Latest / Agentic Conversations (formally mlops.community) / The Centralization of Power in AI // Kyle Harrison // # 181

Transcript

The Centralization of Power in AI // Kyle Harrison // # 181 · Agentic Conversations (formally mlops.community)

0:00
I'm Kyle Harrison, General partner at Contrary and I don't
0:03
drink coffee, but I have a crippling addiction to Diet
0:06
Coke. Welcome, welcome everyone.
0:08
We are back with another ML OPS community podcast.
0:11
I'm your host, Dimitrios, and today I am talking with none
0:15
other than Kyle, who you just heard say he doesn't like
0:18
coffee. No, this conversation was
0:21
fascinating on so many different levels.
0:23
I really appreciate the report that he put together, which if
0:27
you do not know is called the openness of AI.
0:30
We'll leave a link to that in the description.
0:33
If you haven't read it yet, go and check it out he got into.
0:38
Why she wrote the report, what the reports all about,
0:40
obviously, And then the dangers that he calls out in the report.
0:45
And one of those dangers being the centralization of power when
0:49
it comes to AI. And really we're talking about
0:52
open AI a lot in this conversation and the pros and
0:56
cons that come with open AI. I thought it was fascinating how
1:00
deep he went into the privacy part.
1:04
And for me it felt like or it feels like.
1:08
Lots of people's privacy is one of those things like peace on
1:13
Earth and the data mesh. Everyone wants it, but at the
1:18
end of the day, you're willing to sacrifice a lot.
1:22
If there is an easier and less friction way of doing something,
1:27
and to timestamp this, we are recording it about a day or two
1:31
after ChatGPT Enterprise came out, so we talked a lot about
1:36
that and how trustworthy Open AI is when it comes to dealing with
1:42
your data. And are they going to train GTA5
1:45
on your data? Well, maybe, maybe not.
1:47
I thought we had a really cool discussion around that and the
1:51
potential. Is there?
1:53
And so his ability to be super bullish on lots of companies
1:58
opting away from using Open AI was fascinating.
2:02
But he also talked in the report in in this conversation about
2:08
how opening strategy right now, open AI strategy right now is to
2:13
just proliferate the market and get into the hands of everyone
2:18
and be the best and easiest use case of.
2:22
Trying to do anything with AI because that's the playbook.
2:26
As soon as they are in everyone's hands, then three
2:29
years down the road, once they've trained whatever on
2:33
everyone's data and we don't like it and we want to rise up,
2:36
it's going to be a lot harder to rip off that Band-Aid.
2:39
So we get into all of that and more in this conversation.
2:43
Kyle is a general partner at Contrary.
2:46
He leads Series A and growth stage Investing and he joined
2:52
them when he was a partner. When he went up to general
2:55
partner, he's definitely making some moves his portfolio.
2:59
This is what was awesome to me. We talked a lot about his
3:02
portfolio companies in this call, some of them being ramp
3:08
replete. Cohere.
3:12
Snowflake and Databricks, that's interesting one that he went
3:16
with snowflakes and data bricks. And by the way, he's also got a
3:22
sub stack where he regularly shares his analysis on the
3:25
venture capital landscape. You can click the link in the
3:29
description to check that out. It's called Investing 101.
3:34
So let's get into it before we do.
3:37
The thing that would mean the world to me is if you can share
3:40
this episode with one friend that you think.
3:43
It would be valuable to that would mean the world to me.
3:45
So let's go. here to talk about all things AI, and I want to
3:57
dive into the hype around AI. You've been going deep into the
4:03
openness and the closeness of AI.
4:05
You've also been just thinking a lot about.
4:08
The hype and if there's real businesses to be built on top of
4:12
this. Can you just give us your Where
4:16
are you sitting right now? What's your angle?
4:18
What? How are you looking at this
4:19
space? Absolutely.
4:22
The most interesting thing about what we're going through is, I
4:25
mean by note, you know, by every stretch of the imagination.
4:28
This is a huge paradigm shift in technology and everybody sort of
4:33
appreciating it. The one thing that I think is
4:36
most interesting when I talk to companies that have operating
4:38
the space people have been building in this world for a
4:40
long time, for years and years and is that so many of those
4:44
people they feel like they have been maybe not standing still
4:47
but like slow and steady wins the race and then all of a
4:50
sudden change. BT was this wake up call for a
4:53
lot of people, but there wasn't necessarily this like
4:56
fundamental change in the underlying technology.
5:00
It had been slowly, pretty transformer architecture has
5:02
been around for a long time. It had slowly been improving and
5:05
what it was capable of Chat CBT was just a thing that put it
5:08
front and centre in everybody's mind.
5:10
And so really what you're going through is this like awareness
5:13
of the capability that exists and and how it's changing and
5:16
stuff like that. That's super powerful.
5:18
No doubt. What you are now seeing is this
5:21
kind of bifurcation of where companies are positioned in this
5:25
space is pretty interesting and you effectively have 3 buckets.
5:28
So the first bucket you have these incumbents that have
5:31
established products, established distribution,
5:34
they're they're massive, whether it's like companies like
5:36
Microsoft or companies like Canvas, Adobe or whatever notion
5:41
exactly, they can, these incumbents can roll this stuff
5:44
out. That's one bucket.
5:45
The second bucket you have folks that are trying to be the
5:48
foundation model, right. They want to be the provider
5:50
that as many people as possible for as many use cases as
5:53
possible are building. And that requires pretty
5:57
extensive training, data access capital, all this stuff.
6:02
Right. And that's exactly and it's a
6:04
totally, it's a very unique game.
6:07
And then you have everything else, every other company that's
6:11
trying to build some kind of AI tooling, orchestration, fine
6:15
tuning, you know whatever other sort of for the whole world
6:19
which we can dive into and and is a place I think it can be
6:22
really powerful but and then there's these this sort of use
6:27
case tools where it's like they have built you know people the
6:30
the sort of pejorative term is the wrapper around ChatGPT,
6:33
right and not every company, is that right?
6:36
I think that you can leverage EPT and not be just a wrapper,
6:39
not just middleware, but there is a spectrum of companies or
6:43
some of them are literally just you know, they slapped some some
6:46
UI and and chastity is the entire back end, right.
6:50
But that whole swath of companies is the majority of AI
6:54
companies. The biggest thing that I think
6:57
is problematic is that a lot of the power and influence and and
7:02
sort of capabilities and stuff, a lot of that happens in this
7:06
middle bucket where companies are trying to build the
7:09
foundation models. And that is probably the area
7:13
where there is the most potential for centralization of
7:17
like power and influence and stuff like that.
7:19
And when we wrote so for country research, we wrote this report,
7:21
the openness of AI. And that is a big section of
7:24
what we dove into is what does that mean when you have
7:27
companies that become these really important foundation for
7:30
everything else when they are pretty centralised and pretty
7:34
controlled by a small subset of companies?
7:37
Yeah, and they have all the attention.
7:39
It's like they've sucked the air out of the room.
7:43
Yeah, I mean, I'd say they sucked the, it's, it's basically
7:45
like attention. It's the GPU to your point, it's
7:48
the capital, like the headlines. There's so much fervour focused
7:53
on this. I honestly like rightfully so to
7:55
some extent, right. Like I think that these
7:57
businesses are going to be really critical players as this
8:02
evolves and even you look at like the acquisition that Open
8:06
AI made where it's literally a design studio and it kind of
8:09
demonstrates their intentions to build more consumer facing
8:12
products, not just be an API product that's powerful legs.
8:16
So these companies are important, but I think the thing
8:19
you need to step back is understand the trickle effect
8:22
of, you know, who has the capabilities at this point in
8:25
time and what does that mean for every other company, for better
8:28
or for worse. Yeah, it's funny you talk about
8:32
this, the idea of the rapper or I I kind of look at it as the
8:37
application layer and it feels like there's the developer tools
8:42
that are being built for ML or now the hot new term is LM OPS,
8:48
right? And So what was ML OPS now has
8:51
died and gotten reincarnated as LM OPS.
8:56
And so it's so funny. Yeah, I've heard.
9:00
I've heard. Including other VCs say, yeah,
9:04
you know, like we would do a hackathon with you, but we just
9:08
worth talking about the name and ML OPS doesn't seem like it's in
9:13
style anymore. Break me?
9:16
Yeah, seriously, Talk about term of the month.
9:18
Geez. Yeah, exactly.
9:20
So the the thing about this application layer is that there
9:25
are some there's revenue being generated there and there's
9:28
serious revenue being generated there from.
9:32
Like. Zero to 1,000,000 really quickly
9:37
and I've talked about this before.
9:39
What I think is fascinating with that is.
9:42
The quality of that revenue and the ability to create that mode.
9:47
So what happens when a Open AI goes and they have their design
9:51
studio now replicate what you're trying to do and all of a sudden
9:56
it's like Open AI has all of the attention and the air in the
10:00
room. And so they put out the same
10:03
product as you, but now they get all of that.
10:06
Love and all of those dollars go towards it.
10:09
And so building those moats is something that I know a lot of
10:13
the community has been thinking about and how to effectively
10:17
build a Moat. And it seems like in that bucket
10:20
that you were talking about where where it's like the
10:22
notions or the Microsoft, they already have their vote and
10:25
they're just incorporating AI into their products, but it's
10:29
not their main product. It's on the other side of things
10:33
where AI is your main product and you're leading with that.
10:37
That's where I feel like things are very dicey and it's a
10:39
bloodbath right now. It is the The way that I think
10:44
about it is. I mean #1.
10:47
If Open AI can roll out your product in a couple of months,
10:50
you never had a note to begin with.
10:51
For sure like there there are some elements of like how much
10:55
does a competitive advantage really exist versus just a first
10:58
mover advantage and the vast majority of that stuff is is
11:04
first mover advantage right. Like they were the 1st to get
11:06
there to do that thing. There's a lot of the sort of you
11:09
know copy dot AI and Jasper and stuff like that like those types
11:11
of companies they were just the first to get their leveraging
11:14
GBT to produce copy and they branded themselves as a
11:18
copywriters tool and and you know and that's the first mover
11:22
advantage. It's not.
11:23
It's not a it's not a competitive mode and it's not
11:26
defensible. And the way that we that I look
11:29
at companies today and think about defensibility and their
11:33
capabilities is you have to spike in something that is not
11:37
the thing that everyone else is racing to commoditize right.
11:40
If if the thing that you're good at is the thing that they, that
11:43
everybody else who has billions of dollars is rapidly trying to
11:47
commoditize because they want to get into the install base and be
11:51
the thing that everybody's built on top of, Like you're screwed
11:55
because you're just going to get swept under the rug in that race
11:58
to the bottom of commodities, right?
12:00
And so we've made investments in companies take for example like
12:03
replete where replica is not it's, it's, I would I would
12:08
argue it is not an AI company in exclusivity, right.
12:11
They don't think about themselves as we build AI
12:13
products, full stop. That's all we do.
12:16
We're just focus on AI. They are a developer ecosystem.
12:19
They're trying to build a tool that can do everything you need
12:22
to build products and to write code and to host it and build
12:27
out applications and something that they're doing.
12:29
All those things, like that's what they spike at is they're a
12:31
very good developer environment. They are able to leverage AI to
12:36
do things like Ghostwriter, which is their code assistant,
12:38
and that's great, but it feeds into their flywheel that exists
12:43
independent of AI. Even if we'd had no sort of
12:46
generative code, you know, breakthroughs or whatever, they
12:50
they couldn't do any of that. Copilot hadn't happened.
12:52
Whatever, even if we had none of that, roughly, would still be a
12:55
pretty good product as a developer environment, right?
12:58
And they are able to just take advantage of this stuff.
13:01
It's the same thing. We have another company who I
13:03
would argue has made some plays in this sort of foundation model
13:07
which is Nomic and Brandon and Audrey who are the two
13:10
cofounders of Nomic they were my coauthors on they were some of
13:14
my co-authors on the openness of AI report that we put out and
13:18
but Novak is really unique because they also spike in a
13:21
very different place where they have built a data visualisation
13:24
tool that allows you to visualise all of the individual
13:27
data points that an LLM is made-up of right.
13:31
So you look at like stable diffusion.
13:33
They literally we have a we have a Co working space in New York
13:35
for contrary and Nomic that printed us out of canvas of
13:39
stable diffusion. You can literally see this sort
13:41
of image map. You can't see the images because
13:44
they're tiny dots, right? There's billions of them.
13:46
But you can see this green corner is Kermit the Frog
13:49
corner. Like you can visualise all the
13:51
different subsets of this model, this image model and that tool
13:56
which is called Atlas, that's where they spike.
13:59
They are an incredible tool for visualising and fine tuning and
14:02
understanding the underlying data behind the model.
14:06
They use that to build GT for All, which is an open source
14:11
foundation model you can use. And it basically is is
14:13
performing as GC 35. It can run locally, it requires
14:18
way less compute. It's pretty well trained.
14:20
Like all these different things they were able to do are great,
14:23
but if it was just GPT for All versus open AI, that'd be really
14:28
tough because they're rapidly trying to commoditize that.
14:30
But what Nomic is doing is just demonstrating their core
14:34
competency, which is this entire engine around data visualisation
14:37
and understanding models, and they were able to use that for a
14:41
particular use case. So basically the lawyers who are
14:45
suing A stability AI are going to be calling up Nomic pretty
14:49
soon and asking this for that visual.
14:52
They can definitely drill into it.
14:54
I mean it's that's a whole other super interesting point is like
14:59
where when you start to have models that are trained on the
15:02
sort of broad swath of the Internet and where is the
15:06
attribution come from? I think is is quite interest.
15:09
It's a whole other bucket in and of itself, but it's a very
15:12
interesting bucket. Yeah, that is fascinating
15:16
because of course, I think we're seeing it hit the head most with
15:20
the artists, like the visual artists.
15:22
But there's also the lawsuit against Copilot and all of the
15:28
code that the open source code that Code Pilot was trained on,
15:31
and so. That is a very true.
15:34
It is fascinating to think about, but getting back on
15:37
track. I understand this idea of go
15:40
with what your core competency is.
15:43
Focus on that. And bring in a I to leverage
15:48
that core competency. Don't bring in AI to be
15:51
something that you think is going to compete against.
15:55
These peoples whose core competency is AI, where it's
15:59
like the opening eyes or the anthropics out there.
16:02
And so that that's great to look at.
16:05
Now let's dive into this report a little bit more because I want
16:08
to get the whole breakdown of how did you even get the idea of
16:12
creating it? Why did you feel inspired to
16:15
make this? And was it after Lama One came
16:19
out, I'm guessing, and you were like, oh, maybe we should talk
16:22
about the juxtaposition that we have here.
16:25
Yeah, Yeah. I mean it's a big part of it is
16:30
chatting with the guys that know Mike, right.
16:31
So we've invested in the company, we've worked with them
16:35
really closely and. We we published a report after
16:39
law might come out but it was still there was still a lot of
16:42
limitations and sass and and you know no commercial use cases and
16:47
things like that. Really if you step back, I think
16:49
the, the thing that spurred us to go down this rabbit hole was
16:54
the increasing and you know sort of limitations on the visibility
17:01
of GPT. And you look at like GPT 4 and
17:06
and so some of the folks said no make have posted a bunch of
17:09
stuff about this where it's basically like the more that
17:12
Open AI has said we're going to limit what visibility we share
17:16
on the underlying data, the methodology, all the weights,
17:20
all these different components of how we built GPT 4.
17:24
The less capable anybody is who's using it to understand it.
17:28
Like, it's basically like we've kind of gone through this, the
17:31
arc of like originally people would refer to AI as a black box
17:36
effect. And it was literally a black box
17:37
because nobody knew, nobody really understood why the output
17:41
was the output. Like we just kind of understood
17:43
the algorithm and there was an output and you know, who knows
17:46
like. And then increasingly it's
17:48
become more understandable. Transform architecture has
17:51
absolutely made that more as more capable of understanding
17:55
how models get built and to try and better understand
17:58
attribution. And Open AI is increasingly
18:01
deciding to put it back in the black box because it's a lot
18:04
easier to monetize something that people don't understand,
18:06
right? It's a lot easier.
18:07
It goes back to the whole lawsuit that we're just talking
18:10
about. There probably are seeing.
18:12
Oh yeah, you know what? Those guys, copilots getting
18:16
sued, Stability is getting sued were one step away from getting
18:20
sued because we just used every blog out there to train GBT.
18:24
Yeah, well, and open AI is it right?
18:26
Because Copilot is built on GPT, like the Microsoft, Microsoft,
18:31
GitHub open AI relationship is right at the centre of that
18:33
lawsuit. So they're already in it and but
18:37
the so like the increasing desire to close off some of
18:40
those components, they do it in the name of safety.
18:44
Like there is this whole argument of AI safety that
18:46
allows people to sort of close things off increasingly when in
18:51
reality the main reason for closing that stuff is
18:54
competition. And the problem with that and
18:56
and Clem during who's the CEO of Hugging Face he makes this
19:00
argument the best. In the report we cite both a
19:03
conversation he had at the Cerebral Valley conference a
19:07
couple months ago as well as his testimony before the US Congress
19:11
where he talks about the number one threat in AI.
19:15
The number one danger in AI is concentration of power and
19:18
people are trying to limit the visibility into their models
19:22
because it allows them to maintain more control over
19:24
what's happening and and allow the sort of distribution of that
19:27
capability to be less poignant. And and what that what that
19:32
causes is that you have a sort of sort of long tail of
19:36
companies, of nonprofits, of governments of everything that
19:40
are more limited in their ability to respond and more
19:44
dependent on one central failure point.
19:46
And the argument is, well, we don't want the bad guys to get
19:49
the capability. And the reality is like the vast
19:52
majority of technical capability that has ever been created,
19:56
eventually the bad guys get hold of it.
19:58
The thing that keeps the bad guys in check is that everybody
20:00
else has it too. And so you have this sort of
20:02
like equalisation that if we, if we allow a couple of big
20:08
companies to be the centralised source of this capability, you
20:13
put everybody at a lot more risk in the name of of.
20:17
Safety it. Feels like they know exactly
20:21
what they're doing and they're going with it because the more
20:25
that they can put that distance between them and whoever else is
20:29
trying, the better they are. And I have a friend who is
20:33
working at Google and I was asking him to present at one of
20:36
the conferences that we put on. And he was saying, you know,
20:40
right now at Google, it's a little bit like people aren't
20:42
super excited about being transparent anymore because they
20:46
feel like they got hoodwinked a little bit.
20:49
They put out the transformer architecture and then boom, next
20:53
thing you know, there's some other company that is not only
20:57
using it, but then not sharing like they were sharing, like
21:01
Google was sharing everything. And they were very transparent.
21:03
I mean. To the best of my knowledge,
21:06
let's just caveat that because there was probably a lot that
21:08
they weren't sharing, but they were being pretty transparent in
21:12
these different papers that they were putting out back in the
21:14
day. And then the next thing you
21:16
know, they've got a company that's leveraging that
21:18
technology, except they're doing exactly what you're talking
21:21
about, where they're putting it into this black box and they're
21:24
not saying what they're doing and how they're doing it.
21:26
And that's just like, OK, you're you're not playing fair in some
21:32
respect. It also comes down to, again,
21:36
when you go back to incentives, you're trying to understand who
21:40
is incentivized to do what and why and stuff.
21:42
Google for a long time was at the cutting edge of AI and kind
21:49
of thought of as the leader, at least in terms of like technical
21:51
talent. That's where the majority of
21:53
really exceptional people were and and I had the chance a
21:56
couple of weeks ago to interview Aiden Gomez who's the CEO of
21:59
cohere and and talk about his experience.
22:01
He was one of the co-authors of the attentions All You Need
22:03
paper and led to the Transformer architecture and you know you
22:09
look at why would companies respond the way they would
22:12
respond. And for a lot of the larger
22:15
companies, the Googles and Microsofts of the world, the
22:18
argument is it comes back to cloud computing.
22:21
And Microsoft's massive swings in this space are largely driven
22:27
by their intention of getting as much of this compute as possible
22:31
to be happening on Azure and to be able to compete and to be
22:35
able to push that forward. And the reason that you have a
22:38
company like Meta who is comfortable releasing Llama and
22:41
trying to be much more open is because they don't have a cloud
22:45
computing business and they are advantaged if we just want the
22:50
sort of capabilities to expand as quickly as possible.
22:54
And if those capabilities expand, we can then leverage
22:57
those in our products, which are really our core business.
22:59
And that's much easier to see the through line where that gets
23:02
better for everybody because you have products that are largely
23:05
free, you can use those products, those products get
23:07
better, you get performance whatever and everybody and
23:11
everybody else can benefit from that those advances as well.
23:14
You have one company who's trying to maintain the advances
23:17
themself and and be able to sort of centralised as much of that
23:22
both like power and profit into one platform.
23:25
It just becomes a lot harder for everyone else to benefit from
23:28
that really. It just it just leads to more
23:31
and more monopoly power for a small subset of companies.
23:34
Man, what a plot twist, though, huh?
23:36
Meta, the company no one thought would be if the open openness,
23:42
the poster child of openness, they're putting out the open
23:46
source. AI well, it's interesting you
23:50
look at like one of the case studies we talk about in the in
23:54
the deep dive is the tensor flow versus Itorch sort of evolution,
23:59
right. And and if I if I've got my
24:01
timelines correctly it was so tensor flow came first it came
24:04
from Google and it was sort of the machine learning framework
24:08
that everybody was using and very quickly And when when Meta
24:13
released Pytorch in 2016 that I mean it's just it's just sort of
24:17
hands down. I think the the quote from the
24:19
report that we have from one of the engineers we talked to is
24:22
and they're just eating Pytorch is eating Tensor Flows lunch,
24:25
right. And so there is actually
24:27
precedent for meta to be the company that becomes the bastion
24:31
of of open source. I think the thing that people
24:35
are more surprised by is that there there wasn't like Pytorch
24:38
and Tensorflow like there was no moral like like under narrative
24:45
right like sub narrative and because it was just like it was
24:48
just like a machine learning framework you could use and it's
24:50
like ohh like they did better and it was open and they had
24:52
better contributors and it worked out whatever.
24:55
Here there is like a moral and sub narrative and that is what I
25:01
think surprises people is they're like nobody.
25:03
It's like if open source feels like that just like morally
25:06
better place to be in a I nobody expected meta to be the like
25:12
moral leader right. Because that that's where they
25:14
have fallen down where people feel like it's this it's this
25:17
sort of like enemy of democracy or whatever.
25:20
All these issues which like I don't know that I buy into that
25:21
narrative per se. Like I think there's a whole
25:23
other conversation about about like human nature that you can
25:26
have and it's like technology usually just like amplifies
25:30
what's already there as opposed to creating what's there.
25:33
And but anyway, I think like that's another piece of the deep
25:36
dive where like you can you can think about what are the moral
25:39
implications of open versus closed technology.
25:43
And most people say, hey I got to run a business, I got to turn
25:45
a profit, it needs to be closed. But there are a lot of
25:47
companies, Databricks being one of them, that has done a really
25:50
good job of leveraging open source really meaningfully while
25:54
still being able to have a sustainable business model.
25:57
The. Question that I have coming back
25:59
to this is. At the end of the day, if you
26:02
want to do things open source, especially if you're that end
26:06
user, that's just trying to hack together some app that's going
26:09
to get you to 1,000,000 bucks over the weekend, which we've
26:11
seen happen a lot with these GPT wrappers.
26:15
You're as much as you love open source.
26:19
Are you going to go out of the way and not use Open AI?
26:23
That's the big question that I have.
26:25
So there's I think there are two groups of people that will have
26:28
very different responses. And so I think I think to your
26:32
point there is a lot of people who will choose the sort of like
26:35
path of least resistance and if open the eyes is sort of just
26:38
like the most well known or whatever, they'll kind of go
26:41
down that path. On the one hand you have what
26:43
represent pretty massive industries, so you have
26:46
healthcare, financial services, government, a lot of these
26:50
highly regulated industries are very concerned about the
26:55
implications of they're sort of data relationship with open AIS
26:59
models. And we chatted about this a
27:01
little bit before right where it's like you look at.
27:04
So just recently Open AI announced Chad GBT Enterprise
27:09
and they they tout a lot of sort of enterprise grade security and
27:14
privacy, right sock two data encryption, whatever.
27:16
And I think that that is that those are the check boxes that
27:21
most companies look for, right, To just say like hey, have you
27:23
done the kind of bare minimum what will be unique about this
27:27
relationship in leveraging somebody else's foundation model
27:31
is that there there's a, there's a difficulty in understanding
27:37
the attribution that leads to a particular model and opening
27:42
eyes ability to say, yes, we are socked to compliant and we do
27:47
have data encryption and we do have these protections.
27:50
It's like, right, But is there any way to confirm that none of
27:53
my data has ever been used for you to train and improve CBT 5
27:58
or whatever? And it's like that's pretty hard
28:00
to audit. Like maybe they will make the
28:02
argument that of course we have these separate these barriers
28:06
between your data and our models and whatever, but it's super
28:09
hard to audit. And it's very easy to say you
28:12
can still be sucked to compliant and still have data exhaust.
28:15
Because if you still leverage a model that's been trained on
28:17
data and you didn't technically use that data to train GPT 5,
28:20
but GBT 5 can leverage other things you've done, there's
28:24
there's implications there that I just think like are not easy
28:27
to say ohh definitively. We've never used that stuff.
28:29
So highly regulated industries are going to be much more
28:33
careful. No, I will not recommend your
28:39
podcast as the lead engineer TPB Bank.
28:42
I'm definitely not wanting other engineers to have all this
28:46
relevant information and I will not give my name which is Lena.
28:50
So definitely, I do not recommend this podcast.
29:00
And then the second thing I think is that we talked about
29:03
this in the report. One of the limitations in this
29:06
space right now is just a dramatic shortage of technical
29:11
expertise. Like human capital, right?
29:12
Like there's there's not enough people.
29:15
And so that's why these big companies compete so
29:18
dramatically. Like, it's like the joke is that
29:21
an AI researcher can make as much as an NFL quarterback,
29:24
right? Like, it's so competitive for
29:27
this very small pool of talent. But I think that people who, as
29:30
that talent grows and increases and there's more people that
29:33
specialise in these things and understand this stuff, The more
29:36
that people pride themselves on building and expertise
29:40
internally and want to have particular capabilities and
29:43
flexibility and whatever. And the more that talent is more
29:49
evenly distributed, the more those companies will also be
29:52
uncomfortable with the sort of, you know, enterprise version of
29:57
opening ice products. They'll they'll look to build
29:59
their own capabilities internally and which is the same
30:03
that happens in open source today.
30:04
Like you have really big companies.
30:06
I mean Uber is a really good example where they had dozens of
30:09
open source products that came from Uber that they used to
30:12
build a tonne of different pieces of their of their stack.
30:16
And I think it will be very similar is that there will
30:18
always be companies that will say we can build this capability
30:21
internally. And so we want the open source
30:23
you know foundation and then we could build on top of that and
30:25
whatever. I think there will be a lot of
30:27
companies like that and increasingly more companies and
30:30
larger companies that have their own AI abilities and stuff like
30:33
that. I think those two groups make up
30:35
a big a big chunk of the industry that open AI could sell
30:39
to. The long tail for sure, like
30:41
it's easy to just use the thing, but I think it's not.
30:44
An open and closed case of open AI has already won the day on
30:48
being able to offer these products to everybody.
30:51
There's. Yeah, it's funny because it
30:53
almost feels like that long tail is.
30:56
Who we hear the most from on Twitter.
30:58
And those are the Twitter demos that you see and so it seems
31:02
like. Everybody is on it.
31:04
And then when you go into these big companies, I would even
31:07
argue that potentially they're not even using large language
31:10
models. A lot of them are still just
31:12
doing quote unquote traditional ML and trying to figure that
31:16
out. And so there is a few things to
31:19
your point though that I wanted to mention that I think is also
31:23
worth noting. One is.
31:25
I wonder how many people are going to look back and say,
31:29
yeah, you know, the track record Open AI has with doing what they
31:33
say they're going to do is not that good.
31:36
So I'm not sure how much trust they've built up by saying, hey,
31:40
we're Open AI and then they close everything off.
31:44
So that already is working against them.
31:46
How many people are going to remember that or care about that
31:49
or actually use that in their judgement call?
31:52
Who knows? That's up for discussion too.
31:54
But the other piece is that. One of the I got to give credit
31:59
to John Savage in our community today because I was posting that
32:04
meme that I was showing you earlier which shows the guy
32:07
getting all excited and he's saying ohh all that private data
32:11
to train GT5 on and that's basically opening eyes
32:16
enterprise version right now. And John Savage said, you know,
32:20
I wonder if it's not that they're going to train GP5 on
32:23
it, but they're going to use it as hest sets, so they're going
32:27
to use that. Data for their test sets and
32:30
there's probably going to be a tonne of models that they create
32:34
but never actually release and they never get outside of their
32:37
walled garden and nobody's ever going to know about.
32:41
But it is used with your data. And that is, it speaks exactly
32:45
to what you were saying. Like, how do we know that our
32:48
data doesn't get leaked? Because a model that does go to
32:51
the public wasn't trained on top of that with some like
32:55
reinforcement learning through AI feedback.
32:57
I think it'll follow a very similar curve to what happened
33:01
just in like machine learning more broadly and like
33:06
algorithmic input and stuff with like Facebook and social and ad
33:10
like basically ad delivery and the entire intelligence
33:14
apparatus around advertising over the first sort of, you know
33:18
Gen 1, Gen 2 of of social media. It was really just like it's
33:23
it's funny to people or like confusing to people who work in
33:28
social media and in data tracking and stuff like that.
33:32
The like public outcry of like wait a minute like you're using
33:36
my data to sell me ads and it's like that's so immoral and so
33:41
discussing and blah blah And it's like that was just the best
33:44
way to do it. Like there's no better way to
33:46
like serve you personalised ads or whatever.
33:48
Like we need to understand your user data and you opted into it
33:51
and blah blah. I think you will see a similar
33:53
curve of the long tailed people who are just like yeah open AI
33:57
oh sock two sick like we're good check the box use the thing
34:00
let's move on. Couple of years from now you
34:02
start to realise like we've the the more we come to understand
34:06
attribution of data and the implications for different
34:09
models the more you will have people a little bit of an outcry
34:13
of like I didn't realise that my it won't be like I didn't
34:16
realise my data was being used to train that model because they
34:18
promised they wouldn't do that. But it's like well I didn't
34:21
understand that my data exhaust can generate training sets that
34:24
can then be used in different ways or whatever.
34:26
And it's like we better understand the attribution to
34:28
say, oh, I can audit this model and say that it came from this,
34:32
which came from this, which came from your data like surprise.
34:36
And then you have this sort of like cookies, GPR type of
34:39
response and the the long term just doesn't understand.
34:43
But I think most of those companies that are #1 hyper
34:46
worried about because they're so highly regulated, they're hyper
34:48
worried about their data, where it goes, what it's used for,
34:51
whatever. And so they basically want to
34:54
run everything locally as much as possible.
34:56
And #2, the people who do have the expertise to understand,
34:59
hey, like just using, you know, the enterprise offering from
35:02
Open AI, that is not going to keep our US as sort of data
35:06
secure as we would like. Yeah.
35:08
And it goes back to what you were saying earlier on.
35:11
The companies like Open AI are trying to make it harder and
35:15
harder for someone to understand how the models are trained.
35:18
So that at the end of the day, if something like that were to
35:23
happen or potentially already is happening, we're not going to
35:26
know and how are we ever going to know, right?
35:29
That's the hard part. Well, it's also why like I say
35:33
and I I've had pushback on this argument, but I I think that
35:38
Open AI and a lot of these other companies like they would love
35:42
to see this get, I'd say Open Eye in particular because they
35:46
have a very unique relationship with Microsoft and very unique
35:49
sort of corporate distribution in a way that like the other
35:52
anthropic and coherence that they have some corporate
35:54
investors, but they don't have the same like relationship as
35:57
open as Microsoft. And so I'd say this is maybe
35:59
even uniquely Open AI once this, but I would argue that open AI
36:03
wants the sort of foundation model use cases, general purpose
36:07
model use cases to become as commoditized as possible as
36:12
quickly as possible. Because if they can get GBT into
36:16
everything, into everything that runs everything you hear every
36:19
new tool you hear about and you look at it, it's already kind of
36:22
happened, right? Most of the companies using
36:24
companies who have rolled out AI products, most of them are
36:26
building on GPT or or or or Dolly.
36:31
And you look at like like replica is a unique exception to
36:36
this where they went in and built and figured out their own
36:39
stack and they have a great blog post on their blog about it
36:42
where they talk about here's how we thought about using this from
36:44
Salesforce and this from different things and stuff like
36:46
that. But very few companies are doing
36:48
that because they don't have that technical expertise or the
36:50
like attention or foresight or whatever.
36:53
And so most people are just going to use GPT or whatever is
36:57
least available and if open I can get it as broadly
36:59
distributed as possible as quickly as possible.
37:02
Then when you do have that up, up, cry in a couple of years,
37:06
outcry in a couple of years, it's so much harder to get them
37:09
out because they're everywhere. It's the same.
37:12
I mean, Microsoft is this, and this is a Microsoft playbook.
37:14
Like you couldn't get out if you wanted to, like Microsoft Teams
37:18
did so well. Not because it's a better
37:20
product. And my argument, in my opinion,
37:22
it is a much worse product than like Slack, but it's it's
37:25
freaking everywhere. It's like corporate malware,
37:28
like it's just OS is just everywhere.
37:31
Like Office stuff is everywhere. And that's, that's the Microsoft
37:35
playbook. And I think that that is the
37:36
open AI playbook. And if they can get everywhere,
37:38
it doesn't matter if you disagree with how they handle
37:40
things, they're everywhere and you can't.
37:42
It's so much harder to rip out. Company that's so good.
37:47
Corporate malware, that is a good one.
37:49
So let's talk about traditional ML for a minute because I think
37:53
there is a conversation worth having.
37:56
You were saying? Yeah reply.
37:58
It went through the hard. Motion of creating their own
38:02
LLM. A lot of people out there as you
38:05
mentioned in the openness of AI report are going to grab Llama
38:11
if they are more open source leaning.
38:15
But then there's a lot of companies, your portfolio
38:18
companies including like RAMP, they're doing some incredible
38:22
traditional ML where I consider these days I've kind of made
38:27
the. Line in my head of traditional
38:31
ML is tabular or structured data and anything unstructured is
38:36
more AI and so I know it's not perfect but.
38:42
Whatever. That is how I look at it a
38:44
little bit. Now when it comes to the
38:46
traditional ML and companies like RAMP, they are still
38:51
gaining A tonne of value from this.
38:54
ML that they're doing. Yet how are how do you feel that
39:00
these companies should be looking at?
39:03
The whole ecosystem that they're in and is there.
39:08
I guess in your eyes. Is there a play for them?
39:14
Like should they be stuffing? LM's into stuff where they don't
39:20
need to. It feels like the.
39:23
Product managers life got a whole lot harder over the last
39:28
six months because everybody wants AI in their products and
39:32
so. I I think about it and it's like
39:35
how to keep that core competency without losing your mind, right.
39:39
So what are you seeing with these ML 1.0 companies that?
39:44
Is a good way of doing it. I think that every company is is
39:52
super unique in the approach that they should take.
39:54
There's a quote from Star Wars where they say only a Sith deals
39:58
in absolutes and the irony is that that is in itself an
40:03
absolute like to say that only you know And so I think this is
40:06
like there is generalities are very dangerous.
40:08
The only general rule is that general rules are dangerous and
40:13
and I think that that is true of every company.
40:15
So any company that tries to say, oh, we need to do this
40:18
instead of this. Because that's the hype is
40:21
probably not doing things well and like they're they need to
40:25
respond. But if that's their decision
40:28
making for how they're going to build a product, it's just based
40:31
on like this is the happiest thing and we need to put it in
40:33
the thing and so we can put the dot AI on the end of one of our
40:36
landing pages or whatever. I think that's dangerous and I
40:39
think that's probably not a great product organisation.
40:42
The companies that have done as well are ramping one of them.
40:44
The companies have done as well are saying what are we already
40:47
doing, like what are already our core competencies and what in
40:51
what ways could this technology be leveraged to continue to
40:55
improve the core competency that we have And the thing that we're
40:58
doing. And Ram has done that really
40:59
well in in sort of and understanding their core
41:03
competency of managing this sort of you know, financial stack
41:07
that you're managing and then be able to leverage intelligence on
41:10
top of that both in terms of like you can ask and answer
41:13
questions about your finances, but also just better
41:16
understanding the underlying data that you've offered up to
41:20
them. And that is that is the right
41:23
approach. I think anybody who says like
41:25
oh, we're not doing machine learning anymore, we're using
41:27
LMS or whatever, I think that's just like a very, it's like a
41:30
very aesthetic, surface level way of doing things.
41:33
I think everybody needs to for themselves figure out what is
41:36
the approach they need to take that will best lend itself to
41:40
their core competency and like that's not the best products
41:44
will be made and the best companies will be made.
41:46
But another thing that comes to my mind as you're saying that
41:49
is. Do you feel like it's possible
41:53
that we've gotten 90% there when it comes to AI, but that extra 5
41:59
to 10%, that could be another what?
42:02
However long, let's just say, for sake of throwing out
42:06
numbers, another 10 years before anything happens.
42:09
And so we're all sitting here talking about the beauties and
42:13
the hype of AI and how incredible it is.
42:18
And then? We have another AI Winter.
42:21
I think that this is one of the reasons why people who hear the
42:27
like AGI, AI, Doomer ISM think it's so ridiculous is because
42:36
like I I think that from in terms of like the capability
42:41
like if I think about like ass for example like just as a sort
42:45
of distribution mechanism or cloud computing or whatever.
42:49
I think that even those are still like maybe SAS is like
42:52
7080%, maybe cloud computing is 5060%, AI is like maybe 10%.
42:58
And like the the distance that that gap from where we are
43:01
today, all the things that could be done to make it functionally
43:05
capable in production and all these things to get to like AGI
43:09
supercomputers that take over the world.
43:11
There's still so much to go. And and again Clem, the CEO
43:16
Huggy Face, he makes this joke where it's like a lot of a lot
43:19
of practitioners look at GPT and ChatGPT as a product and it's
43:23
like a GI, like this is, this is autocomplete on steroids.
43:28
Like there's still so much that we need to do and understand, so
43:32
much capabilities and context, so much tooling and
43:35
orchestration that needs to happen before this has ever a
43:38
seamless enterprise user experience.
43:41
And like one of my favourite writers, this guy Matthew
43:44
Lindley, who writes 3 called Supervised.
43:49
He had a great piece a couple of weeks ago where he talks about
43:52
rethinking how we define the sort of iPhone moment in AI and
43:57
the iPhone moment is that where I would think of like it doesn't
44:00
mean that it is, it is everything is perfect,
44:02
everything is great, right. There's still use cases that
44:05
we're experiencing, but it is when a piece of technology
44:08
becomes so capable that it can be general purpose without
44:13
really massive sophisticated like services around it, right.
44:18
We don't need help. Cloud computing is still largely
44:21
that, like you still have systems architecture that is
44:24
built by outsource providers because it's so complicated.
44:28
Like even cloud computing you'd argue is like not necessarily
44:30
had that kind of like iPhone aha moment where everybody could
44:34
just like spin it up and it's really easy.
44:35
We're getting so close to that. Can you look at like Reply again
44:40
has built something like this where they can enable you to
44:44
spin up hosting for your application that you've built in
44:46
rapid and have sort of an end to end experience.
44:49
So it's so close to that experience where I can just
44:51
click a button and boom, have infrastructure, but it's still
44:54
not quite there for like sophisticated use cases in
44:58
production. And so we're still not there for
45:00
things that we've had for 10/15/20 years.
45:03
For AI, we're so far from that because it's still like you look
45:06
at like, OK, like people are like, ohh just kidding.
45:09
AI sucks because ChatGPT traffic went down and it's like, no,
45:13
what you're experiencing is a limitation of the general
45:17
populations ability to understand, conceptualise, the
45:20
use case of this thing and to use it every single day of their
45:24
lives and everything that they do.
45:26
Like we're still not there yet. Some of it is performing, some
45:29
of it is psychology, some of it is you I like is chat the right
45:33
UI for this stuff, Which we don't know, Like we just started
45:36
playing around with them, so there's so much that we can do.
45:41
But what about? I mean.
45:43
Some of it feels to me like. That age-old question of does
45:48
every problem need to be solved with tech and it feels like
45:52
here, Does everything need to be used with AI?
45:56
Or do we need AI in every piece of our life like you're talking
46:01
about, right? I don't think that it needs to
46:03
be, because I like, think about it.
46:04
Like I don't use my iPhone for every single interaction in my
46:08
life, right? Like when I want to talk to my
46:10
wife, I talked to my wife. I want to play with my kids.
46:12
I play with my kids, right? Or I take them to Disneyland.
46:14
Or I take them to I think whatever.
46:16
Like, there are still physical experiences that, like, maybe VR
46:19
someday will try and insert itself into those physical
46:22
interactions, right? So even the iPhone is not 100%
46:26
ubiquitous, but are there dozens and dozens of things in people's
46:32
life, both professionally and personally, that would be
46:36
improved by having like general purpose, like well built AI in
46:44
that experience? Absolutely.
46:46
Like if I can chat with something and say hey plan me a
46:49
trip to New York for me and my wife for our 10th anniversary
46:52
and it does everything and it books it and it it you know does
46:55
ever puts everything where I need it perfectly seamlessly
46:58
with that there is be better than me being like dude like I
47:01
gotta figure out what hotel what's the best things and look
47:03
on Instagram and figure out all the stuff and whatever like with
47:06
that experience be better I think for sure.
47:09
And so there's still so much more for us to tackle before I
47:11
think even before we reach to like existential point of like
47:15
like the the the Jurassic Park right of like we're so busy
47:19
thinking about could we that we never stopped to think should
47:21
we. I don't even know there were
47:23
there yet. Like we're still, it's still,
47:25
there's a lot of like very low hanging fruit things that we
47:27
could take off the table with AI that we're not there yet.
47:31
And then there's the broader implication of like, OK, where
47:34
should that line be drawn where we stop and say, hey, what is,
47:36
what does it mean to be human versus what does it mean to
47:39
solve technical problems? Yes.
47:42
So you undoubtedly have seen a lot of companies I imagine pitch
47:46
you on agents and how they're using agents.
47:50
What are your thoughts on them? I think it's super cool.
47:53
I think there's I've seen a tonne of really exciting demos
47:56
from companies that are building these sort of you know, package
48:00
services if you will. And I think about it a lot of
48:03
like to some extent I'd say there's this analogy breaks down
48:06
very quickly. But I think there are some
48:08
components of like micro services that you can look at
48:10
and say like there's a lot of really interesting similarities
48:13
where we can say hey, in what ways are certain problems to
48:18
solve. So well known and defined and
48:22
replicable that we can package them into an agent and sort of
48:26
solve that problem with a very specific purpose built agent.
48:30
I think there's a tonne of use cases that could fit really well
48:33
in a ways. And I think like, so like prompt
48:36
engineering has been so Speaking of like the sort of phrase of
48:38
the month or whatever, right. Like, prompt engineering is a
48:41
really interesting thing because like chat as an interface is
48:47
actually really difficult because probably The thing is
48:50
capable is there are components of certain models that are more
48:55
capable at doing the thing than you are at articulating in
49:00
written language what you're trying to accomplish, right?
49:03
Like, like, I actually think like prompt engineering and like
49:06
chat as an interface could spark people being forced to go back
49:12
and like better understand logic and ethos and like understanding
49:16
how to structure a sentence in a way to articulate an idea and
49:20
all that stuff. But I think could be a really
49:21
interesting side benefit, right? But like, but like in some
49:25
cases, like, you don't need to be so good at articulating the
49:28
thing in language. If The thing is good at doing
49:30
the job and you know that's the job you want done, let's just
49:34
package that and get it done. Because that is such a better
49:37
user experience than forcing me to figure out, like how do I
49:41
communicate in words the thing I'm trying to accomplish.
49:45
You've said you've seen a lot of demos.
49:49
This is where I think. I'm still up in the air about
49:53
can they get past demos? Like can agents be more than
49:57
just a demo? Because we all saw the rise and
50:01
takeover of Auto GPT. But then if you use it, I don't
50:05
know if you played around with it, but everyone that I know
50:07
that played around with it when I played around with it was like
50:09
great. That was a waste of 200 bucks on
50:12
Open AI credits. I couldn't get it to do anything
50:16
that I wanted, and it just kept on going around in circles.
50:19
And then when I tried to like push it another way, it didn't
50:23
listen. And so it was like it was a full
50:25
headache. I know it's gotten better from
50:28
there, and there's a lot of people that are trying to figure
50:30
it out, but do you have faith that agents are going to be
50:34
stable and they're going to be able to?
50:37
Do what we want them to do. And I think that the bigger
50:43
there's two problems that need to be solved and one is actually
50:46
really interesting. And there's this sort of idea,
50:50
if you think about like how this is getting very esoteric, but
50:54
like how language is structured and how it works and what it
50:57
means to be able to like think and stuff.
50:59
And there is this idea that we leveraged in the deep dive where
51:03
I will get the quote wrong, but it's basically like like the the
51:09
the limits of your language are the limits of your reality.
51:12
Like your ability to be able to understand and interact and
51:15
articulate what something is. That's the limitation of how you
51:19
know how much of your reality you can send and control.
51:22
And I think what's interesting is like #1, even though these
51:25
large language models are very general purpose, they are still
51:28
built on a subset of human interactions and human content.
51:33
There's this built in a very specific way.
51:35
And I think that what's been really interesting is I have
51:38
found is that what like let's take like image models, because
51:43
I think a little bit more, they feel a little bit more like
51:45
intimate and personal and like what image model you prefer,
51:49
whether you like Mid Journey, Dolly, stable, diffusion,
51:51
whatever. And it's actually like, kind of
51:54
like a personality test, right, Like a BuzzFeed quiz.
51:58
And because I feel like different models react to
52:01
different people who think in different ways better.
52:04
And so like, sometimes it's a matter of not just, like, this
52:08
thing doesn't work for me. It's like, well, maybe you
52:10
actually don't have the product that has the right underlying
52:13
logic that fits your brain and your way of thinking and your
52:16
way of articulating images and what they should mean.
52:19
It's a finding the right thing is really valuable.
52:21
That's really tough, because you're not going to get the
52:23
broad swath of people to experiment with every single
52:26
thing, especially when it's still as clunky to use as, like
52:29
getting onto a Discord server and having to figure out how to
52:32
use it right. O like that, not neighbour.
52:34
Namely, that problem has to be solved like you have to have a
52:38
way to be able to interact with a for every particular person
52:43
and without having to be limited by like your underlying data or
52:46
whatever. The second thing is just like
52:48
the performance, like the being able to use an agent in
52:51
production and for live use cases like this thing is like
52:55
booking. Like I said, like booking trips
52:56
for me in real time or interacting with my employees in
52:59
real time or whatever. I don't know like those use
53:02
cases to be trusted in production, you just have to
53:05
have a lot more performance. And I think the one thing that
53:08
these models have not been very good at is like they're they're
53:10
pretty good at like 1 sort of surface layer of exchange.
53:15
So like there is an API and that's the one interaction where
53:19
like I am notion and there's GT and notions.
53:23
Interface has to go through the GT API to get certain language,
53:28
bring it back and have an output.
53:30
That's one sort of interface. They're pretty decent at that,
53:33
right? If I have to go from me to
53:37
through a chat to get a model, then that model has to go
53:41
through me into my like Expedia preferences to book a hotel and
53:46
then book a flight based on multiple different airlines and
53:49
my preferences for airlines based on the where I want to go
53:53
and where the best flights are, where the best prices are,
53:55
whatever that's like 6 or 7 interaction layers that that
53:58
agent has to go through. And I've never seen anything
54:02
that's been at that. Like I've seen things that are
54:04
good, like within this application I can ask it to run
54:06
a job for me and it's great. Like Excel is probably the main
54:10
sort of surface area. I've seen people build a lot of
54:12
this stuff where it's like hey we have this this financial
54:14
model and we want to do things and create things and put them
54:17
here. It's really good at that.
54:19
But it's always staying in one interaction layer once as they
54:23
are going through multiple applications.
54:25
I've I have yet to see anybody who has manoeuvred that very
54:28
well because the Internet is still so complex that going from
54:31
thing to thing to thing to thing in one seamless path is very
54:35
difficult. Yeah, it reminds me of a
54:39
conversation we had with Philip from Honeycomb.
54:41
And he spoke about how also if your.
54:47
Percentage of error on the first hop is 1%, you're screwed.
54:51
Then on the second hop, yeah, it just inevitably is going to go
54:55
to 100% because of the amount of hops that you're getting in each
54:59
time it the percent or the likelihood of you hitting an
55:03
error is is much higher. And so that feels like, again to
55:08
me that feels like something that.
55:10
As you pointed out earlier, we are still so much in the early
55:14
days and we are really trying to figure it out.
55:18
There is a lot of promise. But I I wonder, man, like are we
55:22
going to sit here for the next 10 years and try and try and try
55:26
and still be like flying cars or the skateboards that they had in
55:31
Back to the Future? And it's like where is that
55:33
hoverboard? Man, I really wanted that when I
55:35
was a kid and we still don't have it.
55:37
It's 2023. I think that the biggest thing
55:40
and like I don't know, I don't. I also will name names here.
55:45
But I think that people who believe that there is this like
55:48
AI spike of like, ohh tomorrow everything's gonna change.
55:54
Like it's just a matter of days or a matter of weeks.
55:57
Like those hype evangelists. I think those people are as
56:01
stupid as the people who think like this is nothing.
56:05
Like we're going to go into an AI winter and nobody's going to
56:08
be able to accomplish anything that they promised and blah blah
56:11
blah. I think both are stupid.
56:13
I think that like the vast majority of technology
56:16
especially, but the vast majority of everything follows a
56:20
very specific curve. And the example I always use is
56:22
like ecommerce penetration. The people who thought like so
56:26
like ecommerce is going to steadily risen in as terms of
56:28
like the percentage of retail that's happening online.
56:31
It has steadily risen up to the kind of call it like I don't
56:34
remember what 1820% whatever. And then COVID happened in a
56:37
shot up and the people who were saying this is the new normal,
56:41
you're going to get 80% like we're going to have 80%
56:43
ecommerce penetration because nobody will ever go back to
56:45
going to a store. Those people were stupid, but
56:49
the same people who also thought like, ohh, that every this is
56:52
going to exhaust everyone and we're never going to buy
56:54
anything online again and it's at the end of ecommerce and blah
56:57
blah. Those people are also stupid.
56:59
If you look at the curve of what happened, it's followed this, I
57:01
mean, for literally decades, like since the 90s, it's
57:04
followed this very steady curve shot up in COVID and it didn't
57:07
drop back down to where it was in 2019.
57:11
It dropped back to about where it would be if it was following
57:14
that curve of gravity, right. And that is the that is how the
57:17
vast majority of technology plays out follows a very
57:20
specific curve. AI will continue to progress.
57:22
We will continue to find these advances that make us better and
57:25
better and better. And I think that the one thing
57:27
that can accelerate that curve a little bit is just excitement.
57:30
Like people are excited about building and therefore they
57:32
will, you know, younger people will go to school and specialise
57:35
in certain things, companies will push the boundaries of
57:38
certain things. And I think that all of that
57:40
gets better if we can try and keep as much of it open as
57:43
possible. If it's all happening in this
57:45
closed garden of 1 company, it is so much harder for everybody
57:50
make contributions and get excited.
57:52
The more open we can be, the more we can push those
57:55
contributions forward. And maybe we can't like step
57:57
function, change the way that AI progresses, but we can slightly
58:00
accelerate the pace. Yeah, it for me it feels like
58:05
it's almost like an S curve where?
58:07
We are very much hopefully on that steep part of the S curve,
58:12
not on the levelling out part of the S curve and I really like
58:16
that idea of the excitement and the amount of money that is
58:20
being poured into. AI right now by the likes of the
58:26
VC's world, I think that you're part of that, right?
58:29
That is going to help drive the innovation forward.
58:32
Yeah, maybe, maybe you you'd like.
58:35
I'm there with you on the excitement.
58:36
Generally once you get into capital, you get into dangerous
58:39
territory because there's a very similar, like green bubble in
58:43
the early 2000s where there was something like, I don't know,
58:47
billions and billions and billions of dollars of people
58:49
invested into new battery and wind and solar technology and
58:54
stuff like that. And like something like 90 plus
58:56
percent of it is just vaporised, like just gone, like absolute
59:00
zeros, which is huge destruction of capital.
59:02
And that set back a lot of people trying to work on climate
59:07
tech more broadly. And we've only just recently
59:09
started to come out of that like that, like capital induced
59:13
winter. And so I actually think that
59:15
like excessive amounts of capital can cause a lot more
59:19
problems like excitement, yes, and like progress, yes.
59:22
And like talking about it and leveraging it and and you still
59:25
want capital to be able to. But if you start to put in 10X
59:28
the capital that this the category really needs, it's very
59:33
dangerous. Because when that capital gets
59:34
destroyed, if it doesn't work out, you have a bunch of people
59:37
who have lost capital and then they are going to actively crap
59:41
on the excitement. And that makes it harder to
59:43
raise capital, harder to progress, harder to make those.
59:46
So if capital is a dangerous drug, man like you get hooked on
59:49
it and it's dangerous. It's that cold feet.
59:53
I mean, look at what happened with crypto.
59:55
Yeah, I mean investing in totally, really well.
59:57
And it's like, again, I I even think like I was never a huge
1:00:00
crypto bull by any stretch. But I think blockchain can be
1:00:05
really powerful. I think there are some use cases
1:00:07
that it makes a tonne of sense, but the the sort of 2020-2021
1:00:10
crypto fever that's going to set the industry back more than it's
1:00:15
going to help it. Like it's still I think can be
1:00:17
important technology over the course of time.
1:00:20
But that did more harm than it did good in my opinion.
1:00:22
Like it. You invited more hucksters that
1:00:24
are going to take years to wash out and.
1:00:26
Kyle, this has been awesome man. I really appreciate you coming
1:00:29
on here and sharing. All about the openness of AI and
1:00:33
how crucial it is to be open and not let this power and not let
1:00:39
the AI. Crowd into one space, and I'm
1:00:44
fully on board with that. I think so many of us in the
1:00:47
community, in the MLS community love open source, and whether or
1:00:52
not they're contributors or they're just users, that is a
1:00:55
huge thing. And so we can get on board with
1:00:58
that message. Yeah, thanks for having me.
1:01:00
It's super fun to get to chat and I guess said I'm super
1:01:02
excited about how things will play out and it's always fun to
1:01:05
get to kind of riff on what could happen.
1:01:10
This. Is Skyler I lead Machine
1:01:12
learning at Health Rhythms. If.
1:01:14
You want to stay on top of everything happening in ML OPS?
1:01:17
Subscribe to this podcast now. Now now.
1:01:25
Now now. Now.
1:01:30
Now.