09,755
[Speaker 0]
file though... Or a copyright fight though, and
279
00:11:09,755 --> 00:11:11,595
[Speaker 0]
there's a completely different side to this, which
280
00:11:11,595 --> 00:11:14,475
[Speaker 0]
is the security. That same power, turning AI
281
00:11:14,475 --> 00:11:16,635
[Speaker 0]
loose to tear a program apart, is exactly
282
00:11:16,635 --> 00:11:18,875
[Speaker 0]
what you want for catching malware, the malicious
283
00:11:18,875 --> 00:11:21,295
[Speaker 0]
software that security team spend their lives fighting.
284
00:11:22,080 --> 00:11:24,400
[Speaker 0]
Microsoft built a research system last year called
285
00:11:24,400 --> 00:11:26,160
[Speaker 0]
Project I Earth that does this on its
286
00:11:26,160 --> 00:11:28,080
[Speaker 0]
own. It drives Ghidra and other tools to
287
00:11:28,080 --> 00:11:30,240
[Speaker 0]
pick up our suspicious file and test whether
288
00:11:30,240 --> 00:11:32,320
[Speaker 0]
it's dangerous, writing up its evidence as it
289
00:11:32,320 --> 00:11:35,605
[Speaker 0]
goes. And on one test set, Microsoft reported
290
00:11:35,605 --> 00:11:37,285
[Speaker 0]
caught about ninety percent of the bad files
291
00:11:37,285 --> 00:11:39,285
[Speaker 0]
while wrongly flagging only two percent of the
292
00:11:39,285 --> 00:11:41,605
[Speaker 0]
clean ones. A harder batch that its regular
293
00:11:41,605 --> 00:11:43,925
[Speaker 0]
tools couldn't crack though, it caught about a
294
00:11:43,925 --> 00:11:47,045
[Speaker 0]
quarter. And it cuts both ways. Defenders can
295
00:11:47,045 --> 00:11:49,045
[Speaker 0]
point AI at malware. Attackers can rig the
296
00:11:49,045 --> 00:11:52,690
[Speaker 0]
malware for AI. Researchers have started finding malicious
297
00:11:52,690 --> 00:11:54,690
[Speaker 0]
files with a block of hidden text buried
298
00:11:54,690 --> 00:11:57,649
[Speaker 0]
inside aimed not at the computer, but at
299
00:11:57,649 --> 00:12:00,610
[Speaker 0]
any AI sent to examine it, AKA prompt
300
00:12:00,610 --> 00:12:03,110
[Speaker 0]
injection. And we've talked about this a lot.
301
00:12:03,584 --> 00:12:05,504
[Speaker 0]
The text tries to talk the AI out
302
00:12:05,504 --> 00:12:07,745
[Speaker 0]
of its real job into reporting the file
303
00:12:07,745 --> 00:12:10,704
[Speaker 0]
as safe. The security company, Checkpoint, found an
304
00:12:10,704 --> 00:12:12,625
[Speaker 0]
early clumsy attempt at this in twenty twenty
305
00:12:12,625 --> 00:12:15,264
[Speaker 0]
five, and it flopped. The AI the AI
306
00:12:15,264 --> 00:12:17,270
[Speaker 0]
ignored it. A more polished one turned up
307
00:12:17,270 --> 00:12:19,430
[Speaker 0]
in twenty twenty six. A researcher at the
308
00:12:19,430 --> 00:12:21,510
[Speaker 0]
firm, Sentinel Labs, Phil Stokes, said the lesson
309
00:12:21,510 --> 00:12:23,589
[Speaker 0]
for anyone building these AI tools is to
310
00:12:23,589 --> 00:12:26,570
[Speaker 0]
treat the files they analyze, quote, as adversarial
311
00:12:26,950 --> 00:12:30,885
[Speaker 0]
input, never as instructions, end quote. And that's
312
00:12:30,885 --> 00:12:34,645
[Speaker 0]
a definitely... It's definitely good that you, you
313
00:12:34,645 --> 00:12:38,485
[Speaker 0]
know, you... To point specifically, you know, your
314
00:12:38,485 --> 00:12:41,285
[Speaker 0]
tools, you know, with, I don't know, exact
315
00:12:41,285 --> 00:12:43,225
[Speaker 0]
wording like this. You know, it's an adversary
316
00:12:43,445 --> 00:12:47,740
[Speaker 0]
input rather than instructions. Because if you don't,
317
00:12:47,740 --> 00:12:50,139
[Speaker 0]
you know, it just... It will assume that
318
00:12:50,139 --> 00:12:51,660
[Speaker 0]
anything is on the table. You know, you
319
00:12:51,660 --> 00:12:55,420
[Speaker 0]
you... There there are no... It relies heavily
320
00:12:55,420 --> 00:13:01,435
[Speaker 0]
on the semantics. So this is a pretty
321
00:13:01,435 --> 00:13:03,915
[Speaker 0]
interesting thing to kinda see continuing to blow
322
00:13:03,915 --> 00:13:05,995
[Speaker 0]
up. But, yeah, prompt injection is something we
323
00:13:05,995 --> 00:13:09,515
[Speaker 0]
talk about a lot on the show, and
324
00:13:09,515 --> 00:13:11,755
[Speaker 0]
it's just... It's not surprising to find these
325
00:13:11,755 --> 00:13:15,320
[Speaker 0]
in in these malicious software as a way
326
00:13:15,320 --> 00:13:19,160
[Speaker 0]
to ward off weaker AI tools. But as
327
00:13:19,160 --> 00:13:21,640
[Speaker 0]
an additional safety catch as well, these setups
328
00:13:21,640 --> 00:13:23,880
[Speaker 0]
generally are not sealed off from the rest
329
00:13:23,880 --> 00:13:26,680
[Speaker 0]
of your computer. REA's own security notes say,
330
00:13:26,680 --> 00:13:28,695
[Speaker 0]
quote, this is not a sand box, end
331
00:13:28,695 --> 00:13:31,095
[Speaker 0]
quote. So opening a program you don't trust,
332
00:13:31,095 --> 00:13:32,855
[Speaker 0]
hands it to tools running with your own
333
00:13:32,855 --> 00:13:35,655
[Speaker 0]
access to the machine. Even Hexrays, the IDA
334
00:13:35,655 --> 00:13:37,895
[Speaker 0]
company tells people to run this kind of
335
00:13:37,895 --> 00:13:40,535
[Speaker 0]
thing inside an isolated environment when they're poking
336
00:13:40,535 --> 00:13:42,775
[Speaker 0]
at something they don't trust. So if you're
337
00:13:42,775 --> 00:13:44,560
[Speaker 0]
just, you know, if you're just doing it
338
00:13:44,560 --> 00:13:46,160
[Speaker 0]
for fun and you're trying to, you know,
339
00:13:46,160 --> 00:13:48,240
[Speaker 0]
mess around and recreate one of your your
340
00:13:48,240 --> 00:13:50,720
[Speaker 0]
favorite tools or something for free, then that's
341
00:13:50,720 --> 00:13:54,240
[Speaker 0]
one thing. But a whole other thing would
342
00:13:54,240 --> 00:13:59,205
[Speaker 0]
be to try to do security work on
343
00:13:59,205 --> 00:14:02,645
[Speaker 0]
your regular computer trusting these tools. Please, please,
344
00:14:02,645 --> 00:14:05,785
[Speaker 0]
please, if you wanna start getting your security,
345
00:14:05,925 --> 00:14:07,685
[Speaker 0]
do the proper research on how to set
346
00:14:07,685 --> 00:14:11,000
[Speaker 0]
up your environment. Don't just open up Ida
347
00:14:11,060 --> 00:14:13,300
[Speaker 0]
and try to crack open, you know, some
348
00:14:13,300 --> 00:14:15,380
[Speaker 0]
sort of Trojan horse you found online. Please
349
00:14:15,380 --> 00:14:20,180
[Speaker 0]
do not do that. But, yeah, anyway, for...
350
00:14:20,180 --> 00:14:22,660
[Speaker 0]
As for who is behind REA, that part
351
00:14:22,660 --> 00:14:24,565
[Speaker 0]
is still pretty murky. Is the work of
352
00:14:24,565 --> 00:14:26,485
[Speaker 0]
an anonymous GitHub account that goes by the
353
00:14:26,485 --> 00:14:29,524
[Speaker 0]
name Morluto. There's no real name attached and
354
00:14:29,524 --> 00:14:32,245
[Speaker 0]
no interviews. The project finally started back in
355
00:14:32,245 --> 00:14:34,425
[Speaker 0]
April as a small tool for one program
356
00:14:34,485 --> 00:14:36,870
[Speaker 0]
and only snowballed this month. You can even
357
00:14:36,870 --> 00:14:39,190
[Speaker 0]
watch the hype outrun the author. Partway through
358
00:14:39,190 --> 00:14:40,710
[Speaker 0]
the search, they had to post a notice
359
00:14:40,710 --> 00:14:42,730
[Speaker 0]
that the project has not launched any cryptocurrency
360
00:14:42,950 --> 00:14:45,590
[Speaker 0]
after people apparently started trading tokens using its
361
00:14:45,590 --> 00:14:48,150
[Speaker 0]
name. That is crazy. They also added a
362
00:14:48,150 --> 00:14:50,575
[Speaker 0]
formal disclaimer that it won't help you strip
363
00:14:50,575 --> 00:14:53,135
[Speaker 0]
copy protection, cheat in games, or copy some
364
00:14:53,135 --> 00:14:56,255
[Speaker 0]
commercial software. So where does this all land?
365
00:14:56,255 --> 00:14:58,334
[Speaker 0]
The capability is real and the tools under
366
00:14:58,334 --> 00:15:00,334
[Speaker 0]
are genuinely old news, but what's new is
367
00:15:00,334 --> 00:15:02,334
[Speaker 0]
the two things arriving at once. The packaging
368
00:15:02,334 --> 00:15:03,870
[Speaker 0]
that put all of it in one command
369
00:15:04,029 --> 00:15:05,870
[Speaker 0]
away and the fact that AI models have
370
00:15:05,870 --> 00:15:07,790
[Speaker 0]
finally gotten good at reading that stripped down
371
00:15:07,790 --> 00:15:09,790
[Speaker 0]
lameness code and explaining it back in plain
372
00:15:09,790 --> 00:15:12,210
[Speaker 0]
English. Go back to that line from decrypt.
373
00:15:12,350 --> 00:15:14,830
[Speaker 0]
The oldest lock on software was never really
374
00:15:14,830 --> 00:15:16,990
[Speaker 0]
the law. It was how painful machine code
375
00:15:16,990 --> 00:15:19,275
[Speaker 0]
is to read. And that's... That that that
376
00:15:19,275 --> 00:15:23,035
[Speaker 0]
lock is coming loose, guys. So something to
377
00:15:23,035 --> 00:15:24,235
[Speaker 0]
keep an eye on for sure in the
378
00:15:24,235 --> 00:15:27,115
[Speaker 0]
upcoming years. And, you know, more and more
379
00:15:27,115 --> 00:15:29,355
[Speaker 0]
people are building their own personalized tools using
380
00:15:29,355 --> 00:15:33,020
[Speaker 0]
these... Using, you know, AI helpers or AI
381
00:15:33,020 --> 00:15:37,100
[Speaker 0]
tools, like Cloud or ChatTBT or codecs or
382
00:15:37,100 --> 00:15:43,020
[Speaker 0]
whatever you're using. So that is really exciting
383
00:15:43,020 --> 00:15:46,014
[Speaker 0]
to see. I definitely am still concerned regarding
384
00:15:46,014 --> 00:15:47,935
[Speaker 0]
the security angle of all this, you know,
385
00:15:47,935 --> 00:15:51,615
[Speaker 0]
in terms of people not understanding exactly what
386
00:15:51,615 --> 00:15:54,654
[Speaker 0]
they're putting out there in the world in
387
00:15:54,654 --> 00:15:59,140
[Speaker 0]
terms of software and exactly, you know, how
388
00:15:59,140 --> 00:16:01,060
[Speaker 0]
security is or how... You know, if they're
389
00:16:01,060 --> 00:16:04,020
[Speaker 0]
following all the privacy laws. So just something
390
00:16:04,020 --> 00:16:05,780
[Speaker 0]
to think about if you're messing around with
391
00:16:05,780 --> 00:16:09,380
[Speaker 0]
those tools, looking to OWASP's... OWASP security measures,
392
00:16:09,380 --> 00:16:12,395
[Speaker 0]
looking to privacy laws, make sure if you're
393
00:16:12,395 --> 00:16:14,075
[Speaker 0]
putting tools out there for people to use
394
00:16:14,075 --> 00:16:16,235
[Speaker 0]
that you are protecting yourself from a legal
395
00:16:16,235 --> 00:16:21,515
[Speaker 0]
angle. But up next is that tonight's headlines
396
00:16:21,515 --> 00:16:23,774
[Speaker 0]
right after this quick word from our sponsor.
397
00:16:24,200 --> 00:16:25,800
[Speaker 0]
Every time one of you uses our GoDaddy
398
00:16:25,800 --> 00:16:27,320
[Speaker 0]
link, it's like writing a check to this
399
00:16:27,320 --> 00:16:30,040
[Speaker 0]
show. That's not an exaggeration. That is literally
400
00:16:30,040 --> 00:16:32,600
[Speaker 0]
how this podcast stays alive. Twenty plus years
401
00:16:32,600 --> 00:16:35,400
[Speaker 0]
of independent content made possible by you and
402
00:16:35,400 --> 00:16:37,815
[Speaker 0]
by GoDaddy. We use their products every day
403
00:16:37,815 --> 00:16:40,615
[Speaker 0]
at Geek News Central. Domains, hosting, email, the
404
00:16:40,615 --> 00:16:42,615
[Speaker 0]
whole thing. And after two decades, you don't
405
00:16:42,615 --> 00:16:45,175
[Speaker 0]
keep recommending something unless it actually works. If
406
00:16:45,175 --> 00:16:46,935
[Speaker 0]
you need anything web related, whether it's a
407
00:16:46,935 --> 00:16:49,335
[Speaker 0]
domain, hosting, or a website builder, go through
408
00:16:49,335 --> 00:16:51,755
[Speaker 0]
geek new central dot com forward slash godaddy.
409
00:16:51,920 --> 00:16:53,700
[Speaker 0]
Share the link with friends and family too.
410
00:16:53,760 --> 00:16:56,340
[Speaker 0]
Geek new central dot com forward slash godaddy.
411
00:16:58,640 --> 00:17:01,520
[Speaker 0]
Alright. Let's get into the headlines. We're gonna
412
00:17:01,520 --> 00:17:03,840
[Speaker 0]
start with, Google again, but this one comes
413
00:17:03,840 --> 00:17:07,665
[Speaker 0]
from Google DeepMind posted on. September thirtieth, Google
414
00:17:07,665 --> 00:17:10,065
[Speaker 0]
announced Gemini four, Argon, the first model in
415
00:17:10,065 --> 00:17:13,905
[Speaker 0]
its Gemini four generation. Google calls it its
416
00:17:13,905 --> 00:17:16,465
[Speaker 0]
new frontier model, which is which is the
417
00:17:16,465 --> 00:17:19,025
[Speaker 0]
industry's term for the most capable system a
418
00:17:19,025 --> 00:17:21,210
[Speaker 0]
lab has built so far. You can't actually
419
00:17:21,210 --> 00:17:23,850
[Speaker 0]
use it yet, though. Google says it's rolling
420
00:17:23,850 --> 00:17:25,690
[Speaker 0]
Argon out in phases, and right now it's
421
00:17:25,690 --> 00:17:27,929
[Speaker 0]
limited to Google's own teams and a vetted
422
00:17:27,929 --> 00:17:30,730
[Speaker 0]
group of cyber defense organizations in something called
423
00:17:30,730 --> 00:17:34,555
[Speaker 0]
the Fair Win program. Next in line are
424
00:17:34,555 --> 00:17:37,515
[Speaker 0]
paid developers and subscribers to Google AI Ultra,
425
00:17:37,515 --> 00:17:39,915
[Speaker 0]
its top price consumer plan, but Google hasn't
426
00:17:39,915 --> 00:17:42,495
[Speaker 0]
given a date for any of that. Fairwind
427
00:17:42,555 --> 00:17:44,955
[Speaker 0]
started in early September, and Google says more
428
00:17:44,955 --> 00:17:47,860
[Speaker 0]
than six hundred fifty organizations have joined. Its
429
00:17:47,860 --> 00:17:50,659
[Speaker 0]
priority is governments, national cyber agencies, and the
430
00:17:50,659 --> 00:17:53,860
[Speaker 0]
companies running things like hospitals, power, phone networks,
431
00:17:53,860 --> 00:17:56,580
[Speaker 0]
and banks. Numbers go through background checks and
432
00:17:56,580 --> 00:17:59,299
[Speaker 0]
only their security staff get access. Google's own
433
00:17:59,299 --> 00:18:01,140
[Speaker 0]
staff and those vetted defenders are getting a
434
00:18:01,140 --> 00:18:04,445
[Speaker 0]
version of Argon released, quote, without cyber guardrails,
435
00:18:04,665 --> 00:18:07,065
[Speaker 0]
end quote. Google says it's... It wants the
436
00:18:07,065 --> 00:18:09,225
[Speaker 0]
people guarding critical systems to have the model's
437
00:18:09,225 --> 00:18:11,465
[Speaker 0]
full strength for finding the pack... Finding and
438
00:18:11,465 --> 00:18:14,424
[Speaker 0]
patching security holes, which it's... Yeah. That's fair,
439
00:18:14,424 --> 00:18:16,424
[Speaker 0]
but I would assume that they're still being
440
00:18:16,424 --> 00:18:20,850
[Speaker 0]
heavily monitored, I would assume. But Google's main
441
00:18:20,850 --> 00:18:22,930
[Speaker 0]
example came through a free program run by
442
00:18:22,930 --> 00:18:26,470
[Speaker 0]
the security company, Wiz. Their Google says app...
443
00:18:26,530 --> 00:18:29,330
[Speaker 0]
Argon spotted a critical hole in widely used
444
00:18:29,330 --> 00:18:32,870
[Speaker 0]
hospital software that let people's personal data exposed.
445
00:18:33,455 --> 00:18:35,935
[Speaker 0]
Wow. So a whole older top models had
446
00:18:35,935 --> 00:18:39,135
[Speaker 0]
missed. But Google doesn't name the software or
447
00:18:39,135 --> 00:18:41,615
[Speaker 0]
say whether it's been fixed. And also a
448
00:18:41,615 --> 00:18:45,135
[Speaker 0]
caveat is Google owns Wiz. They they closed
449
00:18:45,135 --> 00:18:47,535
[Speaker 0]
that thirty two billion dollar deal in March,
450
00:18:47,535 --> 00:18:50,040
[Speaker 0]
and the post doesn't really mention it, but
451
00:18:50,040 --> 00:18:51,559
[Speaker 0]
it's clear why they have they have the
452
00:18:51,559 --> 00:18:55,400
[Speaker 0]
software in their hands. Google also posted benchmark
453
00:18:55,400 --> 00:18:57,799
[Speaker 0]
scores. There are the standard... Or these are
454
00:18:57,799 --> 00:19:00,280
[Speaker 0]
the standardized tests labs use to compare models,
455
00:19:00,280 --> 00:19:03,265
[Speaker 0]
and Google's own table has Argonne first or
456
00:19:03,265 --> 00:19:05,925
[Speaker 0]
Tide on fourteen of the nineteen it lists.
457
00:19:06,145 --> 00:19:08,245
[Speaker 0]
One of them comes from Zapier, the automation
458
00:19:08,304 --> 00:19:11,265
[Speaker 0]
company. It tests whether an AI can run
459
00:19:11,265 --> 00:19:13,665
[Speaker 0]
a business workflow from start to finish. And
460
00:19:13,665 --> 00:19:16,385
[Speaker 0]
Argonne scored about fifty one with Anthropic's Claude
461
00:19:16,385 --> 00:19:18,799
[Speaker 0]
Opus five point five, Nex at about forty
462
00:19:18,799 --> 00:19:22,000
[Speaker 0]
three. The same table shows Argonne trailing on
463
00:19:22,000 --> 00:19:24,480
[Speaker 0]
the other five, mostly tough tests of AI
464
00:19:24,480 --> 00:19:26,720
[Speaker 0]
agents working at a command line or operating
465
00:19:26,720 -->
00:19:30,000
[Speaker 0]
computer on their own. The, the text of
466
00:19:30,000 --> 00:19:31,760
[Speaker 0]
the post only talks about the wins, but
467
00:19:31,760 --> 00:19:33,525
[Speaker 0]
all these numbers come from Google's post and
468
00:19:33,525 --> 00:19:35,125
[Speaker 0]
the public can't test on the model yet
469
00:19:35,125 --> 00:19:36,485
[Speaker 0]
to check them. But know that these are
470
00:19:36,485 --> 00:19:40,325
[Speaker 0]
coming out for you a, Google Ultra users,
471
00:19:40,325 --> 00:19:44,725
[Speaker 0]
AI Ultra. Alright. More from Google tonight. This
472
00:19:44,725 --> 00:19:47,220
[Speaker 0]
one's from the blog dot Google posted by
473
00:19:47,220 --> 00:19:49,940
[Speaker 0]
Travis Beals. He's the Google senior director who
474
00:19:49,940 --> 00:19:53,780
[Speaker 0]
leads something called Project SunCatcher. On October first,
475
00:19:53,780 --> 00:19:56,200
[Speaker 0]
a SpaceX Falcon nine carried a small prototype
476
00:19:56,260 --> 00:19:58,820
[Speaker 0]
satellite into orbit for Google, and it's packed
477
00:19:58,820 --> 00:20:01,795
[Speaker 0]
with Google's own AI chips. On launch day,
478
00:20:01,795 --> 00:20:03,715
[Speaker 0]
Beal said that the satellite was talking to
479
00:20:03,715 --> 00:20:06,515
[Speaker 0]
the ground and running normally. Suncatchers what Google
480
00:20:06,515 --> 00:20:09,395
[Speaker 0]
calls a research moonshot. The long term idea
481
00:20:09,395 --> 00:20:11,155
[Speaker 0]
is to run AI computing on fleets of
482
00:20:11,155 --> 00:20:13,875
[Speaker 0]
satellites powered directly by the sun. It comes
483
00:20:13,875 --> 00:20:17,559
[Speaker 0]
down to electricity. The International Energy Agency says
484
00:20:17,559 --> 00:20:19,480
[Speaker 0]
the data centers built for AI use about
485
00:20:19,480 --> 00:20:21,480
[Speaker 0]
fifty percent more power in twenty twenty five
486
00:20:21,480 --> 00:20:23,960
[Speaker 0]
than the year before. Google's paper says that
487
00:20:23,960 --> 00:20:25,880
[Speaker 0]
in the right orbit, a solar panel can
488
00:20:25,880 --> 00:20:28,520
[Speaker 0]
collect collect up to eight times as much
489
00:20:28,520 --> 00:20:30,539
[Speaker 0]
energy in a year as one of the...
490
00:20:31,785 --> 00:20:34,105
[Speaker 0]
As one on the ground across most of
491
00:20:34,105 --> 00:20:36,525
[Speaker 0]
the US or Europe, which is impressive. But,
492
00:20:37,065 --> 00:20:40,825
[Speaker 0]
I mean, I think it's fairly well known
493
00:20:40,825 --> 00:20:44,265
[Speaker 0]
that the lifespan on those solar panels aren't
494
00:20:44,265 --> 00:20:47,640
[Speaker 0]
very amazing. So I wonder how, you know,
495
00:20:47,640 --> 00:20:50,600
[Speaker 0]
they'll combat that in space. I'm sure they
496
00:20:50,600 --> 00:20:55,720
[Speaker 0]
have methodologies for that, but very interesting. But
497
00:20:55,720 --> 00:20:58,040
[Speaker 0]
the... This first satellite carries four of Google's
498
00:20:58,040 --> 00:20:59,880
[Speaker 0]
AI chips according to NPR and the New
499
00:20:59,880 --> 00:21:02,365
[Speaker 0]
York Times. Google built it with Planet, the
500
00:21:02,365 --> 00:21:04,765
[Speaker 0]
satellite imaging company. And it's there to check
501
00:21:04,765 --> 00:21:07,085
[Speaker 0]
whether the hardware survives the shaking of launch,
502
00:21:07,085 --> 00:21:09,245
[Speaker 0]
the radiation, and the heat. Heat is the
503
00:21:09,245 --> 00:21:11,965
[Speaker 0]
tough one because the space, there's no air
504
00:21:11,965 --> 00:21:14,770
[Speaker 0]
to carry it away. Oh, because in space,
505
00:21:14,770 --> 00:21:16,630
[Speaker 0]
there's no air to carry it away. Ah,
506
00:21:16,850 --> 00:21:18,850
[Speaker 0]
I thought heat would be less of a
507
00:21:18,850 --> 00:21:24,930
[Speaker 0]
problem, but that's interesting. A Google spokesperson told
508
00:21:24,930 --> 00:21:26,945
[Speaker 0]
CNN the satellite needs a cool down break
509
00:21:26,945 --> 00:21:33,665
[Speaker 0]
about every twenty minutes. Interesting. The peer reviewed
510
00:21:33,665 --> 00:21:35,745
[Speaker 0]
paper behind the project came out the same
511
00:21:35,745 --> 00:21:39,770
[Speaker 0]
day in the energy journal Joule. That's j...
512
00:21:40,310 --> 00:21:43,910
[Speaker 0]
Well, Joule. Like the energy. Like the... I
513
00:21:43,910 --> 00:21:46,070
[Speaker 0]
don't know. It makes sense if you if
514
00:21:46,070 --> 00:21:48,550
[Speaker 0]
you know energies. It's Joules. You know? But
515
00:21:48,550 --> 00:21:50,790
[Speaker 0]
Google redid its radiation test for it, and
516
00:21:50,790 --> 00:21:52,550
[Speaker 0]
the chip's rate of silent errors came out
517
00:21:52,550 --> 00:21:55,085
[Speaker 0]
about three times higher than Google's first reported.
518
00:21:55,085 --> 00:21:57,085
[Speaker 0]
A silent error is a wrong answer that
519
00:21:57,085 --> 00:21:59,325
[Speaker 0]
comes out looking perfectly normal. The new rate
520
00:21:59,325 --> 00:22:01,905
[Speaker 0]
is roughly one for every three million answers.
521
00:22:02,685 --> 00:22:04,445
[Speaker 0]
The the authors expect that to cause a
522
00:22:04,445 --> 00:22:06,525
[Speaker 0]
little trouble with a... When a finished AI
523
00:22:06,525 --> 00:22:08,930
[Speaker 0]
model is answering questions, But for training new
524
00:22:08,930 --> 00:22:11,190
[Speaker 0]
models, they say it's still an open question.
525
00:22:11,730 --> 00:22:13,810
[Speaker 0]
Then there's the money. The paper target for
526
00:22:13,810 --> 00:22:15,890
[Speaker 0]
the launch is about two hundred dollars a
527
00:22:15,890 --> 00:22:18,690
[Speaker 0]
kilogram, which it projected could could happen around
528
00:22:18,690 --> 00:22:21,650
[Speaker 0]
twenty thirty five. Today, a Falcon nine runs
529
00:22:21,650 --> 00:22:24,265
[Speaker 0]
about thirty six hundred dollars a kilogram. So
530
00:22:24,265 --> 00:22:26,505
[Speaker 0]
TechCrunch pointed out the papers on launch map
531
00:22:26,505 --> 00:22:29,785
[Speaker 0]
needs SpaceX's giant Starship to fly around a
532
00:22:29,785 --> 00:22:32,125
[Speaker 0]
hundred eighty times a year for a decade.
533
00:22:32,425 --> 00:22:35,645
[Speaker 0]
Starship's busiest year so far is five flights.
534
00:22:39,480 --> 00:22:42,840
[Speaker 0]
Wow. That's insane. Google itself isn't promising anything
535
00:22:42,840 --> 00:22:45,560
[Speaker 0]
soon, but, you know, technology moves fast. I
536
00:22:45,560 --> 00:22:47,400
[Speaker 0]
I wouldn't be shocked to see this stuff
537
00:22:47,400 --> 00:22:50,040
[Speaker 0]
happening quicker than we think. Beals told NPR,
538
00:22:50,040 --> 00:22:52,265
[Speaker 0]
quote, I don't see this being something where
539
00:22:52,265 --> 00:22:54,105
[Speaker 0]
it's cheaper to do this in the next
540
00:22:54,105 --> 00:22:55,785
[Speaker 0]
five years. I think it'll take longer than
541
00:22:55,785 --> 00:22:58,025
[Speaker 0]
that, end quote. Google plans to run this
542
00:22:58,025 --> 00:23:00,345
[Speaker 0]
satellite for about a year. Next, Google says
543
00:23:00,345 --> 00:23:02,424
[Speaker 0]
it'll fly two satellites together in twenty twenty
544
00:23:02,424 --> 00:23:04,065
[Speaker 0]
seven to test the lasers that would let
545
00:23:04,065 --> 00:23:06,730
[Speaker 0]
a fleet work like one computer. Wow. So,
546
00:23:06,730 --> 00:23:08,990
[Speaker 0]
yeah, lots of tests underway in that regard.
547
00:23:10,169 --> 00:23:12,570
[Speaker 0]
Up next, Amazon Web Services opened a preview
548
00:23:12,570 --> 00:23:15,450
[Speaker 0]
of something called the well architected agent. This
549
00:23:15,450 --> 00:23:18,010
[Speaker 0]
is from the AWS news blog posted October
550
00:23:18,010 --> 00:23:21,395
[Speaker 0]
first by Chany Yoon. It's an AI that
551
00:23:21,395 --> 00:23:23,955
[Speaker 0]
looks over a company setup on Amazon cloud
552
00:23:23,955 --> 00:23:26,595
[Speaker 0]
or Amazon's cloud and tells them what to
553
00:23:26,595 --> 00:23:29,475
[Speaker 0]
fix. For years, AWS has published something called
554
00:23:29,475 --> 00:23:33,040
[Speaker 0]
the well architected... Well architected framework. It's basically
555
00:23:33,040 --> 00:23:35,280
[Speaker 0]
Amazon's rule book for building on its cloud.
556
00:23:35,280 --> 00:23:38,720
[Speaker 0]
Split into six categories like security, reliability, and
557
00:23:38,720 --> 00:23:41,200
[Speaker 0]
cost. The traditional way to check yourself against
558
00:23:41,200 --> 00:23:43,840
[Speaker 0]
it is a questionnaire where a team answers
559
00:23:43,840 --> 00:23:45,680
[Speaker 0]
questions about one system and gets back a
560
00:23:45,680 --> 00:23:49,044
[Speaker 0]
list of risks. AWS has offered a free
561
00:23:49,044 --> 00:23:51,684
[Speaker 0]
tool for that since twenty eighteen. That reviews
562
00:23:51,684 --> 00:23:54,404
[Speaker 0]
a snapshot of one moment AWS pitches the
563
00:23:54,404 --> 00:23:57,765
[Speaker 0]
agent as it's something that keeps checking. According
564
00:23:57,765 --> 00:24:00,164
[Speaker 0]
to AWS, you tell the agent which accounts
565
00:24:00,164 --> 00:24:02,610
[Speaker 0]
to watch and what you care about written
566
00:24:02,610 --> 00:24:06,290
[Speaker 0]
as plain sentences. AWS's own examples are goals
567
00:24:06,290 --> 00:24:09,010
[Speaker 0]
like getting rid of public storage buckets or
568
00:24:09,010 --> 00:24:11,330
[Speaker 0]
cutting the infrastructure bill by twenty percent a
569
00:24:11,330 --> 00:24:14,385
[Speaker 0]
year. The agent then rescans every week, ranks
570
00:24:14,385 --> 00:24:16,725
[Speaker 0]
what it finds against those goals, and attaches
571
00:24:17,105 --> 00:24:21,105
[Speaker 0]
step by steps fix instructions. AWS says AWS
572
00:24:21,105 --> 00:24:23,425
[Speaker 0]
says it covers best practice checks across more
573
00:24:23,425 --> 00:24:28,419
[Speaker 0]
than sixty five of its services. Documentation says
574
00:24:28,419 --> 00:24:30,820
[Speaker 0]
anything the AI writes stays a suggestion until
575
00:24:30,820 --> 00:24:33,240
[Speaker 0]
a person reviews it and runs it. AWS
576
00:24:33,299 --> 00:24:35,559
[Speaker 0]
even puts a warning in the announcement, quote,
577
00:24:35,779 --> 00:24:38,900
[Speaker 0]
generative AI capabilities produce this recommendation, which may
578
00:24:38,900 --> 00:24:42,715
[Speaker 0]
contain errors or incomplete information, end quote. One
579
00:24:42,715 --> 00:24:45,754
[Speaker 0]
exception is some routine findings from AWS's older
580
00:24:45,754 --> 00:24:48,394
[Speaker 0]
automated checker. Those come with prebuilt scripts that
581
00:24:48,394 --> 00:24:49,995
[Speaker 0]
can run with your approval or on a
582
00:24:49,995 --> 00:24:53,034
[Speaker 0]
schedule. Setup means giving it a read only
583
00:24:53,034 --> 00:24:55,615
[Speaker 0]
access to every account you want scanned. AWS
584
00:24:55,835 --> 00:24:58,809
[Speaker 0]
docs say those access roles can't change anything,
585
00:24:58,809 --> 00:25:01,610
[Speaker 0]
and AWS also suggests giving the agent its
586
00:25:01,610 --> 00:25:04,250
[Speaker 0]
own account, one with nothing live running in
587
00:25:04,250 --> 00:25:08,350
[Speaker 0]
it. AWS says first results show up within
588
00:25:08,409 --> 00:25:11,205
[Speaker 0]
a day or two. Wow. So it's a
589
00:25:11,205 --> 00:25:13,845
[Speaker 0]
public preview for now hosted in three US
590
00:25:13,845 --> 00:25:15,684
[Speaker 0]
regions. To use it, you need a paid
591
00:25:15,684 --> 00:25:17,924
[Speaker 0]
AWS support plan at the business support plus
592
00:25:17,924 --> 00:25:20,164
[Speaker 0]
level or higher, and the tier starts at
593
00:25:20,164 --> 00:25:22,965
[Speaker 0]
twenty nine dollars a month per account. AWS
594
00:25:22,965 --> 00:25:25,205
[Speaker 0]
hasn't said what the agent itself will cost.
595
00:25:25,205 --> 00:25:27,390
[Speaker 0]
It also covers only four of the framework's
596
00:25:27,450 --> 00:25:31,390
[Speaker 0]
six categories. Operational excellence and sustainability aren't included.
597
00:25:33,850 --> 00:25:36,330
[Speaker 0]
Now this next one's from, Anthropic, the company
598
00:25:36,330 --> 00:25:38,730
[Speaker 0]
makes Claude. It went up September twenty third
599
00:25:38,730 --> 00:25:40,490
[Speaker 0]
on anthropic dot com, and it comes out
600
00:25:40,490 --> 00:25:42,784
[Speaker 0]
of the life sciences lab the company opened
601
00:25:42,784 --> 00:25:45,904
[Speaker 0]
this spring in the Bay Area. Tropic says
602
00:25:45,904 --> 00:25:48,225
[Speaker 0]
a team of clogged agents spotted something in
603
00:25:48,225 --> 00:25:50,485
[Speaker 0]
viral DNA that appears to have gone unnoticed
604
00:25:50,544 --> 00:25:55,080
[Speaker 0]
until now. The viruses here are phages. They
605
00:25:55,080 --> 00:25:58,120
[Speaker 0]
infect bacteria, not people, and the bacteria fight
606
00:25:58,120 --> 00:26:00,120
[Speaker 0]
back and the most famous defense they have
607
00:26:00,120 --> 00:26:03,000
[Speaker 0]
is CRISPR. And before it became a gene
608
00:26:03,000 --> 00:26:05,900
[Speaker 0]
editing tool, CRISPR was kind of immune memory...
609
00:26:07,045 --> 00:26:09,625
[Speaker 0]
Was a kind of immune memory. The bacterium
610
00:26:09,685 --> 00:26:12,905
[Speaker 0]
stores snippets of viral DNA between short repeated
611
00:26:12,965 --> 00:26:15,845
[Speaker 0]
sequences and uses copies of those snippets to
612
00:26:15,845 --> 00:26:17,605
[Speaker 0]
reorganize and cut up the virus if it
613
00:26:17,605 --> 00:26:19,945
[Speaker 0]
comes back. Oh, interesting. I didn't know that.
614
00:26:21,230 --> 00:26:23,870
[Speaker 0]
What Claude found has the same repeating look.
615
00:26:23,870 --> 00:26:26,210
[Speaker 0]
It's a long row of short DNA sequences
616
00:26:26,270 --> 00:26:28,990
[Speaker 0]
repeated at every... Or at even spacing sitting
617
00:26:28,990 --> 00:26:31,390
[Speaker 0]
right beside the gene that makes an enzyme
618
00:26:31,390 --> 00:26:35,935
[Speaker 0]
called a reverse transcriptase... Reverse transcriptase. And that's
619
00:26:35,935 --> 00:26:38,575
[Speaker 0]
an endemic... Enzyme that copies RNA back into
620
00:26:38,575 --> 00:26:42,655
[Speaker 0]
DNA, and HIV uses one. There's a partner
621
00:26:42,655 --> 00:26:45,375
[Speaker 0]
gene on the other side too. Anthropic named
622
00:26:45,375 --> 00:26:49,635
[Speaker 0]
the whole thing array associated reverse transcriptases. Wow.
623
00:26:50,440 --> 00:26:53,000
[Speaker 0]
The resemblance is, to CRISPR stopped at the
624
00:26:53,000 --> 00:26:56,120
[Speaker 0]
layout, though. Tropic's own technical report says there
625
00:26:56,120 --> 00:26:58,280
[Speaker 0]
are no CRISPR genes anywhere near these arrays,
626
00:26:58,280 --> 00:27:00,600
[Speaker 0]
and the stretches between the repeats run several
627
00:27:00,600 --> 00:27:03,480
[Speaker 0]
times longer. The team's working guess is that
628
00:27:03,480 --> 00:27:06,255
[Speaker 0]
it's related to a different material defense called
629
00:27:06,255 --> 00:27:10,275
[Speaker 0]
a retron or a retron. In a retron,
630
00:27:10,335 --> 00:27:13,135
[Speaker 0]
a infected cell sacrifices itself to protect its
631
00:27:13,135 --> 00:27:15,775
[Speaker 0]
neighbors. What this new system actually does is
632
00:27:15,775 --> 00:27:18,210
[Speaker 0]
still unknown, and Anthropic says so. I don't
633
00:27:18,210 --> 00:27:20,370
[Speaker 0]
know. At least that's what they're saying. But
634
00:27:20,370 --> 00:27:23,410
[Speaker 0]
anthropic scientists gave Claude one broad assignment to
635
00:27:23,410 --> 00:27:25,410
[Speaker 0]
look for new systems built around this kind
636
00:27:25,410 --> 00:27:27,830
[Speaker 0]
of enzyme, and the assignment never mentioned repeats.
637
00:27:28,130 --> 00:27:30,530
[Speaker 0]
About nine hundred fifty Claude sessions then searched
638
00:27:30,530 --> 00:27:34,585
[Speaker 0]
a database of roughly sixteen billion proteins over
639
00:27:34,585 --> 00:27:37,145
[Speaker 0]
about twenty one hours. And one of those
640
00:27:37,145 --> 00:27:39,705
[Speaker 0]
agents spotted the repeats just by reading raw
641
00:27:39,705 --> 00:27:42,265
[Speaker 0]
DNA. It then checked the published research for
642
00:27:42,265 --> 00:27:44,525
[Speaker 0]
anything similar before writing it up for anthropic
643
00:27:44,585 --> 00:27:47,740
[Speaker 0]
scientists. The one outside voice in the post
644
00:27:47,740 --> 00:27:51,180
[Speaker 0]
is Feng Zhang. He's a CRISPR gene editing
645
00:27:51,180 --> 00:27:54,320
[Speaker 0]
pioneer based at MIT and the Broad Institute.
646
00:27:56,540 --> 00:27:58,460
[Speaker 0]
Darvick says he reviewed an early copy of
647
00:27:58,460 --> 00:28:00,705
[Speaker 0]
the paper. He called the finding, quote, genuinely
648
00:28:00,764 --> 00:28:03,005
[Speaker 0]
intriguing, end quote, and said it's, quote, full
649
00:28:03,005 --> 00:28:07,245
[Speaker 0]
merits... Or merits further investigation, end quote. Most
650
00:28:07,245 --> 00:28:09,404
[Speaker 0]
of the caveats come straight from Anthropic's own
651
00:28:09,404 --> 00:28:11,485
[Speaker 0]
report, but it isn't peer reviewed, and it's
652
00:28:11,485 --> 00:28:13,910
[Speaker 0]
a PDF on Anthropic's own site. The lab
653
00:28:13,910 --> 00:28:16,549
[Speaker 0]
work so far is one experiment showing the
654
00:28:16,549 --> 00:28:19,190
[Speaker 0]
repeat region gets copied into short pieces of
655
00:28:19,190 --> 00:28:21,669
[Speaker 0]
RNA. And in ten more runs of the
656
00:28:21,669 --> 00:28:23,850
[Speaker 0]
same search, none of them found the array.
657
00:28:23,910 --> 00:28:26,630
[Speaker 0]
So I'm sure there's a deeper dive that's
658
00:28:26,630 --> 00:28:29,804
[Speaker 0]
gonna happen with that here soon, but kinda
659
00:28:29,804 --> 00:28:32,924
[Speaker 0]
cool to see that type of gene work
660
00:28:32,924 --> 00:28:37,505
[Speaker 0]
being, you know, kinda processed by these AIs.
661
00:28:37,644 --> 00:28:40,684
[Speaker 0]
But I, yeah, I I don't know. I
662
00:28:40,684 --> 00:28:44,300
[Speaker 0]
wonder if... Well, it's it's hard to assume.
663
00:28:44,300 --> 00:28:45,740
[Speaker 0]
And I I I would think that it's
664
00:28:45,740 --> 00:28:48,140
[Speaker 0]
hard to do such things programmatically because there
665
00:28:48,140 --> 00:28:52,700
[Speaker 0]
are so many, you know, variations. So, I
666
00:28:52,700 --> 00:28:56,125
[Speaker 0]
really don't know. That's kinda cool. Up next
667
00:28:56,125 --> 00:28:59,805
[Speaker 0]
is about elephants, interestingly enough. It's from Science
668
00:28:59,805 --> 00:29:02,045
[Speaker 0]
dot org reported by Julia Vaz on a
669
00:29:02,045 --> 00:29:04,785
[Speaker 0]
study out October first in the journal Scientific
670
00:29:04,845 --> 00:29:07,885
[Speaker 0]
Reports. The biggest group on the team is
671
00:29:07,885 --> 00:29:11,130
[Speaker 0]
from the Earth Species Project. That's a nonprofit
672
00:29:11,190 --> 00:29:13,610
[Speaker 0]
that uses machine learning to study animal communication.
673
00:29:14,150 --> 00:29:16,230
[Speaker 0]
They worked with elephant researchers who have been
674
00:29:16,230 --> 00:29:18,870
[Speaker 0]
recording the herds at Mbeusele National Park in
675
00:29:18,870 --> 00:29:23,110
[Speaker 0]
Kenya for decades. They call the... The the
676
00:29:23,110 --> 00:29:25,215
[Speaker 0]
call they studied is the rumble, the most
677
00:29:25,215 --> 00:29:29,135
[Speaker 0]
common thing elephants say. Study coauthor Micky Pardoz
678
00:29:29,135 --> 00:29:32,495
[Speaker 0]
told Scientific American, it sounds like, quote, an
679
00:29:32,495 --> 00:29:35,855
[Speaker 0]
idling truck engine, end quote. Its bass pitch
680
00:29:35,855 --> 00:29:37,830
[Speaker 0]
is too low for human ears, and the
681
00:29:37,830 --> 00:29:40,070
[Speaker 0]
low ones are... And the loud ones carry
682
00:29:40,070 --> 00:29:44,149
[Speaker 0]
about two kilometers to other elephants. Wow. The
683
00:29:44,149 --> 00:29:46,470
[Speaker 0]
catch is that elephants talk over each other
684
00:29:46,470 --> 00:29:48,390
[Speaker 0]
a lot. The paper says about sixty percent
685
00:29:48,390 --> 00:29:50,809
[Speaker 0]
of rumbles overlap with at least one another
686
00:29:51,405 --> 00:29:54,365
[Speaker 0]
elephant calling at the same time. Researchers call
687
00:29:54,365 --> 00:29:56,845
[Speaker 0]
that the cocktail party problem, picking out one
688
00:29:56,845 --> 00:29:59,405
[Speaker 0]
voice out of a crowd. Since the voices
689
00:29:59,405 --> 00:30:02,045
[Speaker 0]
couldn't be separated, the paper says most earlier
690
00:30:02,045 --> 00:30:05,110
[Speaker 0]
work simply left those overlapping calls out. So
691
00:30:05,110 --> 00:30:07,110
[Speaker 0]
this team skipped the untangling and treated each
692
00:30:07,110 --> 00:30:09,429
[Speaker 0]
course as one sound. They used nine hundred
693
00:30:09,429 --> 00:30:11,750
[Speaker 0]
two recordings made between nineteen eighty six and
694
00:30:11,750 --> 00:30:14,070
[Speaker 0]
twenty twenty two. Most of them from one
695
00:30:14,070 --> 00:30:16,730
[Speaker 0]
famous family led by a matriarch named echo.
696
00:30:17,705 --> 00:30:20,665
[Speaker 0]
An experienced observer had labeled every recording in
697
00:30:20,665 --> 00:30:22,505
[Speaker 0]
the field with what the herd was up
698
00:30:22,505 --> 00:30:24,665
[Speaker 0]
to at the moment. There were six situations,
699
00:30:24,665 --> 00:30:26,425
[Speaker 0]
things like calming a calf, trying to get
700
00:30:26,425 --> 00:30:29,085
[Speaker 0]
the group moving, or reuniting after a separation.
701
00:30:29,899 --> 00:30:31,580
[Speaker 0]
Then they train a computer model to name
702
00:30:31,580 --> 00:30:33,659
[Speaker 0]
the situation from the sound alone. It's a
703
00:30:33,659 --> 00:30:36,220
[Speaker 0]
basic statistical model, and it doesn't translate anything.
704
00:30:36,220 --> 00:30:38,059
[Speaker 0]
It just sorts each recording into one of
705
00:30:38,059 --> 00:30:40,620
[Speaker 0]
those six situations. They even had to shift
706
00:30:40,620 --> 00:30:42,779
[Speaker 0]
the audio up eight times in pitch so
707
00:30:42,779 --> 00:30:45,605
[Speaker 0]
standard audio to software could handle it. Wow.
708
00:30:47,105 --> 00:30:49,505
[Speaker 0]
It'd be random guessing though. On single rumbles
709
00:30:49,505 --> 00:30:51,505
[Speaker 0]
from echo's family, it got about forty four
710
00:30:51,505 --> 00:30:53,905
[Speaker 0]
percent right where chance would land about nineteen
711
00:30:53,905 --> 00:30:58,039
[Speaker 0]
percent. Trained only on single rumbles. And then
712
00:30:58,039 --> 00:31:00,280
[Speaker 0]
tested on choruses, it dropped about twenty five
713
00:31:00,280 --> 00:31:02,840
[Speaker 0]
percent against the thirteen percent per chance. But
714
00:31:02,840 --> 00:31:05,640
[Speaker 0]
still work just less less accurately. In the
715
00:31:05,640 --> 00:31:09,020
[Speaker 0]
paper's own words, quote, more rumbles were classified
716
00:31:09,080 --> 00:31:12,825
[Speaker 0]
incorrectly than correctly, end quote. The paper has
717
00:31:12,825 --> 00:31:15,664
[Speaker 0]
a second finding within an, or within an
718
00:31:15,664 --> 00:31:18,304
[Speaker 0]
exchange, the elephant's calls grew more alike as
719
00:31:18,304 --> 00:31:20,405
[Speaker 0]
it went on, which is... That's kinda interesting,
720
00:31:20,544 --> 00:31:23,505
[Speaker 0]
which researchers call vocal convergence. It showed up
721
00:31:23,505 --> 00:31:26,150
[Speaker 0]
in three of the six situations, greetings among
722
00:31:26,150 --> 00:31:28,470
[Speaker 0]
them, and the author suspect it's mainly a
723
00:31:28,470 --> 00:31:31,110
[Speaker 0]
coordination tool. Parrots, dolphins, and people do the
724
00:31:31,110 --> 00:31:35,030
[Speaker 0]
same thing. For an outside view, Scientific American
725
00:31:35,030 --> 00:31:39,695
[Speaker 0]
went to Eric Kirschenbaum. He's a Cambridge zoologist
726
00:31:39,755 --> 00:31:41,515
[Speaker 0]
who wasn't involved in the study. He told
727
00:31:41,515 --> 00:31:44,235
[Speaker 0]
the magazine, quote, this study emphasizes how vital
728
00:31:44,235 --> 00:31:46,955
[Speaker 0]
it is to pair new computational analysis with
729
00:31:46,955 --> 00:31:50,174
[Speaker 0]
real behavioral observations in the wild, end quote.
730
00:31:50,560 --> 00:31:52,880
[Speaker 0]
Science also posted video of one of the
731
00:31:52,880 --> 00:31:55,440
[Speaker 0]
young elephants in the data. Her name is
732
00:31:55,440 --> 00:31:59,280
[Speaker 0]
Nastya. This clip shows Nastya and her mother
733
00:31:59,280 --> 00:32:01,360
[Speaker 0]
back together after a long stretch of calls.
734
00:32:01,360 --> 00:32:05,735
[Speaker 0]
That's very cute. Up next, Meta has a
735
00:32:05,735 --> 00:32:07,815
[Speaker 0]
privacy system for its AI glasses that it
736
00:32:07,815 --> 00:32:11,115
[Speaker 0]
calls private processing. This is from Meta's engineering
737
00:32:11,174 --> 00:32:13,735
[Speaker 0]
blog at engineering dot f b dot com.
738
00:32:13,735 --> 00:32:15,415
[Speaker 0]
It went up September twenty third, the same
739
00:32:15,415 --> 00:32:18,090
[Speaker 0]
day as Meta's big connect event. Meta says
740
00:32:18,090 --> 00:32:20,650
[Speaker 0]
making a call and other quick requests already
741
00:32:20,650 --> 00:32:24,030
[Speaker 0]
run on the glasses themselves. Bigger AI jobs
742
00:32:24,170 --> 00:32:26,490
[Speaker 0]
go up to Meta's servers, and Meta's examples
743
00:32:26,490 --> 00:32:29,850
[Speaker 0]
include translation and summing up a chat. Meta's
744
00:32:29,850 --> 00:32:31,610
[Speaker 0]
claim is that this system walls that data
745
00:32:31,610 --> 00:32:34,865
[Speaker 0]
off from everyone, Meta included. Meta first built
746
00:32:34,865 --> 00:32:37,024
[Speaker 0]
private processing in twenty twenty five for AI
747
00:32:37,024 --> 00:32:40,565
[Speaker 0]
features in WhatsApp. The glasses version adds memory.
748
00:32:40,625 --> 00:32:42,465
[Speaker 0]
It's meant to remember things about you over
749
00:32:42,465 --> 00:32:45,664
[Speaker 0]
time, but Meta says that the user holds
750
00:32:45,664 --> 00:32:48,700
[Speaker 0]
the keys that locks those memory, and they're
751
00:32:48,700 --> 00:32:52,540
[Speaker 0]
locked before Meta stores them. Oh, interesting. So
752
00:32:52,540 --> 00:32:54,700
[Speaker 0]
Meta says the AI runs inside what's called
753
00:32:54,700 --> 00:32:57,420
[Speaker 0]
a trusted execution environment. It's a walled up
754
00:32:57,420 --> 00:33:00,300
[Speaker 0]
zone inside a service processor where the chip
755
00:33:00,300 --> 00:33:02,765
[Speaker 0]
itself encrypts a memory, so even the people
756
00:33:02,765 --> 00:33:05,164
[Speaker 0]
running the server can't look in. Before your
757
00:33:05,164 --> 00:33:07,325
[Speaker 0]
device sends anything, it checks a signed proof
758
00:33:07,325 --> 00:33:09,245
[Speaker 0]
of exactly what software is running in there
759
00:33:09,245 --> 00:33:11,645
[Speaker 0]
against a public record. A failed check means
760
00:33:11,645 --> 00:33:15,150
[Speaker 0]
nothing leaves your device according to Meta. Meta
761
00:33:15,150 --> 00:33:17,230
[Speaker 0]
also says your request traveled through an outside
762
00:33:17,230 --> 00:33:20,350
[Speaker 0]
relay company like CloudFlare or Fastly, so its
763
00:33:20,350 --> 00:33:22,990
[Speaker 0]
servers never see your Internet address. And it
764
00:33:22,990 --> 00:33:25,470
[Speaker 0]
says that inside that zone, engineers get no
765
00:33:25,470 --> 00:33:27,630
[Speaker 0]
debugging tools and no logs of what goes
766
00:33:27,630 --> 00:33:30,265
[Speaker 0]
in and out. They only see overall stats
767
00:33:30,265 --> 00:33:33,165
[Speaker 0]
like processor and memory use. According to Wired,
768
00:33:33,305 --> 00:33:35,065
[Speaker 0]
none of this is on any glasses yet.
769
00:33:35,305 --> 00:33:37,705
[Speaker 0]
Pritam Shah is a director of engineering at
770
00:33:37,705 --> 00:33:39,385
[Speaker 0]
Meta, and he told the magazine there's no
771
00:33:39,385 --> 00:33:41,625
[Speaker 0]
timeline. On the features in the post, he
772
00:33:41,625 --> 00:33:44,200
[Speaker 0]
said, quote, these are examples, end quote. He
773
00:33:44,200 --> 00:33:45,880
[Speaker 0]
said the reason for going public this early
774
00:33:45,880 --> 00:33:48,120
[Speaker 0]
is to bring in security researchers and an
775
00:33:48,120 --> 00:33:51,640
[Speaker 0]
outside auditor. That's kinda cool. There are limits
776
00:33:51,640 --> 00:33:53,880
[Speaker 0]
even once it ships. This protects the wearer's
777
00:33:53,880 --> 00:33:56,280
[Speaker 0]
data from meta. The people around you being
778
00:33:56,280 --> 00:33:58,605
[Speaker 0]
filmed are a separate problem, and Wired points
779
00:33:58,605 --> 00:34:01,325
[Speaker 0]
out that the backlash these glasses have already
780
00:34:01,325 --> 00:34:04,845
[Speaker 0]
faced. Free Arrays at Privacy Guides wrote that
781
00:34:04,845 --> 00:34:08,205
[Speaker 0]
confidential computing, quote, does offer some protection, end
782
00:34:08,205 --> 00:34:10,525
[Speaker 0]
quote, but the data, quote, must still be
783
00:34:10,525 --> 00:34:13,290
[Speaker 0]
decrypted at some point, end quote. But Meta's
784
00:34:13,290 --> 00:34:15,210
[Speaker 0]
own account in its twenty twenty five write
785
00:34:15,210 --> 00:34:17,450
[Speaker 0]
up, security researchers have broken through these chip
786
00:34:17,450 --> 00:34:22,570
[Speaker 0]
load protections more than once. So the ongoing
787
00:34:22,570 --> 00:34:27,525
[Speaker 0]
battle against privacy, I feel like, definitely... I
788
00:34:27,525 --> 00:34:29,684
[Speaker 0]
I like to see that they're continuing to
789
00:34:29,684 --> 00:34:32,325
[Speaker 0]
try to find new ways
to continue to
790
00:34:32,325 --> 00:34:34,405
[Speaker 0]
protect people and that people are holding them
791
00:34:34,405 --> 00:34:37,444
[Speaker 0]
true to that. And, hopefully, with, you know,
792
00:34:37,444 --> 00:34:40,800
[Speaker 0]
outside auditors pushing them as well, they'll develop
793
00:34:40,800 --> 00:34:42,960
[Speaker 0]
something that, you know, is acceptable to most
794
00:34:42,960 --> 00:34:46,880
[Speaker 0]
people. I'm not sure. I'm not really sure
795
00:34:46,880 --> 00:34:49,520
[Speaker 0]
how I feel about the glasses in general
796
00:34:49,520 --> 00:34:55,305
[Speaker 0]
personally. Though, I don't know. I don't like
797
00:34:56,165 --> 00:34:58,165
[Speaker 0]
the recording being on all the time if
798
00:34:58,165 --> 00:34:59,685
[Speaker 0]
that's a thing. I I don't know about
799
00:34:59,685 --> 00:35:02,245
[Speaker 0]
that, but, you know, if that's all selective,
800
00:35:02,245 --> 00:35:03,445
[Speaker 0]
then sure. I don't mean... I don't I
801
00:35:03,445 --> 00:35:08,039
[Speaker 0]
don't know why not. But, yeah, let's move
802
00:35:08,039 --> 00:35:11,160
[Speaker 0]
on. Staying on the privacy route, this one
803
00:35:11,160 --> 00:35:13,400
[Speaker 0]
is a rumor about Apple, and it's on
804
00:35:13,400 --> 00:35:15,900
[Speaker 0]
the verge dot com written by Stevie Bonifield.
805
00:35:16,520 --> 00:35:19,625
[Speaker 0]
The source is Mark Gurman. He's Bloomberg's chief
806
00:35:19,625 --> 00:35:21,385
[Speaker 0]
Apple reporter, and he brought it up on
807
00:35:21,385 --> 00:35:23,805
[Speaker 0]
the first episode of his new Bloomberg podcast,
808
00:35:23,945 --> 00:35:27,145
[Speaker 0]
Power On. That episode came out on October
809
00:35:27,145 --> 00:35:29,385
[Speaker 0]
first. Gurman says Apple is working on a
810
00:35:29,385 --> 00:35:32,185
[Speaker 0]
home security camera that can't record video at
811
00:35:32,185 --> 00:35:34,920
[Speaker 0]
all. The code the code name is j
812
00:35:34,920 --> 00:35:37,080
[Speaker 0]
four fifty. The plan is a set of
813
00:35:37,080 --> 00:35:39,560
[Speaker 0]
small cameras spread around the house. Gurman says
814
00:35:39,560 --> 00:35:42,140
[Speaker 0]
each one, quote, looks like a metal giant
815
00:35:42,200 --> 00:35:45,240
[Speaker 0]
Chapstick, end quote. AI watches the scene and
816
00:35:45,240 --> 00:35:47,400
[Speaker 0]
in Gurman's words, quote, you're not getting anything
817
00:35:47,400 --> 00:35:49,815
[Speaker 0]
but text descriptions, end quote. He says it
818
00:35:49,815 --> 00:35:51,735
[Speaker 0]
will use facial recognition to tell you who
819
00:35:51,735 --> 00:35:53,655
[Speaker 0]
walked into which room or who left the
820
00:35:53,655 --> 00:35:55,415
[Speaker 0]
house and even where the dog is. He
821
00:35:55,415 --> 00:35:57,575
[Speaker 0]
also says the cameras are already in thousands
822
00:35:57,575 --> 00:36:00,395
[Speaker 0]
of Apple employees' homes for testing. Oh, wow.
823
00:36:00,535 --> 00:36:02,500
[Speaker 0]
It's It's part of Apple's bigger smart home
824
00:36:02,500 --> 00:36:04,660
[Speaker 0]
push. I knew it. I told you guys
825
00:36:04,660 --> 00:36:06,500
[Speaker 0]
this last year that they are gonna be
826
00:36:06,500 --> 00:36:08,260
[Speaker 0]
doing a big smart home push, and here
827
00:36:08,260 --> 00:36:09,940
[Speaker 0]
it is. They said we're seeing more of
828
00:36:09,940 --> 00:36:12,339
[Speaker 0]
it coming out. So the smart home isn't
829
00:36:12,339 --> 00:36:14,965
[Speaker 0]
dead. The first piece is a hub with
830
00:36:14,965 --> 00:36:18,005
[Speaker 0]
a roughly six inch touchscreen. The camera is
831
00:36:18,005 --> 00:36:20,405
[Speaker 0]
further out, and Gurman's earlier reporting put a
832
00:36:20,405 --> 00:36:24,485
[Speaker 0]
camera like this in twenty twenty seven. So
833
00:36:24,485 --> 00:36:27,470
[Speaker 0]
interesting. Gramer said Apple, quote, could pitch this
834
00:36:27,470 --> 00:36:29,550
[Speaker 0]
as an alternative to that very disturbing idea
835
00:36:29,550 --> 00:36:31,869
[Speaker 0]
of mass surveillance, end quote. He also raised
836
00:36:31,869 --> 00:36:33,790
[Speaker 0]
the downside himself. His example was a break
837
00:36:33,790 --> 00:36:35,630
[Speaker 0]
in where a written description gives the police
838
00:36:35,630 --> 00:36:37,310
[Speaker 0]
little to go on. As he put it,
839
00:36:37,310 --> 00:36:39,885
[Speaker 0]
quote, you need the actual tape, end quote.
840
00:36:39,965 --> 00:36:42,365
[Speaker 0]
According to Apple's own support pages, its current
841
00:36:42,365 --> 00:36:44,765
[Speaker 0]
camera or current home camera system can already
842
00:36:44,765 --> 00:36:47,405
[Speaker 0]
write searchable text descriptions of your rec... Recorded
843
00:36:47,405 --> 00:36:49,725
[Speaker 0]
clips if you pay for at least the
844
00:36:49,725 --> 00:36:52,285
[Speaker 0]
two terabyte iCloud plan. What would be new
845
00:36:52,285 --> 00:36:54,285
[Speaker 0]
is a security camera with no footage behind
846
00:36:54,285 --> 00:36:57,220
[Speaker 0]
those descriptions at all. Apple hasn't announced any
847
00:36:57,220 --> 00:37:00,440
[Speaker 0]
of this and hasn't commented. Since Gurman's episode,
848
00:37:00,500 --> 00:37:03,780
[Speaker 0]
Apple has actually held the event, on October
849
00:37:03,780 --> 00:37:06,260
[Speaker 0]
thirteenth at a show it called Welcome Home.
850
00:37:06,260 --> 00:37:09,000
[Speaker 0]
Apple unveiled the hub now officially named HomeView.
851
00:37:09,285 --> 00:37:11,684
[Speaker 0]
It's a seven inch touchscreen, a little bigger
852
00:37:11,684 --> 00:37:13,944
[Speaker 0]
than the rumors that had it. It recognizes
853
00:37:14,085 --> 00:37:15,684
[Speaker 0]
which member of the household is standing in
854
00:37:15,684 --> 00:37:17,125
[Speaker 0]
front of it and pulls up their own
855
00:37:17,125 --> 00:37:20,565
[Speaker 0]
messages, calendar, and reminders. CNBC reports it runs
856
00:37:20,565
--> 00:37:22,244
[Speaker 0]
on an a eighteen chip and leans hard
857
00:37:22,244 --> 00:37:25,010
[Speaker 0]
on Apple's new Siri. The... A new Apple
858
00:37:25,010 --> 00:37:27,330
[Speaker 0]
TV and a new HomePod mini came out
859
00:37:27,330 --> 00:37:29,650
[Speaker 0]
of the same event, and those two hit
860
00:37:29,650 --> 00:37:33,250
[Speaker 0]
stores October thirtieth with HomeView following around the
861
00:37:33,250 --> 00:37:35,890
[Speaker 0]
middle of November. The no video camera from
862
00:37:35,890 --> 00:37:38,115
[Speaker 0]
the rumor wasn't part of any of it.
863
00:37:38,195 --> 00:37:40,035
[Speaker 0]
My grims are reporting, and that one is
864
00:37:40,035 --> 00:37:42,195
[Speaker 0]
still a twenty twenty seven story. So this
865
00:37:42,195 --> 00:37:46,595
[Speaker 0]
new HomeView camera still has video. Okay. It's
866
00:37:46,595 --> 00:37:48,595
[Speaker 0]
a little bit of a runaround there. I
867
00:37:48,595 --> 00:37:51,075
[Speaker 0]
just, I I put this all together a
868
00:37:51,075 --> 00:37:53,155
[Speaker 0]
little bit sooner than the event and... But
869
00:37:53,155 --> 00:37:55,790
[Speaker 0]
I'm glad that the event actually came through
870
00:37:55,930 --> 00:37:58,350
[Speaker 0]
and we were able to cover the follow-up
871
00:37:58,410 --> 00:38:03,930
[Speaker 0]
at the same time. But another one on
872
00:38:03,930 --> 00:38:05,610
[Speaker 0]
the Apple side, this one is from Apple
873
00:38:05,610 --> 00:38:08,325
[Speaker 0]
Newsroom posted September twenty ninth. It's a round
874
00:38:08,325 --> 00:38:12,025
[Speaker 0]
of updates to Apple Creator Studio, Apple subscription
875
00:38:12,085 --> 00:38:14,965
[Speaker 0]
bundle of creative apps, and that includes Final
876
00:38:14,965 --> 00:38:17,045
[Speaker 0]
Cut Pro for video, Logic Pro for music,
877
00:38:17,045 --> 00:38:19,285
[Speaker 0]
and Pixel Matter Pro for images for twelve
878
00:38:19,285 --> 00:38:20,805
[Speaker 0]
ninety nine a month or a hundred twenty
879
00:38:20,805 --> 00:38:23,110
[Speaker 0]
nine dollars a year. Creator Studio was last
880
00:38:23,110 --> 00:38:25,350
[Speaker 0]
on the show in July on episodes eighteen
881
00:38:25,350 --> 00:38:28,630
[Speaker 0]
sixty eight. Apple's headline says these updates are
882
00:38:28,630 --> 00:38:31,510
[Speaker 0]
coming, but its release notes most... Or show
883
00:38:31,510 --> 00:38:33,850
[Speaker 0]
most of the apps updated that same day.
884
00:38:34,275 --> 00:38:36,515
[Speaker 0]
On the video side, the changes for cinematic
885
00:38:36,515 --> 00:38:38,755
[Speaker 0]
mode. That's the iPhone setting that blurs the
886
00:38:38,755 --> 00:38:40,674
[Speaker 0]
background like a film camera and lets you
887
00:38:40,674 --> 00:38:42,994
[Speaker 0]
change what's in focus after you shoot. Apple
888
00:38:42,994 --> 00:38:45,315
[Speaker 0]
says regular video from the new iPhone eighteen
889
00:38:45,315 --> 00:38:47,154
[Speaker 0]
Pro can get that treatment on the Mac
890
00:38:47,154 --> 00:38:49,974
[Speaker 0]
now. In Final Cut Pro and In motion.
891
00:38:50,330 --> 00:38:52,170
[Speaker 0]
Inside the editing timeline, you pick where the
892
00:38:52,170 --> 00:38:54,270
[Speaker 0]
focus lands and how much the background blurs.
893
00:38:54,330 --> 00:38:57,610
[Speaker 0]
Wow. According to Apple's release notes, the footage
894
00:38:57,610 --> 00:38:59,450
[Speaker 0]
has to be shot in Apple's free cut
895
00:38:59,450 --> 00:39:02,090
[Speaker 0]
final cut camera app on an iPhone eighteen
896
00:39:02,090 --> 00:39:04,165
[Speaker 0]
Pro. It only works on the Mac and
897
00:39:04,165 --> 00:39:09,925
[Speaker 0]
it needs the newest version, Mac OS. Apple's
898
00:39:09,925 --> 00:39:13,125
[Speaker 0]
export app compressor also got new preset... New
899
00:39:13,125 --> 00:39:15,845
[Speaker 0]
pre... Presets for social media. It also dropped
900
00:39:15,845 --> 00:39:18,425
[Speaker 0]
a feature that split big export jobs across
901
00:39:18,450 --> 00:39:20,690
[Speaker 0]
several Macs, and the press release doesn't mention
902
00:39:20,690 --> 00:39:24,930
[Speaker 0]
that. On the iPad, Apple says Pixel Matter
903
00:39:24,930 --> 00:39:27,329
[Speaker 0]
Pro now has the same cutout tools as
904
00:39:27,329 --> 00:39:29,490
[Speaker 0]
the Mac version. So you can lift a
905
00:39:29,490 --> 00:39:32,130
[Speaker 0]
subject off its background cleanly down to hair
906
00:39:32,130 --> 00:39:34,685
[Speaker 0]
and fur. Podcasters get a new set of
907
00:39:34,685 --> 00:39:37,005
[Speaker 0]
cover art templates too, and Apple says they
908
00:39:37,005 --> 00:39:39,405
[Speaker 0]
already fit what the big podcast apps ask
909
00:39:39,405 --> 00:39:43,485
[Speaker 0]
for. Wow. I'll have to make sure I
910
00:39:43,485 --> 00:39:47,250
[Speaker 0]
bring this up to, my coworkers. But Keynote
911
00:39:47,310 --> 00:39:49,390
[Speaker 0]
slide generator can now add a generated picture
912
00:39:49,390 --> 00:39:51,570
[Speaker 0]
wherever a slide has room for one too.
913
00:39:51,710 --> 00:39:54,290
[Speaker 0]
It's still labeled beta and it's for subscribers,
914
00:39:54,510 --> 00:39:56,590
[Speaker 0]
but Apple support page says the feature runs
915
00:39:56,590 --> 00:39:59,550
[Speaker 0]
on OpenAI's model, not Apple's own. And the
916
00:39:59,550 --> 00:40:02,130
[Speaker 0]
suite's other image tools run on Google Cloud.
917
00:40:02,645 --> 00:40:05,285
[Speaker 0]
Apple says that you you send... Well, what
918
00:40:05,285 --> 00:40:07,365
[Speaker 0]
you send can't be used to train either
919
00:40:07,365 --> 00:40:10,025
[Speaker 0]
company's AI. Okay. That's a good caveat there.
920
00:40:11,444 --> 00:40:13,305
[Speaker 0]
But if you change his land for everybody,
921
00:40:14,165 --> 00:40:16,165
[Speaker 0]
subscriber or not, you can now request access
922
00:40:16,165 --> 00:40:18,619
[Speaker 0]
to someone else's shared document instead of waiting
923
00:40:18,619 --> 00:40:21,039
[Speaker 0]
for an invite. Pages on the Mac gets
924
00:40:21,339 --> 00:40:25,099
[Speaker 0]
a reading focused layout, and separately, Apple's free
925
00:40:25,099 --> 00:40:28,079
[Speaker 0]
form whiteboard app picks up dark mode folders
926
00:40:28,140 --> 00:40:30,799
[Speaker 0]
and AI writing tools that work on handwritten
927
00:40:31,180 --> 00:40:36,515
[Speaker 0]
notes. Cool. I have one final headline for
928
00:40:36,515 --> 00:40:38,835
[Speaker 0]
you folks tonight. This one comes from science
929
00:40:38,835 --> 00:40:41,555
[Speaker 0]
dot nasa dot gov. NASA put together a
930
00:40:41,555 --> 00:40:43,635
[Speaker 0]
movie night list of ten picks from its
931
00:40:43,635 --> 00:40:46,589
[Speaker 0]
streaming service, NASA plus. NASA says it's free
932
00:40:46,589 --> 00:40:49,150
[Speaker 0]
with no subscription and no ads. The two
933
00:40:49,150 --> 00:40:51,869
[Speaker 0]
longest are full features. Cosmic Dawn is the
934
00:40:51,869 --> 00:40:54,029
[Speaker 0]
story of building and launching the James Webb
935
00:40:54,029 --> 00:40:56,349
[Speaker 0]
Space Telescope, and it runs about ninety seven
936
00:40:56,349 --> 00:40:58,990
[Speaker 0]
minutes. Planetary Defenders follows the people who track
937
00:40:58,990 --> 00:41:01,005
[Speaker 0]
asteroids that can hit Earth. The rest are
938
00:41:01,005 --> 00:41:04,145
[Speaker 0]
shorter episodes including visits to Jupiter's moon, Europa,
939
00:41:04,204 --> 00:41:06,525
[Speaker 0]
and Saturn's moon, Titan. There's also one that
940
00:41:06,525 --> 00:41:09,404
[Speaker 0]
turns telescope data into sound and thirty minutes
941
00:41:09,404 --> 00:41:11,984
[Speaker 0]
of calm music over footage of the moon.
942
00:41:12,700 --> 00:41:15,900
[Speaker 0]
One set in Hawaii is called, our alien
943
00:41:15,900 --> 00:41:19,280
[Speaker 0]
Earth, the, the lava tubes of Mauna Loa.
944
00:41:19,820 --> 00:41:22,060
[Speaker 0]
On the big island, four signs seems to
945
00:41:22,060 --> 00:41:24,875
[Speaker 0]
work inside caves that that flowing lava left
946
00:41:24,875 --> 00:41:27,454
[Speaker 0]
behind, trying out ways to spot life underground
947
00:41:27,675 --> 00:41:30,875
[Speaker 0]
as practice for Mars. Yeah. So I I
948
00:41:30,875 --> 00:41:33,915
[Speaker 0]
definitely recommend checking these out. Ten awesome, awesome
949
00:41:33,915 --> 00:41:38,120
[Speaker 0]
movies. They're they they... They're gorgeous. So definitely
950
00:41:38,120 --> 00:41:40,920
[Speaker 0]
check this out if you're interested in some
951
00:41:40,920 --> 00:41:43,800
[Speaker 0]
space stuff. But that's all I got for
952
00:41:43,800 --> 00:41:45,400
[Speaker 0]
you folks tonight. A little bit of housekeeping
953
00:41:45,400 --> 00:41:47,640
[Speaker 0]
before I let you go. Become a GNC
954
00:41:47,640 --> 00:41:49,560
[Speaker 0]
insider at geek news central dot com forward
955
00:41:49,560 --> 00:41:52,365
[Speaker 0]
slash insider. Email me at geek news at
956
00:41:52,365 --> 00:41:54,285
[Speaker 0]
g mail dot com or Chris if you
957
00:41:54,285 --> 00:41:57,485
[Speaker 0]
wanna email Chris. But, subscribe via your paper...
958
00:41:57,485 --> 00:41:59,165
[Speaker 0]
Favorite podcast app and sign up for the
959
00:41:59,165 --> 00:42:01,405
[Speaker 0]
newsletter at geek news central dot com. Grab
960
00:42:01,405 --> 00:42:03,965
[Speaker 0]
a matter... Modern podcast app at podcast apps
961
00:42:03,965 --> 00:42:06,140
[Speaker 0]
dot com. Get all the new podcast two
962
00:42:06,140 --> 00:42:08,780
[Speaker 0]
point o features. And and a big, big,
963
00:42:08,780 --> 00:42:10,540
[Speaker 0]
big thanks to GoDaddy for over twenty years
964
00:42:10,540 --> 00:42:12,140
[Speaker 0]
of keeping this show on the air. Support
965
00:42:12,140 --> 00:42:14,060
[Speaker 0]
us at geek new central dot com forward
966
00:42:14,060 --> 00:42:18,275
[Speaker 0]
slash GoDaddy. Every click counts. But thanks for
967
00:42:18,275 --> 00:42:20,695
[Speaker 0]
tuning in, guys. Have a great great evening
968
00:42:20,915 --> 00:42:22,435
[Speaker 0]
or day or whatever time it is for
969
00:42:22,435 --> 00:42:25,155
[Speaker 0]
you folks. Next time we'll be, we'll be
970
00:42:25,155 --> 00:42:29,315
[Speaker 0]
in the studio, back to video and testing
971
00:42:29,315 --> 00:42:31,635
[Speaker 0]
some of our new Blueberry features, hopefully, that,
972
00:42:32,480 --> 00:42:34,240
[Speaker 0]
I'm pretty excited about now that I'm gonna
973
00:42:34,240 --> 00:42:38,000
[Speaker 0]
be doing some more, video work. So, look
974
00:42:38,000 --> 00:42:40,319
[Speaker 0]
forward to that, and, we'll see you guys