#796: Own Your Intelligence Stack with Tommy Eastman

1
00:00:00,000 --> 00:00:04,000
I've been very much looking forward to that.
2
00:00:04,000 --> 00:00:13,000
I was very excited when we first talked a couple weeks ago, because I've been using your product for four months now, and it has changed my life.
3
00:00:13,040 --> 00:00:13,580
You saw me.
4
00:00:13,800 --> 00:00:15,420
When you walked in here, you thought I was on the phone.
5
00:00:15,480 --> 00:00:16,580
I was talking to my Hermes agent.
6
00:00:16,640 --> 00:00:17,640
No, talking to your Hermes agent.
7
00:00:18,120 --> 00:00:19,220
Better than talking to somebody.
8
00:00:19,860 --> 00:00:21,120
I like the human.
9
00:00:21,260 --> 00:00:22,640
I like the human element, okay?
10
00:00:23,060 --> 00:00:26,380
I do like talking to my wife, Greg, who's off camera here.
11
00:00:26,440 --> 00:00:28,080
He's okay to talk to every once in a while.
12
00:00:28,080 --> 00:00:36,900
but uh it's uh it's changed how i work pretty pretty quickly i mean i'm 35 i've been doing
13
00:00:36,900 --> 00:00:42,120
tftc for almost a decade now and within six months my full workflow has changed
14
00:00:42,120 --> 00:00:49,360
and hermes is a critical part of that yeah what let's take a step back
15
00:00:49,360 --> 00:00:54,100
and talk about you personally how you ended up at news research which
16
00:00:54,100 --> 00:01:01,520
built the Hermes agent um you were in the bitcoin industry the crypto industry at foundry
17
00:01:01,520 --> 00:01:05,120
before this and that's i mean that's a whole nother conversation i think it's really interesting
18
00:01:05,120 --> 00:01:09,480
how a lot of these up-and-coming ai companies are seated with people that have been
19
00:01:09,480 --> 00:01:17,740
in the bitcoin and crypto industry before yep but how did you end up at noose what drew you to it
20
00:01:17,740 --> 00:01:20,980
and what have you been doing there since you got there?
21
00:01:21,540 --> 00:01:23,480
Yeah, so I was at Foundry,
22
00:01:23,860 --> 00:01:25,820
touched the Bitcoin stuff a fair bit,
23
00:01:25,920 --> 00:01:29,960
but was really focused on distributed and decentralized AI there.
24
00:01:29,960 --> 00:01:35,240
I just thought that AI, obviously an incredibly powerful tool,
25
00:01:35,760 --> 00:01:44,700
very important that we get the ideological part of AI shaped correctly.
26
00:01:44,700 --> 00:01:54,260
That was my first real exposure to that, thinking about the importance of AI for people, for the individual, the sovereignty that AI enables.
27
00:01:56,420 --> 00:02:08,700
I didn't really like a lot of the kind of first manifestations of that, like the first projects that kind of tried to champion that I wasn't that impressed with.
28
00:02:08,700 --> 00:02:11,100
I found Noose eventually.
29
00:02:11,300 --> 00:02:19,220
They were focused on post-training open models to make them more pliable, more human-like.
30
00:02:19,300 --> 00:02:22,340
This was when the meta models were state-of-the-art, right?
31
00:02:22,400 --> 00:02:23,220
The llama models.
32
00:02:26,560 --> 00:02:30,520
I quickly found that the team was just incredible.
33
00:02:30,720 --> 00:02:32,820
One, very ideologically driven, right?
34
00:02:32,820 --> 00:02:48,060
At Noose, we truly believe that the ability to own your intelligence stack is the most important technological innovation that we've ever faced as humans.
35
00:02:48,380 --> 00:02:51,740
But also that the team was just incredibly sharp.
36
00:02:52,320 --> 00:02:54,620
Everybody at Noose is just a killer.
37
00:02:55,540 --> 00:02:59,280
In love with AI, very driven.
38
00:02:59,420 --> 00:03:01,900
It's just an incredible, incredible experience.
39
00:03:02,820 --> 00:03:05,1000
And yeah, for me, I focus on a handful of things.
40
00:03:05,1000 --> 00:03:07,660
We all kind of wear a bunch of different hats, right?
41
00:03:07,660 --> 00:03:09,880
We're still pretty small, pretty grassroots.
42
00:03:10,1000 --> 00:03:13,800
But so I touch a bunch of stuff on the product side,
43
00:03:13,800 --> 00:03:16,880
onboarding enterprises and manage all of our
44
00:03:18,180 --> 00:03:19,800
our compute relationships as well.
45
00:03:19,800 --> 00:03:23,840
So we have our managed product, which is news portal, which powers
46
00:03:23,840 --> 00:03:27,020
which can power Hermes agent and powers Hermes agent for a lot of users.
47
00:03:27,800 --> 00:03:30,520
So managing all the inference and tools behind that.
48
00:03:30,520 --> 00:03:35,880
So let's take a step back and just explain Hermes.
49
00:03:35,960 --> 00:03:37,740
Let's focus on Hermes specifically.
50
00:03:38,240 --> 00:03:55,420
And let's, let's, because we've had this conversation off air a couple times now, but speaking to the person who thinks that the creme de la creme of their interaction with AI is opening chat GPT and hitting chat and thinking that they're getting inexperienced.
51
00:03:55,420 --> 00:04:00,620
what does hermes bring to the table and how should people view the hermes hermes harm is and how to
52
00:04:00,620 --> 00:04:06,300
utilize it yeah well that's the that's the incredible piece right is that we're so early
53
00:04:06,300 --> 00:04:11,100
we feel like ai is everywhere and everybody's talking about ai but there's only a small slice
54
00:04:11,100 --> 00:04:15,660
of people actually using ai and then the vast majority of that slice of people are just using
55
00:04:15,660 --> 00:04:22,300
ai basically as advanced google right google on steroids um like my mom loves perplexity she uses
56
00:04:22,300 --> 00:04:26,540
perplexity all the time but she's basically just she likes it because it's google but a little nicer
57
00:04:26,540 --> 00:04:36,300
um so where you start where you make that jump from ai chatbot to agentic ai is when you start
58
00:04:36,300 --> 00:04:43,180
to enable actual autonomy so these agents start to actually be able to make some decisions for you
59
00:04:43,180 --> 00:04:47,980
um and carry out tasks without you having to sit there and monitor them 24 7.
60
00:04:47,980 --> 00:04:51,980
You can go and have an agent book reservations for you.
61
00:04:51,980 --> 00:04:54,740
You can have an agent look at data that you have
62
00:04:54,740 --> 00:04:56,780
and build dashboards.
63
00:04:56,780 --> 00:04:58,780
You can start to really automate real tasks.
64
00:04:58,780 --> 00:05:01,280
At Noose, we use it for a bunch of different things,
65
00:05:01,280 --> 00:05:07,280
ranging from managing our CRM of enterprise customers
66
00:05:07,280 --> 00:05:20,160
to taking ideas that we have for content generation and working through a whole creative pipeline to create our end product Why Hermes Agent itself is I think very special
67
00:05:21,600 --> 00:05:23,360
Two major reasons.
68
00:05:23,540 --> 00:05:26,600
One being that, I'll say three major reasons.
69
00:05:26,940 --> 00:05:29,620
One, the first being that it's very simple to use.
70
00:05:29,800 --> 00:05:31,020
It really just works.
71
00:05:32,320 --> 00:05:35,320
I think a lot of open source projects end up with this issue
72
00:05:35,320 --> 00:05:37,820
where the repo gets bloated.
73
00:05:37,900 --> 00:05:39,600
They want to be inclusive of everything.
74
00:05:40,280 --> 00:05:46,540
But we've taken a stance where, yes, we want Hermes agent to be able to bend to your will
75
00:05:46,540 --> 00:05:51,100
and be completely pliable, but that it's really important that we maintain the sanctity of
76
00:05:51,100 --> 00:05:55,820
that repo and are somewhat opinionated in what kind of exists in that core repo to protect
77
00:05:55,820 --> 00:05:56,960
the user experience.
78
00:05:58,940 --> 00:06:02,580
So working out of the box is a huge reason why it's been successful.
79
00:06:02,580 --> 00:06:06,100
Two is the memory piece, the self-evolution piece.
80
00:06:06,380 --> 00:06:14,600
So as you use Hermes Agent, you'll notice market improvements in the quality of output over time.
81
00:06:15,440 --> 00:06:31,500
And this is a really important piece because I think where people end up frustrated and what prevents people from using ChatGPT in autonomous ways is you can't do tasks reproducibly with the same results.
82
00:06:31,500 --> 00:06:37,340
right so you you ask chat to go do a task it tries to do the task it might figure it out the first day
83
00:06:37,340 --> 00:06:40,540
and then the next day you have it do the same task and it can't figure it out right and it's a super
84
00:06:40,540 --> 00:06:46,380
frustrating experience hermes agent stores the primitives of that path where it found the right
85
00:06:46,380 --> 00:06:51,100
solution and then recalls them every single time so if you figure something out once with hermes
86
00:06:51,100 --> 00:06:54,540
agent you're going to be able to do it again and again and again and again and that's where you
87
00:06:54,540 --> 00:06:58,780
can truly say okay agent go do this and you don't have to sit there and monitor and worry and you
88
00:06:58,780 --> 00:07:01,560
And you can have other tasks be dependent on that task.
89
00:07:03,180 --> 00:07:14,500
The third piece of why Hermes agent, I think, is so beautiful, and this is like the ideological holy grail, right, is it's able to be fully model agnostic.
90
00:07:14,1000 --> 00:07:17,620
And that is important for a lot of reasons.
91
00:07:17,840 --> 00:07:20,480
And we've thought it's been important for a really long time at Noose.
92
00:07:21,480 --> 00:07:27,660
You've seen more and more attention come to this issue now, though, of data sovereignty.
93
00:07:27,660 --> 00:07:29,900
and provider lock-in.
94
00:07:30,120 --> 00:07:32,900
Nobody wants to be locked in to a single big lab.
95
00:07:33,420 --> 00:07:36,320
Nobody wants these labs training on their data.
96
00:07:37,380 --> 00:07:40,460
People want to own that stuff because that's the secret sauce, right?
97
00:07:40,460 --> 00:07:43,400
That's the sauce of my company is this data.
98
00:07:44,520 --> 00:07:48,260
And if I'm sending all of that to a lab that's going to use that to train in the next model,
99
00:07:48,360 --> 00:07:51,580
I'm giving away the secret codes.
100
00:07:53,260 --> 00:07:56,640
So Hermes Agent enables you to own your full stack.
101
00:07:56,640 --> 00:07:58,920
You can choose to have stuff managed if you want.
102
00:07:59,020 --> 00:08:00,960
Your cloud instance is managed or whatever.
103
00:08:01,720 --> 00:08:10,600
You can basically craft whatever solution meets your sovereignty and privacy needs, and you can go implement that.
104
00:08:10,660 --> 00:08:12,940
And if things change, you can pivot it as well, right?
105
00:08:12,960 --> 00:08:15,680
So there's no provider lock, and you can use any model in the world.
106
00:08:16,120 --> 00:08:17,620
You can use much cheaper models.
107
00:08:17,820 --> 00:08:24,200
Like there's so many open source models that are incredibly performant for the vast majority of tasks now and at a fraction of the price point.
108
00:08:24,200 --> 00:08:34,140
um so yeah it's i think it enables it enables you to just have a much more sovereign ai experience
109
00:08:34,140 --> 00:08:42,920
and really own the full the full stack from the floor up and it's uh i mean we the so my my hermy
110
00:08:42,920 --> 00:08:52,120
story my name is marty ben i'm a hermy so i i open claw over christmas break people like open claw
111
00:08:52,120 --> 00:08:56,600
i'm like one it snowed like 30 inches in philly one sunday and the kids were watching a movie i
112
00:08:56,600 --> 00:09:02,440
was like all right i'll try this out so like first i've ever set up a uh vm and set up a cloud server
113
00:09:03,960 --> 00:09:09,400
set up the hetzner server like went through this youtube store set up open club and connected it
114
00:09:09,400 --> 00:09:15,240
to anthropic at the time when you could oauth in before they uh before they prevented that from
115
00:09:15,240 --> 00:09:22,040
happening and was immediately blown away like holy crap this thing is insane so for about four months
116
00:09:22,040 --> 00:09:39,420
I was building out that OpenClaw instance and integrating more of our workflows into the agent setup and the harness in the cloud and seeding the file system in the cloud with everything that I wanted this computer to be able to access.
117
00:09:39,420 --> 00:09:46,800
and along the way like April May I was talking about open claw and I heard people
118
00:09:46,800 --> 00:09:51,780
it was probably actually March April people began to be like hey you're using open claw you should
119
00:09:51,780 --> 00:09:55,740
like try out Hermes and then like another similar situation or I was just a Sunday afternoon
120
00:09:55,740 --> 00:10:03,360
and uh I was sitting there I was like okay like I've had 10 people like reply to some of my tweets
121
00:10:03,360 --> 00:10:08,640
like you should try her and I was like okay we love our community literally went to my open claw
122
00:10:08,640 --> 00:10:15,120
agent and telegram i was like hey spin up a hermes agent uh in your server and uh let me know when
123
00:10:15,120 --> 00:10:19,680
that's finished and then i hooked up telegram and then within like 24 hours i was like okay
124
00:10:20,320 --> 00:10:30,019
hermes agent shut down the open claw agent and uh we gonna we gonna begin working and and it been insane what we been able to do what I been able to
125
00:10:30,019 --> 00:10:35,239
do on the team at TFTC over the last four or five months just really seeding
126
00:10:35,239 --> 00:10:41,059
the Hermes agent but to your point I think that's what I really want to get
127
00:10:41,059 --> 00:10:44,779
through to people when we discussed this on our call a couple weeks ago is I'd
128
00:10:44,779 --> 00:10:47,839
be interested to get your thoughts on how you view this like I think there's
129
00:10:47,839 --> 00:10:51,359
three legs to the stool of an agentic system,
130
00:10:51,359 --> 00:10:52,679
particularly if you're doing it for a business,
131
00:10:52,679 --> 00:10:56,699
you have a file system or the second brain,
132
00:10:56,699 --> 00:11:01,859
and there's ways in which you can make that second brain
133
00:11:01,859 --> 00:11:04,479
stronger with things like semantic search for QMD
134
00:11:04,479 --> 00:11:08,959
or an obsidian vault for a knowledge graph with associations.
135
00:11:08,959 --> 00:11:11,859
But I would bucket those into one thing,
136
00:11:11,859 --> 00:11:14,399
second brain, essentially a file system,
137
00:11:14,399 --> 00:11:21,219
the harness the agentic system in our case Hermes and then the models the two
138
00:11:21,219 --> 00:11:25,199
most important things are the second brain and the harness and the models
139
00:11:25,199 --> 00:11:29,259
like you said you can be model agnostic and that's what I love about Hermes is
140
00:11:29,259 --> 00:11:33,199
it makes it extremely easy to switch models and I'm hitting slash model
141
00:11:33,199 --> 00:11:38,519
changing for different tasks when I don't want to burn tokens or burn
142
00:11:38,519 --> 00:11:44,019
expensive tokens or if I want to use a more private open source model I don't
143
00:11:44,019 --> 00:11:49,919
think most people understand like i think we're at this it's not that i don't think i know most
144
00:11:49,919 --> 00:11:55,179
people don't understand that this uh this sort of three-pronged approach and what is the value
145
00:11:55,179 --> 00:11:58,619
what is the relative value of the different prongs within that three-pronged approach
146
00:11:58,619 --> 00:12:05,139
and i think that's where the opportunity lies and why i love hermes i think you mentioned like the
147
00:12:05,139 --> 00:12:11,199
self-referential learning as part of the memory system and i think there's a ton of people
148
00:12:11,199 --> 00:12:14,059
out there saying, all right, AI is here.
149
00:12:14,879 --> 00:12:16,339
I've been convinced it's not going anywhere.
150
00:12:16,459 --> 00:12:19,559
I need to implement it in my business.
151
00:12:19,639 --> 00:12:22,259
So just focusing on businesses specifically,
152
00:12:22,659 --> 00:12:24,459
like how would you, like, is that,
153
00:12:24,459 --> 00:12:28,119
is my inclination of that three-prong approach
154
00:12:28,119 --> 00:12:29,359
right in your mind?
155
00:12:29,719 --> 00:12:32,019
And if so, how do you articulate that,
156
00:12:32,059 --> 00:12:33,619
particularly to like enterprise clients
157
00:12:33,619 --> 00:12:34,519
who walk them through this?
158
00:12:34,719 --> 00:12:35,339
Yeah, yeah.
159
00:12:35,619 --> 00:12:36,979
I think you're completely right.
160
00:12:38,679 --> 00:12:40,819
There's so much value.
161
00:12:40,819 --> 00:12:43,839
The returns are compounding.
162
00:12:43,979 --> 00:12:49,859
The quality of intelligence is compounding the more and more, as the agent gets more and more intelligence.
163
00:12:50,159 --> 00:12:53,379
So that file system is incredibly valuable.
164
00:12:54,319 --> 00:12:57,679
The agents are only as good as the data that they are able to access.
165
00:12:58,199 --> 00:13:00,119
It's not some magic.
166
00:13:00,119 --> 00:13:13,139
And so the ability to give your agent access to as much context as possible and basically just let it go to work is the secret sauce.
167
00:13:13,299 --> 00:13:16,379
That's the real value in these systems.
168
00:13:17,699 --> 00:13:24,519
And I think it's almost hard to articulate sometimes to enterprises, like, why would I use Hermes Agent?
169
00:13:25,079 --> 00:13:27,359
It's like, well, you can use it for everything.
170
00:13:27,359 --> 00:13:32,699
and that can that's it's like it's almost too abstract right we're too good to be true yeah
171
00:13:32,699 --> 00:13:38,639
yeah yeah it's like okay what does that mean and it's like the reality is though you really can and
172
00:13:38,639 --> 00:13:43,639
you've seen it i'm sure you really can use it for anything any monotonous task you may have
173
00:13:43,639 --> 00:13:48,559
um anything you want done just throw it at your hermes agent and see what happens
174
00:13:48,559 --> 00:13:54,219
right i think the the low-hanging fruit where a lot of people immediately say wow this is incredible
175
00:13:54,219 --> 00:13:56,539
is working with big data.
176
00:13:58,119 --> 00:14:00,779
If you have, whether it's your Stripe data,
177
00:14:00,919 --> 00:14:01,839
your customer data,
178
00:14:03,059 --> 00:14:05,679
the ability to create, to organize that,
179
00:14:05,759 --> 00:14:07,939
to create visualizations, to do modeling
180
00:14:07,939 --> 00:14:14,299
is stuff that these agentic systems excel at, right?
181
00:14:14,299 --> 00:14:17,519
They are just so much better than we can possibly be at.
182
00:14:18,859 --> 00:14:21,759
And so I think that's where you see a lot of the time
183
00:14:21,759 --> 00:14:23,539
people immediately go, okay,
184
00:14:23,539 --> 00:14:25,039
this is a game changer.
185
00:14:25,119 --> 00:14:28,319
When they have some big data problem and they make visualizations out of it,
186
00:14:28,339 --> 00:14:29,719
it works right away in one shot.
187
00:14:31,339 --> 00:14:37,079
But yes, to your original point, the structure is your company.
188
00:14:37,819 --> 00:14:41,359
In its entirety, it should serve as your company brain.
189
00:14:41,679 --> 00:14:45,579
Those three pieces of the stool, like you said, are the company brain.
190
00:14:46,779 --> 00:14:51,899
I think I agree with you that it's context or file system,
191
00:14:51,899 --> 00:14:56,279
harness and then model i don't want to discount the model too much because model quality does
192
00:14:56,279 --> 00:15:02,539
matter a lot right and but they're very additive um but they're becoming commoditized that's where
193
00:15:02,539 --> 00:15:07,279
like i'm oh in the long tail for sure i think in the long tail you can be like over time the ability
194
00:15:07,279 --> 00:15:12,139
to be model agnostic is only going to increase yes yeah i mean it's a reaching parity nine nine
195
00:15:12,139 --> 00:15:17,599
months ago you couldn't use an open model for anything like they were basically worthless right
196
00:15:17,599 --> 00:15:27,199
And now you, I like to say, I think about 90% of users can do 90% of their work with open models.
197
00:15:27,799 --> 00:15:48,159
There is definitely stuff that you want the biggest guns for but tons of work can go to open models And yes that gap is continuing to close and it only going to become more and more commoditized Yeah And at what point do you like what is the minimum AI IQ level that you need to do all the tasks Like do you really like
198
00:15:48,499 --> 00:15:54,159
say a year from now, are you going to need a frontier model to do your modeling or your CRM?
199
00:15:54,399 --> 00:15:59,299
Yeah. The vast majority of people are never going to do tasks that require these top tier models.
200
00:15:59,299 --> 00:16:06,439
it's just not part of part of what they do no and the so let's take a step back and talk about
201
00:16:06,439 --> 00:16:14,419
AI broadly like where do you think we are um what trends are you most interested right now
202
00:16:14,419 --> 00:16:20,179
is it the data center compute side is it this model race is it the harness uh
203
00:16:20,179 --> 00:16:27,079
what are you most interested right now in the world of AI outside of Hermes and noose yeah I
204
00:16:27,079 --> 00:16:34,939
I think it's really interesting to follow how the open source models have changed the market.
205
00:16:35,719 --> 00:16:38,979
Like I said, nine months ago, you really couldn't use open source AI.
206
00:16:40,839 --> 00:16:48,239
Now, if we transport back to January, February, you saw a bunch of prominent CEOs talking about,
207
00:16:48,239 --> 00:16:52,319
our employees need a token max, token max, token max, unlimited spend.
208
00:16:52,539 --> 00:16:56,719
Very quickly, they were like, whoa, way too much money.
209
00:16:57,079 --> 00:16:59,759
and want to find cheaper solutions.
210
00:17:00,079 --> 00:17:03,099
And that is timed well with open models
211
00:17:03,099 --> 00:17:05,539
reaching the point where they can be these cheaper solutions.
212
00:17:05,899 --> 00:17:08,539
I think there's a lot of stigma around them for some people still
213
00:17:08,539 --> 00:17:10,659
about being mostly Chinese models,
214
00:17:10,819 --> 00:17:13,899
like China is completely championing the open source model,
215
00:17:13,899 --> 00:17:15,099
the model race.
216
00:17:16,659 --> 00:17:21,019
But you're seeing numbers in Hermes Agent
217
00:17:21,019 --> 00:17:22,739
on open models go up and up,
218
00:17:22,859 --> 00:17:25,999
on open router, open source models going up and up.
219
00:17:25,999 --> 00:17:37,079
Companies are recognizing that AI spend is out of control, that being locked into OpenAI and Anthropic is prohibitively expensive to power your business.
220
00:17:38,759 --> 00:17:41,599
And open models are really the only solution.
221
00:17:42,319 --> 00:17:48,439
And there's such an incredible free market of these open models.
222
00:17:48,699 --> 00:17:54,359
There are so many different GPU providers in different locations, different compliance requirements.
223
00:17:54,359 --> 00:18:04,359
So you can tailor whatever you need. If you need SOC 2, if you need GDPR, if you need them in the U.S., there's all these different things that you can tailor to fit your needs.
224
00:18:04,819 --> 00:18:09,039
But it's a much just more freer market. Right. So you're seeing that price discovery.
225
00:18:09,159 --> 00:18:19,519
You're seeing model prices decrease and you're seeing people be able to access intelligence much, much cheaper, which is obviously just incredible for for the economy, for business, for small businesses, especially.
226
00:18:20,159 --> 00:18:26,239
no and i think i just discussed this with jordy visser who was on before you rolled in but i think
227
00:18:26,239 --> 00:18:31,439
the the latham and walk-in story from a couple weeks ago the law firm basically said we're not
228
00:18:31,439 --> 00:18:35,599
using anthropic or open ai we're gonna buy our own gpus and self-host her model that i mean
229
00:18:36,319 --> 00:18:41,439
that's a massive signal not only the models are good enough to do that but also i think a trend
230
00:18:41,439 --> 00:18:48,719
that many people are not um not appreciating enough and i think alex carp's been beating the
231
00:18:48,719 --> 00:18:53,759
the drum about this but like particularly if you're a enterprise the size of latham and walk-ins or
232
00:18:53,759 --> 00:19:02,639
any fortune 500 company like the feeding of your company ip to anthropic and serenai is
233
00:19:02,639 --> 00:19:08,439
i think it's going to get to a point where shareholders or board boards of directors
234
00:19:08,439 --> 00:19:11,859
going to say this is like a breach of fiduciary responsibility i think you're just handing over
235
00:19:11,859 --> 00:19:19,459
secret sauce to yeah these model providers who are uh proving or who are literally launching
236
00:19:19,459 --> 00:19:23,859
products that they're training off of companies that are doing this like they're going to try to
237
00:19:23,859 --> 00:19:28,819
compete with you and mobilize your business while it's providing you value in the interim and so
238
00:19:29,379 --> 00:19:36,579
to your point about open models like i think there will be um a massive inflow of usage for
239
00:19:36,579 --> 00:19:42,179
an increased usage of these open models from like a fiduciary perspective just to preserve your ip
240
00:19:42,739 --> 00:19:46,979
yeah i mean people are just throwing their trade secret it's actually a really interesting
241
00:19:46,979 --> 00:19:52,419
psychological phenomenon right like yeah businesses are doing it but also that the average user the
242
00:19:52,419 --> 00:20:00,499
chat gpt user the anthropic user people are people share so so much with these models and almost
243
00:20:01,059 --> 00:20:06,499
like completely unaware of of privacy or how that obviously could could could backfire and the
244
00:20:06,499 --> 00:20:10,659
different ends that that could go to it's happening at the at the individual level it is happening at
245
00:20:10,659 --> 00:20:15,859
the business level as well but i think at least now the more and more businesses are very aware
246
00:20:15,859 --> 00:20:21,379
of that that's something that really resonates um with businesses that we talk to about about hermy's
247
00:20:21,379 --> 00:20:28,819
agent is they want to ever since that alex alex carp um spiel he went on businesses want to own
248
00:20:28,819 --> 00:20:30,819
They're very, very aware.
249
00:20:30,819 --> 00:20:33,819
They're very, very concerned about where their data is going.
250
00:20:33,819 --> 00:20:38,819
I think the general public is lagging behind, but I think there'll be a rude awakening for them as well.
251
00:20:38,819 --> 00:20:39,819
Agreed.
252
00:20:39,819 --> 00:20:40,819
Agreed.
253
00:20:40,819 --> 00:20:44,819
It's funny, I was taking the MetaMuse agent for SPIN this morning.
254
00:20:44,819 --> 00:20:45,819
I was like, wow, this thing's powerful.
255
00:20:45,819 --> 00:20:49,959
But then, how much of this information do I want to give to Meta?
256
00:20:50,011 --> 00:20:54,091
suck and i think that's a question that a lot of people are not to grab this but i mean
257
00:20:55,851 --> 00:21:00,091
history shows particularly in the digital age most people don't care um that's why i think like
258
00:21:00,091 --> 00:21:03,611
focus on enterprise like where you have to care because money and shareholder value for
259
00:21:03,611 --> 00:21:08,891
lack of better term is on on the line um but then for the individual like running individual agents
260
00:21:08,891 --> 00:21:16,731
like i think uh hopefully uh team at news research and others can get the the ux
261
00:21:16,731 --> 00:21:23,091
uh to parody with something like a muse agent or just works out of the box which i think to your
262
00:21:23,091 --> 00:21:26,831
point like that was one of the most impressive things so like i literally told my open call agent
263
00:21:26,831 --> 00:21:32,491
set up a hermese agent and had it working within like 10 minutes yep yeah i think that um
264
00:21:32,491 --> 00:21:38,131
it's really important to make the product easier and easier for people to use right because
265
00:21:38,131 --> 00:21:44,511
if it's you know hours of setup versus instant setup and the privacy is the only differentiator
266
00:21:44,511 --> 00:21:48,431
that's been proven time and time again that that's not enough.
267
00:21:48,631 --> 00:21:51,071
I do think on the individual privacy aspect,
268
00:21:51,431 --> 00:21:57,511
I think it probably needs some sort of massive event
269
00:21:57,511 --> 00:22:04,331
to catalyze people growing aware of this concern.
270
00:22:04,631 --> 00:22:06,391
But I do think we're at a stage where,
271
00:22:06,871 --> 00:22:10,191
yes, it was bad when you shared all of your information
272
00:22:10,191 --> 00:22:12,311
in this digital age previously.
273
00:22:12,311 --> 00:22:20,611
But now the ability for agents to access it, to compile it, access it, spread it, it's seconds, right?
274
00:22:20,691 --> 00:22:32,611
So I think there are way worse potential outcomes on the table now that agentic AI is as powerful as it is if people continue to be so free with their data.
275
00:22:32,611 --> 00:22:42,471
I mean, imagine what percentage of chat GPT users would be horrified if their chats all got leaked.
276
00:22:42,491 --> 00:22:51,211
Well, we see, wasn't there the ability to like Google's, like if you were sharing the output and you create like a share link that was getting cached in Google search.
277
00:22:51,551 --> 00:22:53,071
Yeah, yes, yes, yes.
278
00:22:53,071 --> 00:22:54,971
Something similar happened to Meta as well.
279
00:22:55,591 --> 00:22:56,071
Yeah.
280
00:22:56,431 --> 00:22:59,731
Where they thought they were like having a private conversation with the AI or they're posting it publicly.
281
00:23:00,991 --> 00:23:02,011
Yeah, it's horrifying.
282
00:23:02,011 --> 00:23:09,771
horrifying horrifying yeah and it's so how would you like that's what i'm i've been trying to
283
00:23:09,771 --> 00:23:17,571
discern because things are changing so quickly like last year at this time like opus 4-6 wasn't
284
00:23:17,571 --> 00:23:23,491
even out yet correct so you didn't have like there was that frontier model or breaks remember
285
00:23:23,491 --> 00:23:27,391
it was like oh wow now this stuff can like really cook and then a month later it was like
286
00:23:27,391 --> 00:23:31,351
they were working on open claw you guys been working on hermes but like people didn't become
287
00:23:31,351 --> 00:23:35,911
aware of like this harness thing until that happens we're only like nine months into this what
288
00:23:36,951 --> 00:23:41,831
where does it like how do you think this looks uh even three months from now a year from now
289
00:23:41,831 --> 00:23:45,831
and what trends should people be paying attention to particularly around
290
00:23:46,471 --> 00:23:52,951
user experience and and workflows that they can incorporate using this stuff yeah i well
291
00:23:54,151 --> 00:23:59,271
what you said is exactly right right like at this time last year opus 4 6 wasn't out i think the
292
00:23:59,271 --> 00:24:05,431
best way to describe that change, that clear step function increase in AI capability was
293
00:24:06,151 --> 00:24:13,431
at this time last year, the top 10 developers I knew, 0% of their code was written with AI.
294
00:24:14,871 --> 00:24:20,151
It wasn't better than them, too many issues, too many hallucinations, no ability to do long-running
295
00:24:20,151 --> 00:24:28,871
tasks. Now, 100% of their code is written with AI. There's almost no reason to be writing code
296
00:24:28,871 --> 00:24:30,011
except with AI now.
297
00:24:30,371 --> 00:24:33,891
And people are doing incredibly long-running tasks with agents.
298
00:24:35,871 --> 00:24:37,991
And that was less than a year ago.
299
00:24:37,991 --> 00:24:40,771
Like you said, like nine months ago, that change happened.
300
00:24:41,251 --> 00:24:44,051
And so to think about where we are from the next nine months,
301
00:24:44,491 --> 00:24:46,791
it's really, really hard to be predictive.
302
00:24:46,791 --> 00:24:56,531
I think that what's very clear is we're early on the adoption of AI,
303
00:24:56,531 --> 00:25:02,811
and we're going to start to see like just those basic levels like engineers all using ai we're
304
00:25:02,811 --> 00:25:07,871
going to start to see that all be integrated into every enterprise this like this year from this
305
00:25:07,871 --> 00:25:13,731
time now this time next year i think every enterprise all code is written by ai right
306
00:25:13,731 --> 00:25:19,091
like there's really no there will be really no excuse at that point i think it will be abundantly
307
00:25:19,091 --> 00:25:24,511
clear that you have to do this to keep up to ship because the companies that choose to do it are
308
00:25:24,511 --> 00:25:28,671
going to, it won't happen overnight, but it will happen fairly quickly, will drastically
309
00:25:28,671 --> 00:25:32,611
outperform and be able to cut expenses and be able to ship faster than companies who
310
00:25:32,611 --> 00:25:36,691
do not choose to use AI.
311
00:25:36,691 --> 00:25:44,491
I think as far as novel behavior that emerges, I think we're getting close to starting to
312
00:25:44,491 --> 00:25:51,091
see more and more autonomy be dropped into the hands of AI.
313
00:25:51,091 --> 00:25:54,351
There's always the examples of, oh, my AI is shopping for me.
314
00:25:54,351 --> 00:26:07,928
stuff it really isn happening yet for the most part right like you can kind of manufacture some text but it by no means a smooth user experience And so that plumbing has to continue to improve
315
00:26:09,928 --> 00:26:13,188
for you to actually make that transformation
316
00:26:13,188 --> 00:26:14,888
where it is as simple as text your Hermes agent,
317
00:26:14,888 --> 00:26:17,548
like, oh, I need this purchased
318
00:26:17,548 --> 00:26:19,668
and the Hermes agent goes and purchases it.
319
00:26:19,668 --> 00:26:21,928
The next iteration of that, which is even cooler,
320
00:26:21,928 --> 00:26:23,408
and this is where we're heading,
321
00:26:23,408 --> 00:26:26,168
is Hermes agent is always on
322
00:26:26,168 --> 00:26:28,368
and actually being predictive, right?
323
00:26:28,528 --> 00:26:32,748
So if I, for example, and this kind of can happen now,
324
00:26:32,808 --> 00:26:34,548
but I think it's going to become more and more prevalent,
325
00:26:34,708 --> 00:26:38,048
is I have to go to SF next week.
326
00:26:38,168 --> 00:26:39,208
It's on my calendar.
327
00:26:39,648 --> 00:26:41,748
I forgot to book flights.
328
00:26:41,868 --> 00:26:43,348
I forgot to book hotel.
329
00:26:43,728 --> 00:26:46,928
Hermes agent just goes and does all of that for me,
330
00:26:46,948 --> 00:26:48,868
and it's handled, and I don't even have to worry about it.
331
00:26:48,908 --> 00:26:51,848
And then I get in the routine where, yep, oh, Marty texted me
332
00:26:51,848 --> 00:26:53,288
and said he wants to hang out in Philly.
333
00:26:54,288 --> 00:26:55,268
Hermes agent's going to go.
334
00:26:55,268 --> 00:26:58,968
it's going to book my flights. It's going to book my hotels. It's going to do it all correctly.
335
00:26:59,368 --> 00:27:02,548
It's going to send Marty the information that he needs to know. It's going to send Marty's
336
00:27:02,548 --> 00:27:08,828
Hermes agent the information it needs to know. Right. So I think, um, all of these minor tasks
337
00:27:08,828 --> 00:27:14,748
that you have to do in your life can be pushed off to agents, right? It can be foisted onto
338
00:27:14,748 --> 00:27:18,688
agents and clear you up. So I like, I hate booking up hotels. I'm terrible at it. I book wrong hotels
339
00:27:18,688 --> 00:27:22,668
all the time. I book wrong flights all the time. Hermes agent should do all that for me. And I think
340
00:27:22,668 --> 00:27:28,248
we're getting really close to it being able to and i think like that three six month window um
341
00:27:28,248 --> 00:27:32,368
of where it's in an always on state it's being predictive and it's being autonomous i think
342
00:27:32,368 --> 00:27:40,228
that's that's very very reasonable i'm a massive uh flight hotel train procrastinator yeah yeah
343
00:27:40,228 --> 00:27:44,688
same it's horrible i've got to go to i'm gonna go somewhere in a couple weeks so that i should
344
00:27:44,688 --> 00:27:48,388
have booked the flight months ago yeah i booked it yesterday i was like oh i'm paying the premium
345
00:27:48,388 --> 00:27:54,148
for this because it's been it's been on my calendar and it's uh what are you seeing
346
00:27:55,348 --> 00:28:00,468
internally so this is one thing selfishly i want to get out like obviously you interact with
347
00:28:01,028 --> 00:28:06,068
some of individuals and enterprises leveraging hermes agent what are some of the most creative or
348
00:28:06,868 --> 00:28:12,628
mind-blowing ways or workflows that people have incorporated using hermes yeah it's a good
349
00:28:12,628 --> 00:28:18,308
question um or another way to frame it is like who's utilizing it the best and why is the way
350
00:28:18,308 --> 00:28:23,828
they're utilizing it the best way yeah well i think news research utilizes it the best no but
351
00:28:23,828 --> 00:28:29,428
but in all seriousness we i think we it's really really important for us right if we're gonna
352
00:28:29,428 --> 00:28:36,308
deliver this product to enterprise that we are the number one power user of it um and so we use it we
353
00:28:36,308 --> 00:28:37,208
We use it for everything.
354
00:28:38,848 --> 00:28:43,008
I couldn't even try to quantify the productivity increase
355
00:28:43,008 --> 00:28:47,788
that Hermes Agent has applied to Noose.
356
00:28:47,908 --> 00:28:49,528
But it's funny when people ask us how many,
357
00:28:49,828 --> 00:28:50,648
you know, how big is the team?
358
00:28:50,708 --> 00:28:53,428
And we say, you know, 30 people, 35 people.
359
00:28:53,808 --> 00:28:55,408
They're often blown away.
360
00:28:55,408 --> 00:28:57,388
Just because if you've watched the rate
361
00:28:57,388 --> 00:28:59,448
that we've shipped on the Hermes Agent repo
362
00:28:59,448 --> 00:29:01,568
and the cloud products and the enterprise offering,
363
00:29:02,268 --> 00:29:03,788
it's something that would, you know,
364
00:29:03,788 --> 00:29:09,948
take an order of magnitude more people in the past um so we use hermes agent for
365
00:29:11,068 --> 00:29:18,428
you know managing our full our full sales pipeline um you know no need for these you know the standard
366
00:29:18,428 --> 00:29:26,908
kind of you know crm crm sass hermes agent can manage that for our content creation pipeline um
367
00:29:27,948 --> 00:29:32,188
we haven't i don't know if you've seen our i'm sure you have our our design work and the the
368
00:29:32,188 --> 00:29:33,108
The videos we put out.
369
00:29:33,408 --> 00:29:34,068
Very aesthetically pleasing.
370
00:29:34,148 --> 00:29:35,828
The aesthetic, we want things to look pretty.
371
00:29:37,308 --> 00:29:41,568
The stuff we put out, all of that is generated in Hermes agent pipeline.
372
00:29:41,868 --> 00:29:43,308
It's very reproducible.
373
00:29:43,508 --> 00:29:49,108
It allows our designers to have a ton of control over the end product.
374
00:29:49,108 --> 00:29:55,328
And it takes a ton of manual graphic design work and labor off of their plate.
375
00:29:55,328 --> 00:30:04,008
one of the most interesting tools we have is an agent internally that all our employees can use
376
00:30:04,008 --> 00:30:11,568
and it basically has a bunch of key data sources of ours so if a you know if a user complaint or if
377
00:30:11,568 --> 00:30:18,028
a customer complains about a billing issue or oh some something isn't working correctly a news
378
00:30:18,028 --> 00:30:24,028
portal or my hermes agent there's some bug we can just ask we can just ask our internal hermes agent
379
00:30:24,028 --> 00:30:33,608
Hermes Agent debugs it, can push fixes, and basically is able to access all of our data backends at once, whatever it may need.
380
00:30:33,648 --> 00:30:40,328
And it becomes over time really, really capable and competent at accessing this data and turning it into an actionable item for the user.
381
00:30:41,028 --> 00:30:48,488
So rather than somebody having to manually sift through infinite logs basically and find issues, Hermes Agent can find it in seconds.
382
00:30:48,828 --> 00:30:52,928
And that's actually a good way of describing what Hermes Agent is good at.
383
00:30:52,928 --> 00:30:56,768
It can do any task that a human could do,
384
00:30:56,768 --> 00:30:59,188
but would require infinite patience, right?
385
00:30:59,188 --> 00:31:02,028
Like nobody's actually able to sift through millions
386
00:31:02,028 --> 00:31:02,928
of pages of logs.
387
00:31:02,928 --> 00:31:05,388
Like you could, but you would go insane.
388
00:31:05,388 --> 00:31:07,308
But Hermes agent does it instantaneously
389
00:31:07,308 --> 00:31:09,088
and Hermes agent doesn't complain either.
390
00:31:09,088 --> 00:31:20,604
So it a pretty fun solution Yeah That what I wondering I would love again selfishly to give your team access to my Hermes agency Like where am I fucking up
391
00:31:21,144 --> 00:31:24,664
Because that's what I'm continuously trying to figure out is,
392
00:31:24,764 --> 00:31:26,144
all right, how can I better utilize this?
393
00:31:26,264 --> 00:31:28,644
I always feel like I'm not utilizing,
394
00:31:29,504 --> 00:31:30,744
just in AI generally,
395
00:31:30,884 --> 00:31:33,624
it's like I'm not utilizing everything to the best of its ability.
396
00:31:33,624 --> 00:31:38,684
And it's trying to explore where the boundaries and the edges are.
397
00:31:38,684 --> 00:31:45,184
and i feel like every week i'm an aha i'm like holy crap like i just like uh one example this
398
00:31:45,184 --> 00:31:50,304
will be the second time you're hearing the story uh in two two consecutive episodes if you listen
399
00:31:50,304 --> 00:31:55,404
to the geordie viscer episode but i'll tell it again i was at the bar we published this
400
00:31:55,404 --> 00:32:03,584
podcast via fountain um and so like our rss or video or audio file video file description
401
00:32:03,584 --> 00:32:11,544
everything goes up there trans trans uh transcript as well they uh they have a it's not publicly
402
00:32:11,544 --> 00:32:17,744
available yet and i hope they don't get mad at me for disclosing that they have a built an mcp that
403
00:32:17,744 --> 00:32:27,044
i have uh mpc mcp mcp mcp that i have a beta access to and so bound has video the transcripts
404
00:32:27,044 --> 00:32:33,104
and i've connected the mcp to my hermes agent and i can literally just voice the text like hey
405
00:32:33,104 --> 00:32:40,564
uh tommy and i the podcast is published now we were we were um let's get meta here let's see if
406
00:32:40,564 --> 00:32:46,464
my hermes agent will pick this up by reading the transcript uh i want you to use the fountain mcp
407
00:32:46,464 --> 00:32:54,364
to clip out the the section of the conversation with tommy where we're talking about uh how to
408
00:32:54,364 --> 00:33:00,684
connect fountains mcp what i what i use it for uh take a take like the the most engaging quote
409
00:33:00,684 --> 00:33:04,924
from that and then a one sentence brief description tag tommy in it and then post it on x and
410
00:33:05,484 --> 00:33:10,444
i did something similar to that last night within five minutes the hermese agent had used fountain to
411
00:33:11,244 --> 00:33:16,684
find exactly what i described in the podcast i published yesterday clip it out send it to me
412
00:33:16,684 --> 00:33:21,404
to review and then like send tweet it's incredible within five minutes it's incredible like i'm at
413
00:33:21,404 --> 00:33:26,364
the bar having a beer yep and getting work done yeah yeah for a media company it's like hey that
414
00:33:26,364 --> 00:33:31,164
his work we're getting uh yeah well no right it's i mean that's a non-negligible amount of labor
415
00:33:31,164 --> 00:33:36,204
that you would you would have had to do it's kind of a pain and a tedious task and it's like it's
416
00:33:36,204 --> 00:33:40,684
like death by a million paper cuts when you have to do all those things all the time and the amount
417
00:33:40,684 --> 00:33:47,484
of time that it frees up by you being able to just dump all of that on it is i mean how much i guess
418
00:33:47,484 --> 00:33:53,804
i'd ask you can you quantify the impact on business that you know hermy's agent has had
419
00:33:54,364 --> 00:34:01,964
for you all over the last four months uh yeah i mean i think from
420
00:34:03,564 --> 00:34:08,604
it's increased our profit margins increased our revenue uh one thing i've hated historically i've
421
00:34:08,604 --> 00:34:16,524
told this to people many times like like we're a small shop it's me i ran pftc by myself for
422
00:34:16,524 --> 00:34:21,244
six seven years and it's only within like last three or four years that we've brought on any
423
00:34:21,244 --> 00:34:28,924
employees and there's five of us and some contractors uh i did all for many years i did
424
00:34:28,924 --> 00:34:36,044
like the recording the editing the publishing uh but the bane of my existence and as funny as what
425
00:34:36,044 --> 00:34:40,924
i did at barstool sports when i worked there was ad sales uh and like putting together decks is
426
00:34:40,924 --> 00:34:46,364
something i'm i appreciate good aesthetics but i don't know i don't know how to manifest them
427
00:34:46,364 --> 00:34:51,324
in the digital world uh very well so like all my decks were like google slide like slapdick
428
00:34:51,324 --> 00:34:58,364
screenshots yeah yeah trust me it's a good show um but uh went and found like a beautiful slide
429
00:34:58,364 --> 00:35:03,084
deck skill connected it to the youtube abi connected it to a fountain api connected it to
430
00:35:03,884 --> 00:35:11,244
um a bunch of other things that are pertinent to putting together a media deck and basically say hey
431
00:35:11,244 --> 00:35:12,984
hey, for example,
432
00:35:13,084 --> 00:35:16,224
like if I was going to try to sell you on an ad deal,
433
00:35:16,304 --> 00:35:16,724
I'd be like, hey,
434
00:35:17,384 --> 00:35:21,444
I would like News Research to be an ad partner with us.
435
00:35:22,204 --> 00:35:26,384
Think about ways in which we would be a good brand fit for them
436
00:35:26,384 --> 00:35:29,264
based off of our demo data,
437
00:35:29,484 --> 00:35:31,724
like our stats on YouTube and whatever,
438
00:35:31,924 --> 00:35:33,384
and like come up with a deck five minutes later.
439
00:35:33,904 --> 00:35:36,944
It's got a HTML page on Vercel,
440
00:35:37,184 --> 00:35:39,084
beautiful deck that can actually sell.
441
00:35:39,324 --> 00:35:40,224
Like on top of that,
442
00:35:40,224 --> 00:35:42,984
It's like, let's make sure we're pricing this right.
443
00:35:43,084 --> 00:35:45,704
Like run market research about podcasts of our size.
444
00:35:45,884 --> 00:35:46,744
Like what is CPM?
445
00:35:47,224 --> 00:35:51,184
Is there any premium on our audience specifically that could do that research
446
00:35:51,184 --> 00:35:52,844
and come up with like accurate pricing?
447
00:35:52,924 --> 00:35:57,704
And that has led to more closed ad deals for us.
448
00:35:57,904 --> 00:35:58,704
So yes, it has.
449
00:35:59,124 --> 00:36:00,544
Along with the way saying yes, it has.
450
00:36:00,844 --> 00:36:04,764
Yeah, but I like what you said too about, you know,
451
00:36:04,764 --> 00:36:08,504
you have a hard time personally manifesting aesthetically pleasing stuff
452
00:36:08,504 --> 00:36:09,024
in the digital age.
453
00:36:09,024 --> 00:36:10,984
but Hermes agent enables you to do that.
454
00:36:10,984 --> 00:36:15,804
And that's what's so cool about it is if you're a high agency,
455
00:36:15,804 --> 00:36:17,404
if you're a high agency person right now,
456
00:36:17,404 --> 00:36:20,904
there's no better time in the history of the world to exist.
457
00:36:21,964 --> 00:36:43,160
Like I can write code anymore but I manage our compute clusters and I can get all of the visualizations all the telemetry that I could possibly want from it just by interfacing in natural language right and build dashboards that help inform our business decisions and allow us to see how we acquiring new users and how our compute clusters health is
458
00:36:43,160 --> 00:36:48,960
and that you know that would take data scientists front-end engineers it would it would take a you
459
00:36:48,960 --> 00:36:53,620
know a series of people to do that in the past and now i'm just prompting hermes agent and it's
460
00:36:53,620 --> 00:36:58,560
building out what a full team would have to build historically so it empowers you to do pretty much
461
00:36:58,560 --> 00:37:04,400
anything in any dimension that you want right in any domain that you want i mean it's it's yeah
462
00:37:04,400 --> 00:37:08,740
it's unbelievable well i mean let's shift to this the compute side what are you what are you seeing
463
00:37:08,740 --> 00:37:15,140
the company you're talking about your cluster what uh on the physical side of things like you
464
00:37:15,140 --> 00:37:20,020
actually need the gpus need to download like what what are you seeing there like what is it like
465
00:37:20,020 --> 00:37:24,500
managing a cluster maybe we'll start there all sorts of headaches managing managing clusters
466
00:37:24,500 --> 00:37:32,340
and computes compute for sure um what i think is really interesting about um the way compute
467
00:37:32,340 --> 00:37:38,020
is shifted it's shifted with use which is i guess obvious right but i think it's worth noting again
468
00:37:38,020 --> 00:37:40,660
We said this time last year, Opus 4.6 wasn't out.
469
00:37:41,220 --> 00:37:43,020
Developers weren't using it to write code.
470
00:37:43,140 --> 00:37:44,600
Good developers weren't using it to write code.
471
00:37:44,680 --> 00:37:47,720
I was using it to write code, but good developers were not using it to write code.
472
00:37:49,960 --> 00:38:04,920
At this time last year, if you needed a cluster of whatever, 128 GPUs, I could send a text out and have 10 different providers bidding to win that.
473
00:38:05,280 --> 00:38:06,640
Now it's the inverse.
474
00:38:06,640 --> 00:38:11,300
I basically would have to go and text 10, 20 different people and say, I need a cluster.
475
00:38:11,420 --> 00:38:11,980
I need a cluster.
476
00:38:12,060 --> 00:38:12,560
I need a cluster.
477
00:38:12,740 --> 00:38:12,960
Nope.
478
00:38:13,320 --> 00:38:14,140
We don't have any.
479
00:38:14,180 --> 00:38:14,620
We don't have any.
480
00:38:14,680 --> 00:38:15,260
We don't have any.
481
00:38:15,540 --> 00:38:22,220
It's unbelievable how squeezed the HPC industry is right now.
482
00:38:22,820 --> 00:38:29,820
I think we're up against a lot of land and power limitations.
483
00:38:30,540 --> 00:38:35,780
I think you're seeing the regulatory environment has just really squashed the ability to continue
484
00:38:35,780 --> 00:38:36,520
to grow that.
485
00:38:36,640 --> 00:38:47,200
But also just the usage was such a parabolic increase in inference needs that we weren't prepared for it, right?
486
00:38:47,200 --> 00:38:55,220
Yeah. And so to that point, who is getting access, obviously, outside the frontier model?
487
00:38:55,680 --> 00:39:02,920
Because you have the training, obviously, you have the inference, and obviously, Anthropic, OpenAI, Meta, Google, SpaceX.
488
00:39:03,240 --> 00:39:05,380
They all have incredible access to this stuff.
489
00:39:05,380 --> 00:39:08,240
But I think we've discussed like open router.
490
00:39:08,380 --> 00:39:11,460
I think Stripe's acquisition of open router is a very strong signal.
491
00:39:11,680 --> 00:39:21,640
And I think this neocloud sector is going to, I think, basically having open market for different models is a very smart idea.
492
00:39:21,740 --> 00:39:28,540
Especially if we believe that being model agnostic is going to be a trend that strengthens in the future.
493
00:39:28,760 --> 00:39:32,360
But to that point, like how do you get priority?
494
00:39:32,360 --> 00:39:38,060
like how do you both on okay let's back up infrastructure side what makes you a good
495
00:39:38,060 --> 00:39:43,480
cluster what do you need to do right for your end customer to make sure people come back to you and
496
00:39:43,480 --> 00:39:46,980
buy compute for you because there's a bunch of stuff with like networking and latency and all
497
00:39:46,980 --> 00:39:55,720
that then to how um for the for the off taker like how do you how are these people the neoclouds
498
00:39:55,720 --> 00:40:00,940
that own the gpus and will inevitably host them how do they decide whether or not you're a good
499
00:40:00,940 --> 00:40:05,340
you're a good off taker that deserves access to you yeah does it even matter it's like you're
500
00:40:05,340 --> 00:40:14,940
paying you yeah um so on the first piece yes it's managing managing these nodes is difficult um
501
00:40:15,900 --> 00:40:20,620
there's a lot of entropy in the management of these nodes that i don't really fully understand
502
00:40:20,620 --> 00:40:26,860
um but there's a lot of down time um with if you're talking about managed inference specifically
503
00:40:26,860 --> 00:40:36,700
there's throughput and latency, throughput latency uptime. These providers do want to work to
504
00:40:37,580 --> 00:40:41,660
have a good experience for the end user because it can be a huge headache and people are paying a
505
00:40:41,660 --> 00:40:45,340
lot of money for these nodes. If their throughput isn't met, if their latency requirements isn't
506
00:40:45,340 --> 00:40:49,740
met, it really ruins the end user experience really, really quickly and people are going to stop
507
00:40:50,300 --> 00:40:55,340
people are going to stop paying for paying for machines from a specific provider right now the
508
00:40:55,340 --> 00:41:01,580
market i think is so squeezed that there's way more wiggle room and tolerance for that stuff
509
00:41:01,580 --> 00:41:09,180
just because it's like where else are you going to go um and to your second question i know that
510
00:41:09,180 --> 00:41:16,220
that plays a role in like who is the end user who is the buyer who is the off taker plays a role in
511
00:41:16,220 --> 00:41:18,320
in the decision-making process of these companies.
512
00:41:18,320 --> 00:41:23,020
I've heard people talk about how, like, this sales cycle for these AEs
513
00:41:23,020 --> 00:41:25,920
is basically flipped before it's like, I need to go find people
514
00:41:25,920 --> 00:41:27,260
to sell this compute to.
515
00:41:27,260 --> 00:41:29,920
Now it's like, I have all these people to go sell this compute to.
516
00:41:29,920 --> 00:41:33,380
I have to build a case of why my people to internal teams,
517
00:41:33,380 --> 00:41:37,420
why my people are worthy of getting the right to this compute contract.
518
00:41:37,420 --> 00:41:39,960
So the thing is completely inverted.
519
00:41:39,997 --> 00:41:41,137
It's fascinating, right?
520
00:41:41,157 --> 00:41:43,157
And I think it's always company.
521
00:41:43,757 --> 00:41:47,597
It's internal politics and kind of company dependent on that stuff, right?
522
00:41:47,597 --> 00:41:52,477
There's different partnerships between companies and favors had, of course.
523
00:41:53,997 --> 00:41:59,897
I think a lot of these companies do try to care about open source
524
00:41:59,897 --> 00:42:03,597
and try to support different parts of the ecosystem, right?
525
00:42:03,597 --> 00:42:16,337
I think people recognize that if all of this compute ends up very, very consolidated, that's a dangerous situation for a multitude of reasons.
526
00:42:18,257 --> 00:42:27,937
So I think, yeah, economics play a big part, but there's also a bunch of kind of strategic initiatives that drive where this compute ends up.
527
00:42:27,937 --> 00:42:30,877
And I'm happy we're discussing this now because I meant to bring this up earlier.
528
00:42:30,877 --> 00:42:33,397
I think it's important to clarify for people
529
00:42:33,397 --> 00:42:34,717
we were talking about open source models
530
00:42:34,717 --> 00:42:36,357
many of which were predominantly
531
00:42:36,357 --> 00:42:39,377
trained in China and then released
532
00:42:39,377 --> 00:42:40,837
people hear that
533
00:42:40,837 --> 00:42:43,537
they're like oh I'm sending all my data to China
534
00:42:43,537 --> 00:42:45,317
now that's not always the case
535
00:42:45,317 --> 00:42:47,497
like if you're connecting directly to
536
00:42:47,497 --> 00:42:49,257
DeepSeq servers like yes you are
537
00:42:49,257 --> 00:42:51,617
but I think the understanding
538
00:42:51,617 --> 00:42:53,057
around like no like it's open source
539
00:42:53,057 --> 00:42:55,757
so you can have American companies
540
00:42:55,757 --> 00:42:57,657
go buy GPUs and download
541
00:42:57,657 --> 00:42:59,577
the model there and you're not serving
542
00:42:59,577 --> 00:43:05,097
it to the chinese servers it's a chinese model run on american servers correct massively different
543
00:43:05,097 --> 00:43:11,097
right and that's been a huge education piece that we're trying to drive home is you can have all of
544
00:43:11,097 --> 00:43:15,577
these cost savings you can get really good performance at you know a fraction of the cost
545
00:43:15,577 --> 00:43:23,257
you know 60 70 80 90 cheaper than u.s closed frontier labs um and you can do it with yes it's
546
00:43:23,257 --> 00:43:28,537
a chinese model but it's running on infrastructure owned by a u.s company in a u.s data center on u.s
547
00:43:28,537 --> 00:43:33,417
soil right and those are where your data flowing is two totally different things yeah well you can
548
00:43:33,417 --> 00:43:37,257
say that people like it's running home to china though right it's like yeah no no it's not it's
549
00:43:37,257 --> 00:43:42,857
definitely not and then uh like the prior again going like tying in privacy to this too this is
550
00:43:42,857 --> 00:43:51,177
what uh i know we were talking about maple ai uh venice has obviously exploded in popularity uh ppq
551
00:43:51,177 --> 00:43:52,077
Here's another one.
552
00:43:55,437 --> 00:43:58,897
I am very bullish on running these open source models
553
00:43:58,897 --> 00:44:00,557
in trusted execution environments,
554
00:44:00,657 --> 00:44:02,057
just like to increase the privacy layer.
555
00:44:02,137 --> 00:44:04,217
Because that, shout out to Mark Goodwin,
556
00:44:04,937 --> 00:44:07,657
who's been on the tip on Twitter saying,
557
00:44:07,757 --> 00:44:09,037
like, even if you're using open source,
558
00:44:09,077 --> 00:44:09,937
it doesn't mean decentralized.
559
00:44:10,197 --> 00:44:12,397
There's still somebody hosting those servers
560
00:44:12,397 --> 00:44:13,197
at the end of the day.
561
00:44:13,197 --> 00:44:15,277
But with like trusted execution environments,
562
00:44:15,357 --> 00:44:18,777
being able to run the inference in a secure enclave
563
00:44:18,777 --> 00:44:25,737
and have everything end-to-end encrypted from your device that's making the prompts to the input
564
00:44:26,937 --> 00:44:32,297
in transit is encrypted then you do everything on the secure enclave you re-crypt it set it back like
565
00:44:32,297 --> 00:44:37,257
i think that's a massive opportunity too um i'm just rambling now yeah and i think i think that
566
00:44:37,257 --> 00:44:42,537
and i think i think local models local models as well yeah and local models are continuing to get
567
00:44:42,537 --> 00:44:48,457
better hardware consumer hardware is continuing to get better and better right and so the ability
568
00:44:48,457 --> 00:44:54,057
for you to run a model on your local machine and do again it's not going to be a hundred percent of
569
00:44:54,057 --> 00:44:58,537
your work but a good chunk of your work especially if you have sensitive workloads like you can run
570
00:44:58,537 --> 00:45:03,577
those on local models and and get get real work done well this is actually something i'm happy you
571
00:45:03,577 --> 00:45:09,177
brought this up this is something i've been like thinking like i'm not smart i'm maybe i'm not
572
00:45:10,377 --> 00:45:16,217
you're smart i won't denigrate myself i'm smart but i do not have the skills to post train an
573
00:45:16,217 --> 00:45:20,537
an open source model on my own data as like something of like her if noose
574
00:45:20,957 --> 00:45:23,937
figure out a way to take what I built with my Hermes agent and be like, okay,
575
00:45:23,937 --> 00:45:30,437
I want to use this model. Um, this local, this SLM, like locally, uh,
576
00:45:30,557 --> 00:45:32,897
Hermes agent, you have all the context of what we've been working on over the
577
00:45:32,897 --> 00:45:36,857
last six months. Like, can you help me post train this model on our specific
578
00:45:36,857 --> 00:45:40,397
data? Like that, I think, at least for my purposes, that's, that's one thing
579
00:45:40,397 --> 00:45:44,697
I've been, uh, somewhat obsessed with for the last week or two is like, okay,
580
00:45:44,697 --> 00:45:45,537
I've gotten to this point.
581
00:45:45,637 --> 00:45:46,277
Like, what else can I do?
582
00:45:46,357 --> 00:45:48,497
Yeah, how can you squeeze more performance out of this?
583
00:45:48,497 --> 00:45:49,957
I want to post-trained a model on all of our stuff.
584
00:45:50,117 --> 00:45:50,977
Like, how do I do that?
585
00:45:51,077 --> 00:45:51,197
Yeah.
586
00:45:52,217 --> 00:45:53,097
That's what I'm looking for.
587
00:45:53,197 --> 00:45:54,457
Anybody who wants to help me, let me know.
588
00:45:54,677 --> 00:45:55,457
Yeah, we can do that.
589
00:45:55,577 --> 00:45:56,417
We can get you.
590
00:45:56,517 --> 00:46:00,837
So that actually was part of the reason that Hermes Agent was created.
591
00:46:00,837 --> 00:46:04,177
It wasn't created to be this outward-facing product.
592
00:46:04,357 --> 00:46:08,257
It was created really out of, like, a completely utilitarian birth.
593
00:46:08,257 --> 00:46:16,217
our co-founder Technium created this Hermes agent to help as a small AI research lab, right? You're
594
00:46:16,217 --> 00:46:23,277
always trying to find like these moonshot ideas to push you to the frontier. And Technium created
595
00:46:23,277 --> 00:46:32,037
this self-evolving Hermes agent to help us, you know, become frontier model trainers. So the
596
00:46:32,037 --> 00:46:38,017
original birth of Hermes agent was really to do, was in part to do that. And that's something we
597
00:46:38,017 --> 00:46:44,037
should do a job of highlighting um how you can use hermy's agent in that stuff as well yeah so
598
00:46:44,037 --> 00:46:59,013
what uh what the plan with noose like you have the agent you have this portal we got this enterprise what uh what are you guys striving for um in terms of what this company looks like um and i know you said this on the
599
00:46:59,013 --> 00:47:04,613
phone like who knows what any company looks like like what are your what are your north stars and
600
00:47:04,613 --> 00:47:10,133
who are you building for i think yeah yeah i think um who are we building for is is everybody and i
601
00:47:10,133 --> 00:47:16,933
think i i answer that really intentionally that way um because i think we've we've built this
602
00:47:16,933 --> 00:47:23,473
We've built this product in a way that it meets you wherever you are.
603
00:47:24,053 --> 00:47:34,753
The open source repo, if you want to run a local model on your local network and never touch any news research managed product, that's great.
604
00:47:34,753 --> 00:47:39,953
The Hermes agent repo will always be able to empower you to do that and will continue to improve it and maintain it.
605
00:47:41,233 --> 00:47:45,553
If you want certain things managed, great, we'll do all that as well.
606
00:47:45,553 --> 00:47:50,753
We are the best place to have your Hermes agent hosted and managed.
607
00:47:51,233 --> 00:47:56,333
If you want different pieces managed, but you want to own certain pieces or do some work locally, great.
608
00:47:56,473 --> 00:47:57,993
We'll meet you wherever we are.
609
00:47:58,053 --> 00:48:01,693
So we really are building this for everybody.
610
00:48:01,873 --> 00:48:09,553
I think the two sort of key outcomes over the next handful of months for us,
611
00:48:09,553 --> 00:48:16,793
which lead to where we want to be in the end is one, making, decreasing the amount of time that
612
00:48:16,793 --> 00:48:23,273
you go from Googling Hermes agent or having your friends send you a Hermes agent to realizing,
613
00:48:23,273 --> 00:48:30,753
wow, this is a magical, magical experience, right? How can we get that time? So how can we
614
00:48:30,753 --> 00:48:34,833
make it a frictionless onboarding? How can we meet people on whatever device they are?
615
00:48:34,833 --> 00:48:40,373
how can we enable people to send others there, send others a Hermes agent? Oh, Marty hasn't
616
00:48:40,373 --> 00:48:45,353
used Hermes agent. Let me send it to you. You can click a link, open it up and ask it one question
617
00:48:45,353 --> 00:48:50,393
and be like, wow, and not have to do, you know, set up, not have to configure anything. So that's
618
00:48:50,393 --> 00:48:53,853
really, really important, right? Is how can we make sure it is frictionless as possible and make
619
00:48:53,853 --> 00:48:59,613
it so that you as quickly as possible realize that you have a tool that you've never experienced
620
00:48:59,613 --> 00:49:08,933
before in your hands. And then the second piece is delivering Hermes agent to enterprise in a way
621
00:49:08,933 --> 00:49:16,293
that they are able to control their AI stack. It is their sovereign AI stack with all the bells and
622
00:49:16,293 --> 00:49:25,213
whistles that enable employees to be much higher agency, much more output, and do it in a way that
623
00:49:25,213 --> 00:49:31,513
isn't prohibitively expensive, that leverages the plethora of models and tools out there
624
00:49:31,513 --> 00:49:41,033
that can power Hermes agent in a meaningful way, delivering that product to enterprise at scale.
625
00:49:41,193 --> 00:49:45,473
So yeah, those two things. Can we trim down that timing? Can we deliver to enterprise at scale?
626
00:49:46,473 --> 00:49:52,273
The meet people on their device, that's been one of the interesting observations I've had
627
00:49:52,273 --> 00:49:58,353
over the course of the year is uh who attracts to what specific form factor like i was telling you
628
00:49:58,353 --> 00:50:05,553
like i haven't used cloud code or codex desktop in seven months people like are you doing that like
629
00:50:05,553 --> 00:50:10,353
are you are you trying to get left behind i'm like no i literally it's connected to all these
630
00:50:10,353 --> 00:50:16,513
different models and my go-to form factors is like voice to text in telegram i have a group chat with
631
00:50:16,513 --> 00:50:21,473
the agent just the agent of me but the group chats allow you to create topics yep so we have topics
632
00:50:21,473 --> 00:50:29,633
to represent different sessions that are ad ops bitcoin brief uh or the commoner newsletter builder
633
00:50:29,633 --> 00:50:35,713
um uh if we have a business meeting okay like biz ops what do we need to get done blah blah blah
634
00:50:36,273 --> 00:50:41,313
uh and then there's something about that form factor is just vibes with me like just being able
635
00:50:41,313 --> 00:50:46,593
to like on the go on a walk around the neighborhood and you saw me literally when you walked in i had
636
00:50:46,593 --> 00:50:49,093
my newsletter session open,
637
00:50:49,213 --> 00:50:51,313
like talking about how I want to frame today's newsletter
638
00:50:51,313 --> 00:50:55,293
and just that form factor being able to walk on the go
639
00:50:55,293 --> 00:50:56,773
and do what my phone clicks with me.
640
00:50:57,033 --> 00:50:57,873
For others, it's like,
641
00:50:58,593 --> 00:51:00,593
and I think a lot of people are talking about vendor lock
642
00:51:00,593 --> 00:51:04,633
and people are locked into Cloud Code
643
00:51:04,633 --> 00:51:07,333
and Codex specifically, those desktop apps.
644
00:51:07,453 --> 00:51:07,933
And they think like,
645
00:51:08,013 --> 00:51:09,453
this is the only way that you can leverage this.
646
00:51:09,593 --> 00:51:13,493
I'm like, no, you got to get yourself off that tip
647
00:51:13,493 --> 00:51:14,893
as quickly as possible
648
00:51:14,893 --> 00:51:20,413
because you're going to get vendor locked in and that's not what you want um but like everyone but
649
00:51:20,413 --> 00:51:25,773
something like ed on our team he loves the uh he loves the hermese desktop app that's how he
650
00:51:25,773 --> 00:51:30,093
interacts with it and i'm i'd never frankly i never use it yeah just use it through telegram
651
00:51:30,093 --> 00:51:37,373
um and i think there is it's becoming clear to me that there's going to be different preferences
652
00:51:37,373 --> 00:51:41,453
different form factors for how you interact with this and thinking about a company like
653
00:51:41,453 --> 00:51:45,433
like news research trying to design for all those different ways which people are going
654
00:51:45,433 --> 00:51:49,053
to prefer to interact with it is it's a crazy problem to think about.
655
00:51:49,053 --> 00:51:51,773
Yeah, it's a it's a it's a massive problem, right?
656
00:51:51,773 --> 00:51:54,273
And then you're exactly right.
657
00:51:54,273 --> 00:52:10,309
There not there isn some you know golden solution where everybody is going to use this as the way to interface right At least certainly not yet And I think the stance that we generally take is don be like we definitely opinionated on some of the best ways to use these things
658
00:52:10,309 --> 00:52:13,769
but we don't wanna ever lock in unless it's blatantly obvious
659
00:52:13,769 --> 00:52:15,949
that that is the place that we need to lock in, right?
660
00:52:15,949 --> 00:52:20,449
So can we, so we allow WhatsApp, we allow Telegram,
661
00:52:20,449 --> 00:52:23,149
we allow email, like wherever you want this thing to live
662
00:52:23,149 --> 00:52:26,909
and meet you is important because people use it,
663
00:52:26,909 --> 00:52:27,949
people use it in different ways.
664
00:52:27,949 --> 00:52:29,889
There's a lot of people that would never wanna talk
665
00:52:29,889 --> 00:52:34,449
want to talk as much as you right like they would never want to do that everybody is so different
666
00:52:34,449 --> 00:52:38,369
and everybody finds different ways to get the performance out of these agents that they want
667
00:52:38,369 --> 00:52:45,649
that you you need to build for a huge range of different um platforms yeah yeah it's funny when
668
00:52:45,649 --> 00:52:53,169
we like as you're mentioning that like email the hermes agent has uh an agent.2 mail uh email
669
00:52:53,169 --> 00:52:59,489
address um instead like for my wife whenever she's like ah we're in we're in the market for a new car
670
00:52:59,889 --> 00:53:04,769
and she's like oh she's like doing it the old way i'm like hey just email martin i gave him
671
00:53:04,769 --> 00:53:08,529
permission to email you back like tell him what you want he'll go do like a search for you so like
672
00:53:08,529 --> 00:53:14,609
my wife has emailed martin the more sophisticated marty is uh is doing all this work but it's like
673
00:53:14,609 --> 00:53:19,009
it's been fun watching my wife just i mean just she'll be like oh i emailed martin today and he
674
00:53:19,009 --> 00:53:23,969
like did some good work for me like that's pretty does she like martin more than you no no of course
675
00:53:23,969 --> 00:53:30,849
i hope not not yet at least maybe at some point maybe when robotics get here shoot
676
00:53:32,289 --> 00:53:39,809
i'm kidding i'm kidding the uh but um no it's just like the yeah like that's one form factor
677
00:53:39,809 --> 00:53:43,809
where it's like hey honey instead of like you take my phone and getting telegram just email martin
678
00:53:43,809 --> 00:53:48,849
yeah yeah he'll get you a list of cars in the area that are within the the range of
679
00:53:49,649 --> 00:53:53,649
what you're looking for yeah she's just been going back and forth with them in an email thread
680
00:53:53,649 --> 00:53:58,929
that's yeah it's probably a magical experience for it right yeah yeah yeah and awful and awful
681
00:53:58,929 --> 00:54:03,649
it's all of those monotonous tasks like going and searching and getting quotes from six different
682
00:54:03,649 --> 00:54:08,769
from six different vendors and finding the car you want and the reviews all of that is it goes from
683
00:54:08,769 --> 00:54:14,609
these incredibly laborious processes to you know a few minutes texting your hermes agent
684
00:54:14,609 --> 00:54:22,129
no no we have uh ed on our team he has his agent uh and our agents are emailing each other yeah
685
00:54:22,129 --> 00:54:27,969
like stuff we have stuff for the business like ed in his telegram or his hermes desktop app with
686
00:54:27,969 --> 00:54:33,889
his age like all right email martin um and tell him that we did this thing in this part of the
687
00:54:33,889 --> 00:54:39,729
business and make sure he updates his second brain to yeah to recognize that it's uh and i feel i still
688
00:54:39,729 --> 00:54:45,329
feel like i'm not utilizing it like i'm like we do that and something like a workflow like that will
689
00:54:45,329 --> 00:54:48,929
uh emerge i'm like is this the right way to do something like this like is there there are more
690
00:54:48,929 --> 00:54:55,489
efficient way and so the point being is there's so much to uh to explore in terms of all these
691
00:54:55,489 --> 00:55:00,289
different workflows in these cases particularly teams like thinking about teams interact with
692
00:55:01,249 --> 00:55:08,609
agents and i think dorsey's interview with sequoia earlier this year really uh really
693
00:55:08,609 --> 00:55:13,889
cemented my thinking about like how to view a company in the world of agentic ai or something
694
00:55:13,889 --> 00:55:19,809
the company is like essentially like a knowledge layer and your job is to give it the best most
695
00:55:19,809 --> 00:55:28,209
robust knowledge that every employee can access and so um i haven't let everybody into the group
696
00:55:28,209 --> 00:55:32,449
chat on telegram however we do have like discord and the agents run wild and discord and they can
697
00:55:32,449 --> 00:55:39,489
talk there and interact with them so that that use case of creating a second second brain for
698
00:55:39,489 --> 00:55:46,209
a company specifically and then getting access the right access to individuals on the team is a
699
00:55:46,209 --> 00:55:51,969
massive problem uh not a problem massive opportunity uh because when you think about like permissions
700
00:55:51,969 --> 00:55:55,809
especially if you're like selling enterprises like okay i want becky from finance to have access to
701
00:55:55,809 --> 00:56:00,769
this data but not this data yeah um and like just building the guardrails for that yep totally
702
00:56:00,769 --> 00:56:05,729
massive opportunity yeah it's a massive opportunity i think there's a bunch there too as far as um
703
00:56:05,729 --> 00:56:10,929
how can you use Hermes agent to get more out of every employee?
704
00:56:11,569 --> 00:56:14,189
And I think what you see is certain employees,
705
00:56:14,249 --> 00:56:16,349
it's a skill to be able to prompt these agents.
706
00:56:16,949 --> 00:56:19,689
Different people get different performance out of agents.
707
00:56:19,849 --> 00:56:21,369
It's very much a feel thing.
708
00:56:21,909 --> 00:56:27,229
And I think you have people that are incredibly, incredibly proficient,
709
00:56:27,309 --> 00:56:28,509
like your top performers.
710
00:56:30,289 --> 00:56:33,369
Hermes agent lets you bring other people up to their level.
711
00:56:33,369 --> 00:56:46,629
And that's what we're rolling out for enterprises is this concept of a collective wisdom, a collective brain where you can take the performant employees' top skills and deliver those to relevant employees.
712
00:56:46,909 --> 00:56:48,849
And they immediately get boosts in their performance.
713
00:56:48,949 --> 00:56:50,429
Their agents are immediately better.
714
00:56:50,769 --> 00:56:56,789
They're shaped in the way that you want them shaped, however your top employees perform, if you want other people like that.
715
00:56:56,949 --> 00:56:59,769
It will shape other people's agents in that way, too.
716
00:56:59,769 --> 00:57:12,109
And when you have new hires, instead of having like this cold start on-ramp experience, you can equip them with a skill set of tools that are formulated, you know, in the way that TFTC wants things written, right?
717
00:57:12,249 --> 00:57:22,786
And you know all of that stuff And you can have your agents molded kind of to your voice and the way that you want things done your guidelines
718
00:57:23,246 --> 00:57:26,366
And you can give that to new employees, existing employees, et cetera,
719
00:57:26,366 --> 00:57:28,346
to kind of help shape people in your direction
720
00:57:28,346 --> 00:57:30,366
and also give them that productivity boost.
721
00:57:30,766 --> 00:57:30,926
Yeah.
722
00:57:32,826 --> 00:57:34,466
What do you think this does for the job, Mark?
723
00:57:34,826 --> 00:57:39,306
Are you, the permanent underclass is quickly approaching
724
00:57:39,306 --> 00:57:42,346
or a world of a billion flowers are going to bloom?
725
00:57:42,486 --> 00:57:47,106
I think we could have a whole nother podcast on that.
726
00:57:47,106 --> 00:57:51,566
I think I land somewhere in between, right?
727
00:57:51,626 --> 00:58:01,306
I think that what this does is, you know, it enables you to not do all of these monotonous tasks, right?
728
00:58:01,406 --> 00:58:03,506
If you've, I think, my bigger...
729
00:58:03,506 --> 00:58:05,086
Oh, my God, I'm going to lose my laptop job.
730
00:58:05,086 --> 00:58:12,266
Yeah, my bigger concern is, oh, you're worried about not just like, you know, making Excel sheets for the rest of your life or whatever.
731
00:58:12,486 --> 00:58:14,306
Like that's your really worried about the paycheck.
732
00:58:15,506 --> 00:58:15,906
Yeah.
733
00:58:15,906 --> 00:58:24,506
Like I think what this what this should do is free people up to do, you know, more thoughtful and higher agency tasks than sitting there and like button pushing.
734
00:58:26,026 --> 00:58:29,706
So, yeah, I mean, is it going to be this straight shot to a million flowers?
735
00:58:30,006 --> 00:58:30,426
No.
736
00:58:30,726 --> 00:58:39,866
Like there's, of course, going to be bumps in the job market is, you know, these disruptive technologies, you know, move through their natural growth and integration and flow.
737
00:58:39,866 --> 00:58:54,886
So but I think if you the optimistic view, right, is that this will enable humans to do things that are more way more fulfilling than what a lot of people have convinced themselves is is foundational to their lives.
738
00:58:54,886 --> 00:59:04,106
Yeah, the that's sort of philosophical musing I've had over the last year, because you mentioned earlier, like if you're high agency and get a lot out of this.
739
00:59:04,106 --> 00:59:07,026
And that's the question I have.
740
00:59:07,026 --> 00:59:16,666
Is there just like a natural distribution of individuals who are inherently high agency and a very large percentage of which are not?
741
00:59:16,666 --> 00:59:25,726
Or is there an ability to engender high agency in people and do AI tools inspire them to become a high agency?
742
00:59:26,606 --> 00:59:29,386
I don't know. I think that's what we're going to find out in the next decade.
743
00:59:29,666 --> 00:59:32,566
We'll find that out. I'm sure it's some mixture of both.
744
00:59:32,566 --> 00:59:39,106
Yeah. What's your most contrarian take on what's going on right now?
745
00:59:40,146 --> 00:59:48,926
My most contrarian take of what's going on right now. I don't know if it's contrarian. I'm sure it's,
746
00:59:49,606 --> 00:59:53,566
there's a lot of people I think that probably show this viewpoint. I never know how
747
00:59:54,226 --> 00:59:59,286
pigeonholed I am, or sorry, how echo chambered I am. But I think that
748
00:59:59,286 --> 01:00:09,286
if you want to talk about like existential risk and the way that existential risk of AI,
749
01:00:09,286 --> 01:00:13,406
like AI, you know, misaligned AI killing all humans or whatever.
750
01:00:15,006 --> 01:00:19,306
I actually like sometimes believe in there's some, I believe there's some degree of credibility
751
01:00:19,306 --> 01:00:25,286
to that argument as a, as a possibility. But what I think is, I think where we get to crazy
752
01:00:25,286 --> 01:00:31,906
town really fast and it's really, really disingenuous is you see large companies and
753
01:00:31,906 --> 01:00:37,806
lobbyists pushing for a stop of open source, right?
754
01:00:37,806 --> 01:00:44,966
Like open source has become this evil and this scapegoat.
755
01:00:45,906 --> 01:00:51,086
Whereas if you ever actually believed in existential risk as a real threat, open source would be
756
01:00:51,086 --> 01:00:52,566
the last thing you would be concerned about.
757
01:00:52,566 --> 01:00:55,306
You'd be concerned about this massive consolidation of compute.
758
01:00:55,586 --> 01:01:01,746
You'd be concerned about the frontier constantly being pushed by the closed labs.
759
01:01:02,546 --> 01:01:10,226
Distillation to create open source models definitionally can't exceed AGI unless the top labs already have done it themselves.
760
01:01:10,226 --> 01:01:33,046
So I think the way that that argument has been pushed and is certainly kind of the controlling argument in the vast majority of the population, likely not those that are very in tune with AI, but the way that that argument has been framed to the vast majority of the population, I think is just incredibly, incredibly disingenuous.
761
01:01:33,046 --> 01:01:41,206
incredibly and it's almost disgusting at the point because it's like if i i don't think
762
01:01:42,406 --> 01:01:45,846
most of the public is aware because they don't pay close enough attention it's like very clear
763
01:01:45,846 --> 01:01:50,326
to me it's like okay you're looking for the you've raised a ton of capital you're deploying a ton of
764
01:01:50,326 --> 01:01:58,806
capital into this infrastructure built out really for the frontier labs uh and you have a need to
765
01:01:58,806 --> 01:02:01,506
to get a return on invested capital.
766
01:02:01,506 --> 01:02:03,246
And you're seeing these open source models
767
01:02:03,246 --> 01:02:04,306
begin to nibble at your feet.
768
01:02:04,306 --> 01:02:06,446
Not only that, like many people are turning to them
769
01:02:06,446 --> 01:02:11,046
for 90% of the work, but they're individual use cases.
770
01:02:11,046 --> 01:02:12,986
And you're trying to pull up the ladder
771
01:02:12,986 --> 01:02:15,326
and shut the door behind you
772
01:02:15,326 --> 01:02:16,966
just so you can save those profits.
773
01:02:16,966 --> 01:02:21,466
And I think when you consider how profound of a shift
774
01:02:21,466 --> 01:02:23,166
this technology is for humanity,
775
01:02:23,166 --> 01:02:25,866
the fact that there's literally two or three companies
776
01:02:25,866 --> 01:02:27,526
trying to pull up the ladder and control it,
777
01:02:27,526 --> 01:02:29,946
just incredibly
778
01:02:29,982 --> 01:02:32,662
I would say evil at the end of the day.
779
01:02:32,682 --> 01:02:34,502
Well, where did all the data come in the first place?
780
01:02:34,622 --> 01:02:34,862
Right.
781
01:02:35,222 --> 01:02:36,622
Like it's all of our data.
782
01:02:36,782 --> 01:02:40,142
I mean, the fact that they were buying like centuries old books and like cutting off the
783
01:02:40,142 --> 01:02:41,002
bindings and.
784
01:02:41,562 --> 01:02:41,802
Yeah.
785
01:02:41,802 --> 01:02:42,662
And then burning them.
786
01:02:42,742 --> 01:02:43,382
It's like, what the hell?
787
01:02:43,482 --> 01:02:43,682
Yeah.
788
01:02:43,782 --> 01:02:46,642
It was all our data that it came from in the first place.
789
01:02:46,702 --> 01:02:46,902
Right.
790
01:02:46,942 --> 01:02:49,402
And then they, yeah, it's, it's just as ingenious.
791
01:02:49,402 --> 01:02:50,982
Dealing with it with the New York times or not, I believe.
792
01:02:51,102 --> 01:02:51,202
Yeah.
793
01:02:51,442 --> 01:02:57,582
I think that, that, uh, I guess copywriter and copyright case is moving forward.
794
01:02:58,482 --> 01:02:59,262
I mean, it's been a pleasure.
795
01:02:59,262 --> 01:03:05,522
um i'm sure there's more i am gonna pick your brain off uh off air about what i should be doing
796
01:03:05,522 --> 01:03:10,842
with my agent i love it any any final thoughts final final words of wisdom before we wrap up here
797
01:03:10,842 --> 01:03:18,842
i don't know about words of wisdom i think um i think at some point it will be very obvious that
798
01:03:18,842 --> 01:03:28,982
you should care about um privacy in in the age of ai and agentic ai um
799
01:03:28,982 --> 01:03:48,942
I think that the more proactive you can be about it and the steps that you can take to empower a world where you have more ownership as a business and as an individual of your AI and where the data that you give to your AI flows, the better served you'll be.
800
01:03:48,942 --> 01:03:51,702
and you can avoid a lot of headache down the road.
801
01:03:51,822 --> 01:03:54,902
And so I think looking for, obviously, Hermes agent,
802
01:03:55,022 --> 01:03:58,302
but looking for tools that enable you in the AI realm
803
01:03:58,302 --> 01:04:01,962
to have good controls over your data,
804
01:04:01,982 --> 01:04:04,362
but do it in a simple way that isn't super burdensome
805
01:04:04,362 --> 01:04:05,802
and a huge headache for you,
806
01:04:05,802 --> 01:04:08,962
I think will be a really, really valuable investment going forward.
807
01:04:09,642 --> 01:04:10,622
I completely agree with that.
808
01:04:12,222 --> 01:04:13,162
Thank you for joining me.
809
01:04:13,262 --> 01:04:16,162
Thank you and the team at Noose for building Hermes.
810
01:04:16,162 --> 01:04:22,242
has been incredibly uh powerful for us here at tftc and uh excited to see what you guys build next
811
01:04:22,242 --> 01:04:26,162
love it thanks marty thank you peace and love freaks