Latest / Agentic Conversations (formally mlops.community) / From Rules to Reasoning Engines // George Mathew // #297

Transcript

From Rules to Reasoning Engines // George Mathew // #297 · Agentic Conversations (formally mlops.community)

0:00
Hello there. I'm George Matthew.
0:02
I'm a managing director at Inside Partners and I take my
0:07
coffee as a hot or cold been a latte with at least 1% milk.
0:18
Welcome to Round 2 of this conversation that I'm having
0:21
with George Matthew. This is the Envelopes Community
0:25
Podcast, and I'm your host, Demetrios.
0:28
In the 1st edition. When I spoke with George so many
0:31
years ago, we reminisced about what he did before he became one
0:38
of the most predominant VCs in the AI and ML industry.
0:42
And in this edition, we talk all about what he's been seeing on
0:47
the market and his theories for the next couple years.
0:51
Let's. Dive right in.
0:59
It's been 2 years since we chatted.
1:01
I remember when we chatted last, I was in Greece enjoying the sun
1:05
and I was totally oblivious to what was about to happen, which
1:09
I think we talked even before the ChatGPT moment ate up the
1:14
Internet. Now, what else has changed those
1:18
for those last two years? What have you been focusing on
1:21
and thinking about? Oh my God, like nothing happened
1:25
in the last two years, right? It was so quiet and such such
1:28
peaceful moments and everything related to data.
1:31
AIML Dimitri as well, wonderful to see you again.
1:34
It has been too long. Thank you again for having me
1:37
back on the ML OPS podcast. It's a pleasure what you've been
1:41
able to do with this community for the last several years.
1:44
Thanks, man. Coming back to a podcast a
1:46
second time around, particularly after, you know, having some
1:49
perspectives even before that, everything got crazy.
1:52
Yeah. We last talked, so yeah, I.
1:55
I remember distinctly that you talked about, I was asking how
2:00
do you look at companies and what are things that get you
2:02
excited in companies? And there was something you
2:05
said, and I'm paraphrasing here, but it was along the lines of
2:08
you're looking for companies that are going to really be able
2:13
to build generational organizations or?
2:17
Something generational outcomes. Yeah, I think it's a.
2:19
Generational. Outcomes.
2:20
Which was the first time that I had thought in such long terms.
2:25
And I was just like, wow, how do you even?
2:29
Wow? That blew my mind.
2:31
And so maybe we can talk a bit about generational outcomes and
2:34
what you think are the next wave of generational outcomes.
2:39
Sure. Yeah, happy to do that.
2:41
And I think it's a time and moment for me to have that
2:43
conversation because I am entering my fifth year as a
2:48
venture capitalist and certainly had had quite a long career as a
2:53
builder that has been able to scale up companies, particularly
2:57
in the data AIML space prior to of course being a venture
3:01
capitalist. That Insight for four years and
3:04
going on my 5th year, I mentioned that because Insight
3:07
just celebrated our 30th anniversary and we just raised
3:11
our 13th fund, which is a $12.5 billion fund 13.
3:16
And if you look at the arc of software for 30 years and
3:21
Insight has been very focused as investors in software from 1995
3:26
to now. I graduated College in 1995.
3:30
And so for me to, you know, almost think about what I did in
3:36
my career as a builder for as long as I was.
3:38
And now being AVC, it is a dramatic shift to go from the
3:44
moment in time where we start to see the introduction of client
3:47
server in 1995 and the advent of the web, the emergence of mobile
3:54
at scale, the cloud infrastructure that ended up
3:58
defining a good decade plus of software.
4:01
And now we're, we're again at this precipice and we're all
4:04
past it in terms of how AI systems are being built as we
4:08
speak. So I, I, I can't be more excited
4:11
by, you know, not only what we've been through for almost 30
4:15
plus years, if not longer, in terms of how software has become
4:19
the most important substrate for how humanity continues to
4:24
express itself in driving productivity around the world,
4:28
but also what we can do for the next 30-40, fifty years.
4:34
And it's very clear to me that everything, particularly in
4:38
software from this moment on is an AI driven feature.
4:42
And what does it look like yet? Because there's some things that
4:48
I know we were talking about where it's like AI systems are
4:52
taking more than software and it's really thinking about how
4:58
do we go beyond the screen. And it's, I think probably the
5:03
first step is going beyond the chat box and then we can go
5:08
beyond the screen maybe. And so how are you looking at
5:10
that? And what are some things that
5:11
you're seeing that excite you? Yeah, I want to set that up as
5:15
well. In terms of the March towards
5:18
this current generation of AI systems that are being built.
5:22
It goes without saying, we should really highlight that we
5:25
couldn't be building these modern AI systems without the
5:28
emergence of the modern data stack itself, right?
5:31
Because it's very clear now that you need high quality data to be
5:37
able to build your AI systems at scale.
5:40
And certainly that proved itself out with the Transformer
5:42
architecture that really got to the GPT 4 models, the O1 models,
5:46
certainly the things that Metropic has released and what
5:49
many of the foundation model builders have been able to
5:51
capably ascribe in terms of both knowledge and reasoning.
5:55
And that had been to, or let's say more precisely imbedded to
5:58
the models themselves. And so you see the March from
6:01
the modern data stack to now taking those models and defining
6:06
the machine learning operations surrounding how you build and
6:10
scale a bottle. And so more recently around LMS
6:14
and the operations are LM OPS and now the agent based systems
6:20
that are now possible, right? It's not just about what you can
6:23
do in the foundation models themselves, but how do you
6:27
orchestrate those effectively so that you have a agentic approach
6:31
to being able to have multiple reasoning systems, multiple
6:36
knowledge systems converge together to have a new AI system
6:41
being delivered for an existing or a new problem.
6:45
And so that's where I'm starting to get pretty excited right now
6:49
because I realize that most of software up to this point has
6:53
been based on rules and building, you know, highly
6:58
capable rules engines. But over time, rules get very
7:01
brittle because you have to continue to encode what your
7:05
business processes are into those rules.
7:07
But what if you didn't? What if you had almost a
7:10
reasoning system that can continuously reason what's
7:14
happening? And that reasoning could be
7:16
human like, perhaps even superhuman, like in terms of the
7:19
reasoning that you're coming out of an AI system today.
7:24
And that's exciting, right? Because it starts to change a
7:26
lot of where you can see the value proposition and generally
7:32
how software itself is consumed, right?
7:34
Because at some point up to this moment, you had to buy your
7:38
software, you had to enable someone from a services
7:43
standpoint to get that software working inside the enterprise.
7:47
And you had some people that were operating the software as
7:52
the personas that cared about that particular piece of
7:55
software. If you use the CRM software,
7:57
then salespeople, sales manager, sales operations cared about
8:00
that software. If you were an ERP software
8:04
provider, then you had a number of folks that were in finance
8:07
and operations that cared about that software.
8:09
Now this feature that we're barreling towards incredibly
8:13
quickly is 1, where the AI systems are encompassing the
8:19
software value proposition with the services in terms of
8:23
enabling that software in the enterprise.
8:25
And the first advent of digital workers synthetics like things
8:31
that we haven't seen before where there's a full autonomous
8:36
experience of how that software is now functioning inside the
8:40
enterprise without necessarily as much of the cost involved
8:45
with services in terms of putting that software in to
8:48
work. And certainly now even synthetic
8:52
SDRS, synthetic PDRS, synthetic developers, synthetic process
8:58
automation folks that can now, you know, instead of manually
9:03
processing that information, you have a full reasoning system
9:06
that's processing. That's very similar to what the
9:08
investment we made with Crew AI was in terms of just being able
9:11
to automate all the what would be historically manual robotic
9:16
processing with very little smart exception handling.
9:20
And now you have a high reasoning function involved with
9:22
how you do a lot of your back office automation and
9:25
processing. So that arc like software now
9:30
transforming all the way to the emergency of these AI first
9:33
systems and these AI first systems are encompassing
9:37
software and the services and labor.
9:41
I got something we haven't really fully seen yet.
9:44
It's just starting to emerge in some of these key markets and
9:48
I'm pretty excited with what that opportunity looks like.
9:52
Let's talk for a second about a few different papers.
9:55
I know that you pulled up what you really liked about Dario's
9:58
paper. And then there's these different
10:03
almost views of where we could go with the future and and it's
10:08
almost like 1 is hugging machines and the other is being
10:13
strangled by machines in a way. Yeah, that's funny.
10:17
Looks like set the context right.
10:18
So I think it starts certainly with Leopold Aschenbrenner's
10:21
paper in June the 2024, which is situational awareness, right.
10:25
He was outlining the next decade of her pad in terms of what was
10:31
necessary to build this future. And we'll talk about both the
10:35
papers in a moment, but what I think I saw in that paper was a
10:41
very mechanistic view of the future.
10:43
What would it take to scale out these clusters?
10:47
What would it be necessary to go from AGI to super intelligence
10:51
and what the time frame was to do that?
10:52
What does it mean to build a trillion dollar cluster itself?
10:57
Like how much we have to lock the labs down, which is really
11:00
funny that we talked about locking labs and security of AGI
11:03
would particularly the deep sea situation that's even six months
11:06
or whatever. Seven months.
11:08
It's like a profoundly different way.
11:10
At least the rest of the world has reacted to it.
11:12
Then at least of the paper coming came about it.
11:14
What does alignment look like? What does it mean to Dr.
11:18
Democracy in a free world while all this is happening?
11:21
It's just a brilliantly thoughtful paper.
11:24
I think it was also just a little heavy-handed in terms of
11:27
how it thought about the timing of AGI, like what would
11:29
happened. I thought one of the most
11:31
interesting thing was which whether and when it happened or
11:34
not didn't matter. That was really exciting was
11:38
this idea that if you could build a super intelligence, it
11:43
could automate the research output that's necessary to
11:48
create all the possible futures that you would want as humanity.
11:51
And I think that that was my, my, yes, that does make sense.
11:56
What no matter when it happens, that is incredible view of
12:01
unlocking cognitive labor, unlocking additional
12:06
opportunities for science and technology, unlocking robotics,
12:09
unlocking the military edge and more broadly just GDP.
12:15
But it does reflect a very mechanistic view of all this.
12:18
If you do these things, these outcomes will emerge.
12:21
And it is pretty almost brutalistic in terms of how I
12:24
thought about it. Wildly important paper though.
12:28
But then Rio comes out with, to your point to be just like
12:32
machines of loving grace. And it talks about, you know,
12:35
not a completely different future, but it talks about like
12:39
why these machines can actually help us as humans.
12:43
Like, how do we enable a better humanity because the machines
12:49
are helping us get there. It, it actually reminded me of
12:52
the I don't know if you came across the science fiction
12:54
series called the Culture series, right?
12:58
In Culture, it's the the systems, the IAS systems in
13:02
particular are just like infused into how we live, right?
13:06
And it's like a very post economic society that there's an
13:10
infinite amount of resources and how the machines have helped us
13:14
progress beyond our current constraints as humans and what
13:18
is like humanity look like in the future.
13:21
It's a beautiful view of what we think about the coexistence of
13:24
the convergence, the singularity of of humans and machines
13:27
together. And both are are definitely
13:30
possible. But I do find something
13:33
beautiful about Daria's thinking because we have this this
13:38
tremendous angst about the future of AI and its impact on
13:41
humanity. And I think there is a lot to be
13:45
worried about. I have note about about that,
13:47
particularly if we are the stage where we're building systems
13:50
that are not aligned to human values, certainly at scale, but
13:54
there is a possibility of reaching a future where humanity
14:00
itself progresses in an exponential way for the benefit
14:05
of humanity, because our AI systems are here to enable that
14:09
to happen. And I think I would highly
14:11
recommend anyone who hasn't read The Machines of Our Grace to
14:15
spend some time with that paper, in particular, both papers, but
14:18
certainly that one, if you want to believe in a more positive
14:23
future for AI, will transform humanity.
14:27
And, and one thing I wonder about as we think about that
14:31
shift, is it going to be a layer on top of legacy or is it going
14:39
to be built from the ground up? And the reason I ask is because
14:43
the I I was with some friends two weeks ago and they were
14:47
talking about how they tried to create a JIRA agent and it
14:51
worked great in their hackathon. And then when they plugged it
14:55
into their real JIRA instance, it fell flat on its face.
14:58
Because humans in Jira are doing the least amount possible to get
15:03
across what they need to get across.
15:06
And another human can see it and understand it because they have
15:09
all the context around it. But the agent was seeing it and
15:13
it had no idea what was going on.
15:15
And so I asked, would you then go and rebuild it with some kind
15:20
of a graph knowledge base around it so that it could have the
15:23
context? And he said something funny.
15:25
He said, no, I would go and I would just think about it
15:29
differently. Like instead of trying to
15:30
shoehorn an agent into JIRA, I would just think like how can I
15:34
make a better JIRA? Yeah, I think both things are
15:37
going to happen, right. I think there's going to be a
15:39
number of situations where you are going to be able to take
15:43
your existing systems, put that into an agentic workflow where
15:50
you are combining it with some of the reasoning and knowledge
15:53
that's imbued into a foundation model.
15:56
That might be fine-tuned or you might have a which you log into
15:59
generated view of just giving contextualization for how that
16:04
model is going to behave with that data that's provided from
16:08
that existing system. And you can get some tremendous
16:10
benefits for how a compound AI system emerges there.
16:14
And then to make sure to your point, like there's going to be
16:17
like a number of key new systems that are going to emerge that
16:23
you're just not going to want to take out existing process.
16:27
You, you're just going to come up with something completely new
16:32
that you don't even need to wrap your, you know, complexity of
16:37
the existing function towards and like just build something
16:41
like radically different from what you've done before.
16:43
And that that's a, that's an interesting idea.
16:45
Is there a different way of thinking about Jira and
16:50
ticketing and software development that we've been very
16:55
used to a process that could be radically re imagined.
16:58
All right. Yeah.
16:59
And I think it's a fairpoint. There could be quite a few of
17:02
those things that we're just used to doing in a certain way
17:05
and we no longer have to use AI to automate what we're doing,
17:10
but just completely create a new way of doing it.
17:12
Yeah, I'm for me, I feel like the, every time I see an update
17:18
from the cursor folks, I think, man, they're going to eat up
17:23
everything that a developer does.
17:25
And so it feels like there's a world, a future where everything
17:30
that a developer has to touch is in some way, shape or form
17:37
connected with. It doesn't necessarily have to
17:39
be cursor because cursor may not perform if we see another UIUX
17:45
change or shift, as we saw with with Copilot, like losing the
17:51
lead. And so you think about it that
17:54
way. If something is aware of what
17:56
you're doing and what you're trying to do, then wouldn't it
17:59
be able to update tickets for you too?
18:02
That's right. Yeah, yeah.
18:04
Go further downstream in terms of what the responsibility of
18:08
that user, that persona is and what they're doing versus just
18:12
the task at hand, right. Yeah, Yeah, this is where I
18:15
look. I've shifted my thinking on this
18:17
quite a bit even this past year, because I, I think about a year
18:21
and a half ago, a lot of my thought process was that most of
18:25
what this AI based emergence in software is going to look like.
18:29
It's like a bunch of call it Co pilots, right?
18:32
You have work and this thing is going to just help you do that
18:34
work better. And that's generally the the
18:38
mindset that that a lot of folks have had on Co pilots.
18:41
Now I've really embraced the idea that you can see full
18:47
autonomous autopilots, and this full autonomous autopilots don't
18:51
have to even do the work the same way you've been doing it as
18:54
a human. Wait click into that a little
18:58
bit more because it makes me think about how when they let 2
19:02
chat bots loose they created their own language.
19:04
Yeah, that's. The beauty of.
19:06
Some of the probabilistic capabilities of the models that
19:12
we can now see these emergent behaviors that we were not
19:17
expecting just because of the constraints of like how we think
19:21
about it as humans. But these probabilistic models
19:23
can interface with each other in ways that we probably haven't
19:28
even thought of. And so I, I do think there is
19:31
this value in we, we worry about confabulations and
19:36
hallucinations, but I actually think there is some significant
19:39
value in the systems that are emerging here, the probabilistic
19:45
systems that are emerging here to have different conversations
19:50
with each other than the ones that they would typically have
19:54
with existing business processes and existing people that are
19:57
interfacing with. So I, I think there is a lot to
20:00
be said there. Now look there, there's a lot
20:02
more to go because it does feel like the, the vector of
20:08
reasoning. There's certainly quite a bit of
20:11
advancements have been made with the emergence of one model.
20:14
Certainly we can talk a lot about where DeepSeek is.
20:17
But you can sense that the reasoning vector has been really
20:22
kicked off this past year. Even as the knowledge vector,
20:26
the accumulation of knowledge, we're probably slowing down
20:28
because there's less and less data that is pristine to consume
20:34
from the public Internet. And therefore, we're not using
20:38
private data sources, etcetera, to be able to train this next
20:41
generation of models that are coming to market.
20:44
Yeah, there's that vector. Do you feel like there's other
20:46
vectors that we are going to pull on too?
20:49
Because I look at the idea just how humans now are OK with
20:55
waiting 20 minutes for an answer.
20:58
And when Chad GBT first came out, that was unthinkable.
21:04
We wouldn't wait around for 20 minutes.
21:06
But now I use deep research all the time and I'm like, yeah,
21:10
just go and do it, whatever. And so that's another vector of
21:13
the patience of a human. And then you have the actual
21:19
like tech vectors that you're looking at.
21:20
Like you're saying, there's the reasoning or maybe there is some
21:24
kind of there's other things that you've been thinking about
21:26
that we could augment? Keep in mind a lot of ways the
21:29
reasoning vector encompasses some of the things that you just
21:33
mentioned, right? Because if you talk about the
21:35
notion of the chat itself, reasoning inside the chat is
21:39
instantaneous, right? It's synchronous.
21:42
It's immediate. You want that answer right away.
21:45
But you can also think of as you advance into more complicated
21:50
longer form thinking, the short and long form thinking.
21:53
And so you get into some of the longer form thinking, then the
21:56
reasoning can take longer, right?
21:58
You can think about it and come back with an answer that doesn't
22:01
have to be synchronous to the chat.
22:04
It can be quite asynchronous to go to your point, do the
22:07
research and come back. And we're patient about research
22:10
being done as long as their thinking is higher quality at
22:13
that point with a longer form response time.
22:16
So I do think that it's been proven now that they're, just
22:21
like humans, have short form and long form thinking.
22:25
There's no reason why our AI systems shouldn't also have more
22:30
immediate short form as well as longer form reasoning and
22:34
thinking. It's funny to think about the
22:39
idea of hallucinations as a feature and allowing those to
22:45
happen in a way or almost like you're giving the space for
22:50
models to go and get creative. And they can do that over here,
22:55
maybe in a little sandbox and then they can come back with
22:58
some wild ideas. And you have the almost like the
23:02
judge that is going to say, does this work or not?
23:06
And then reason with that idea and might be total trash, might
23:11
not work at all. But in a way like giving it be
23:16
directive to think outside the box and almost like hallucinate
23:21
on this is a funny way of looking at it.
23:24
Yeah. And when you think about the
23:25
type of work we're either augmenting or fully embracing as
23:32
a autonomous system, you can see vast quantities of, of creative
23:37
work where you do want, you know, a highly call it diverse
23:43
output that you can iterate through and just have the
23:46
models, you know, to lesser or greater degree genuinely
23:51
hallucinate and come back with output that you wouldn't have
23:54
thought of and leverage that quite effectively, particularly
23:57
in, in many of these creative domains, right?
23:59
You do want that constraint to be relaxed to, to get more, more
24:03
possible creative things out of the, the model that you might be
24:07
working with. And then in the other cases, if
24:09
you're doing a trade settlement processor or shred through
24:12
process in a back office, that's the last thing you need.
24:15
You just need a very tightly wound model that doesn't go off
24:20
its guardrails. It has the human like reasoning
24:23
to handle and understand like when exception occurs, like how
24:26
to process that. But then everything else, it
24:29
literally does what it needs to complete that process.
24:33
And I think that's where I'm really excited honestly, with
24:38
where the recent developments run, like something like
24:43
DeepSeek has has really opened everyone's eyes.
24:46
There's a lot of things that we could talk about or with
24:49
DeepSeek, which happy to do it. But one of the things that
24:52
really caught my eye was like you can take a very large model
24:58
and drive model distillation to a point and using some really
25:02
compelling techniques to just be able to hyper optimize the
25:07
efficiency of how a model is not fully trained.
25:10
But also inference. Using reinforced learning,
25:14
transfer learning techniques, you can build very interesting
25:18
domain specific models that really stay on the rails, do
25:22
exactly what they need to go do, have high quality human like
25:27
reasoning involved in it and doesn't you know, go off and
25:32
think about 14th century British history.
25:34
So I think that is a, a, a tremendously valuable element of
25:40
what we are going to likely see that it's not just this idea
25:44
that you need a go from a trillion parameter GBT 4 style
25:50
of like large language model to whatever 10 or 20 trillion
25:55
parameter model or maybe even bigger.
25:59
There's a lot of questions about how that's going to emerge.
26:01
And see, I'm sure the, the researchers are, are working
26:04
through that. But there's something to be said
26:07
that these elegant compact domain specific models, and I
26:12
said this years ago, I just never saw a manifestation of it
26:15
like we see with Deep Sea, are going to be just as important,
26:19
if not more important for how the proliferation of AI systems
26:23
function in that elite, the enterprise, but probably most of
26:27
humanity. Yeah.
26:29
Let's keep pulling on the thread of DeepSeek right now because it
26:32
is so fascinating. It's captivated the world.
26:35
And one thing that I am wondering about is your thoughts
26:40
on the other model providers and what that's done to their
26:46
positions? It's got to raise a bunch of
26:49
questions like I'm not inside of any one of the model providers.
26:54
And just to stated it out loud, Insight never invested it to any
26:59
of the foundation model builders, which not because we
27:02
didn't love what the foundation model builders were doing.
27:06
We didn't have deep faith in what was coming out of the model
27:12
builders, factories and research shops.
27:15
We were profound believers in a lot of ways. 1 was just the
27:19
valuations. We never felt right at any
27:22
moment of time and hindsight P20 there was there was probably a
27:25
few that we should have been in right at certain valuations like
27:28
a pretty sure like the $1 billion valuation route that Pat
27:31
Anthropic was a good one and we we had a chance to be in that
27:34
one. So you, you think about that
27:36
these anti portfolios that you build over time, that's another
27:39
one in my anti portfolio for sure.
27:41
But we just had a very strong belief that the data was the
27:47
most important substrate. And I'll talk about some of the
27:50
sort of thematic investments that we've made, particularly on
27:53
the the idea that the data was the the most important substrate
27:56
for building these AI systems. And we also just knew that with
28:02
how much was continuing to popularize in terms of open
28:04
source models was we're we're seeing and hugging face like
28:08
there was always a moment where the best model in the world from
28:14
a knowledge and or reasoning standpoint continuously had to
28:18
call it the near equivalent emerge within 6 to 9 months
28:21
afterwards. The open source community and
28:25
DeepSeek is probably one of the best examples of that, at least
28:28
for the reasoning vector. Like this is equivalent to what
28:31
the O1 models, if not as good or better than what the O1 models
28:34
can do and done with hypothetically a lot less
28:39
resources. So if you're a big model
28:42
builder, you got some were cut out for you at this moment,
28:45
right? Because you got to show this
28:46
vastly incremental traversal of either the reasoning vector or
28:53
the knowledge vector, which again, the the model builders
28:56
are, are are really thinking about as they're training these
28:59
next foundation models, frontier models that that they've been
29:02
working on. But when you do that, you're
29:05
spending hundreds of billions of dollars across the network, the
29:11
GP, US, the power required, the resources required to make it
29:16
happen. Up to this point, every time a
29:20
significant model evolution has occurred, it does seem there is
29:24
a more elegant, efficient way to catch up.
29:27
And and certainly that's what we should worry any of the model
29:30
builders, right, Particularly in the large scale research shops
29:33
that are building these Frontier Foundation models, like when do
29:38
they actually have a true Moat that is sustainable for more
29:41
than six to nine months? And that that should worry
29:44
everybody. I heard, I remember that even
29:47
one of the more efficient model builders that were open sourcing
29:52
like Meta and the lava models, like I think there was a moment
29:56
where Zuckerberg was probably yelling at his teeth about, OK,
30:01
how is it possible that you're asking for why amount of dollars
30:05
to go build this next generation of llama for and beyond models?
30:10
And this is out here doing it at a fraction of the cost with some
30:15
very clever methodologies and distillation techniques.
30:20
So yeah, it's everyone's got to be thinking about this, right?
30:24
But it's everyone who has spent hundreds of billions of dollars
30:28
up to this point have to be thinking probably important,
30:31
right? Important like the, I used to
30:34
talk about this as you needed the space program to build the
30:38
aerospace industry. And yes, we do need to put the
30:43
fundamental dollars into the space program.
30:45
But now we're at a point where it does feel like we should be
30:49
able to build the aerospace industry.
30:52
And so at least in the analogy, what I'd look at is like the AI
30:55
research versus the industry that's forming around AI
30:59
systems. So when we're doing that, should
31:01
we be spending the dollars that we had to spend on the research
31:05
to get there and how much more is really necessary?
31:08
So there's got to be a lot of questions going on and the the
31:11
big research about this exact topic.
31:14
That's a great analogy, bringing the aerospace industry to the
31:19
2025 basically and making sure that we're building on top of
31:25
what has already been established in a way.
31:29
And now it's just time to take off.
31:31
And so there there is another thing that is interesting when
31:37
it comes to these models being built.
31:40
In my eyes that is it's a bit of a crapshoot and you don't really
31:45
know what you're going to come out with after you spend all
31:47
that money. And it could be good, it could
31:51
just be mediocre. And I think we've seen a lot of
31:56
models that come out, they don't really make a splash because
32:00
there it's cool, Yeah, there's another foundational model
32:03
that's out and we're and almost like this base model that you
32:06
can then go and fine tune, but you quickly see them fade away
32:13
into history because they're not that good.
32:18
Yeah, compared to what's out there.
32:20
Yeah, yeah. Now I think the long tail of
32:25
models littered on hugging faces as much in indicator of some
32:33
incredible things clearly in the leaderboard.
32:35
But also there is a as you quite a bit of a long tailor.
32:40
And I think it's just shows that there is a, there is definitely
32:47
room to continue to create these, these beyond expected
32:53
outcomes like we just saw with DeepSeek.
32:56
But to get there, there's a lot of failure that you need to
32:59
iterate through. And sometimes that's happening
33:01
inside of a specific research shop in their work.
33:05
Sometimes it's happening across research shops.
33:07
And certainly it shows in long tail of what we see on Hugging
33:10
Face today. Yeah, the the interesting thing
33:14
too from my side, I guess not as interesting, but I don't know if
33:19
you saw that Mistral dropped a new model a few days ago and I
33:26
didn't mean it was like bad timing, man.
33:29
It's. Like just a complete, yeah,
33:31
tough, tough choice to to have to put a new model out there
33:34
this week, yeah. Yeah, but Mistral was that in
33:37
the past, like Mistral had their moment of they put out an open
33:42
source model. It was all the rage.
33:44
Everyone was talking about it and I think they got a little
33:48
confused that oh, we could do it again and we'll put it out.
33:51
But it wasn't. Again, going back to the quality
33:54
of the model, it's not this groundbreaking thing.
33:57
It doesn't have this whole narrative around it.
33:59
It's not capturing and captivating everyone's eyes.
34:03
And in my eyes, it's like you do have the Mistrels of the world
34:08
that are in a bit of a tough spot right now because of how
34:12
they're trying to do things. And somebody sent me a article
34:17
that said that Mistrell's trying to go public, and it was out of
34:21
all of the model builders, I was not expecting Mistral to be the
34:26
first one that goes public. I couldn't be honest.
34:28
That wasn't on my Dingo card. Yeah, yeah, the, the, the race
34:34
is continuing across the existing model providers.
34:39
And yeah, I would agree that, you know, you'd expect a little
34:45
bit more scale, right? For the few that would end up
34:49
going public, you you certainly would expect a bit more scale in
34:51
terms of where they are at their perspective journeys.
34:54
We tend to see a little bit of this bit from the view of
34:58
weights and biases being a a an investment that my first
35:02
investment, the fact that I've made that insight and being the
35:06
experiment tracking parameters, tuning version control, de facto
35:11
solution. It turns out that open AI,
35:15
anthropic Cohere, Mistral, Meta, NVIDIA are all customers using
35:22
always biases for building high quality models.
35:26
And because of that, you get a pretty interesting sensibility,
35:30
even just in terms of the amount of users there are in these
35:33
respective organizations and how much the leverage they're
35:35
getting from A2 like WB to conversion build their models.
35:40
It is, it is fascinating to see that, you know, there's
35:43
definitely like this, we call it core of fiber 10 research shops
35:48
that are doing the work that's leading the market.
35:51
But then there's these moments where everything gets disrupted
35:55
or reinterpreted when something like Deep Sea comes along and
35:59
then you have to think about what just happened.
36:02
But just seems like there's still great work coming out of
36:04
open AI and Anthropic and coherent with Shell and Meta and
36:10
NVIDIA for sure. Yeah, yeah.
36:12
And you need these moments to raise the bar and keep people on
36:15
their toes. And also it it changes the
36:18
landscape. It makes things a little bit
36:21
more. Oh, cool.
36:23
I can go use DeepSeek now. I'm going to think twice about a
36:27
different model if I can host it on my own.
36:29
But it's not like you just grab the model and then you're good
36:36
and you can be all right with it.
36:37
I think the other pieces our we just had this thread in the
36:42
community that was fascinating to me because there was someone
36:47
asking about the cost differences that people have
36:51
seen between using Sagemaker and not using Sagemaker.
36:55
And somebody chimed in on the thread and said yeah we saved 5
37:01
grand a month from switching from Sage maker AI to just
37:07
regular. I think they were using what is
37:10
it ECS and EC2. Yeah.
37:13
EC2. That was the EC2 instances.
37:16
Oh, so they they didn't have a, call it manage Sage maker
37:22
experience. They just rolled their own tools
37:24
into AW Sempra. And they saw that wow, OK, now
37:28
we're saving 5 grand a month. But what you don't see?
37:32
And they were very transparent about this in the thread was how
37:36
much is the cost of having engineers that know how to do
37:41
this? What are you really trying to
37:43
optimize for here? How big is your team?
37:46
How mature is your team? All of that stuff comes into
37:48
play and it's that classic build verse buy decision.
37:52
And so you can't just think, all right, I'm going to just grab
37:55
DeepSeek now and I'm good. I don't have to worry about
37:57
anything because as another friend told me, he said whenever
38:01
possible at my job as AML platform engineer, I am trying
38:07
to outsource the ML platform to one of the big model providers
38:11
because I don't want to have to deal with that.
38:12
It's such. A headache, Yeah, No, it it
38:15
mean, it certainly makes sense, right?
38:18
Because if you're one of those ML teams, your job is to
38:27
introduce the highest quality AI system that incorporates that
38:31
model, that brings in the data that's necessary to train a
38:37
domain specific model for your needs and get that into a
38:42
production workload that is actually benefiting that
38:47
organization. And yeah, the more you can get
38:50
platforms and tools that help you support that.
38:52
It just, it makes sense. If anything, that was for lack
38:55
of better expression the picks and shovels thesis that we added
38:59
insight for everything in the space.
39:01
We just definitely bleed in on most of the picks and shovels in
39:04
this area. Yeah, and he's even in a
39:07
regulated space and he was saying unless you have data
39:11
issues and regulations, just outsource it, picks and shovels
39:15
it and get rid of it. And then you can spend your time
39:18
on higher priority issues. And so it's fascinating to think
39:23
about that the. Yeah, By the way, I've come
39:25
across certain enterprises where I've seen, Oh yeah, we filled in
39:28
with a lot of pride. Like we built very custom
39:33
bespoke ML OPS pipelines and, and I was like, oh, you, you
39:39
poor things. Like why would you put all that?
39:43
Energy into that. When you could and some of it
39:46
decisions were made a certain time.
39:48
There's like a belief that you couldn't take everything from
39:51
what's available in the market and coerce it and harness it
39:55
together to make it make sense. But yeah, I got to imagine like
39:59
that's not got to be a great use of resources, particularly in an
40:03
organization. It's and you're always having
40:07
the conversations every quarter I imagine or every year you're
40:11
having a conversation like do we continue to support this or do
40:15
we try and migrate to something newer?
40:18
Because in, in a way when you choose your stack, it's a
40:24
snapshot in time of what the best capabilities are at that
40:28
moment in time. And in if you're really going
40:31
all in on it. I remember we had a guy from
40:33
Pinterest on here and a lot of the stuff that they are using
40:38
for their ads platform was chosen in 2018 or 2017.
40:46
And there's been so much advancement and they have to
40:51
decide, do we you manually or maybe you take a piece off here
40:57
and then upgrade it here, but maybe it doesn't fit.
40:59
And so you can't add that new shiny thing, which can be good,
41:04
it can be bad. You don't always necessarily
41:06
need the new shiny things. But but really thinking about
41:09
like how if you decide to go down that route, you're deciding
41:16
to take what is out their best right now and have that snapshot
41:23
be what you're using. And then forgo all the future
41:26
innovation that could come A. Little bit, huh?
41:29
That's the the things that are outside your 4 walls.
41:32
Yeah, that's the thing. You got to be careful about
41:34
these build versus buy decisions.
41:36
And this is not just in your ass and fit in just general.
41:40
If you're going to lock down something now you're going to
41:44
get likely something that is very highly specific to your
41:49
needs, but then eventually just just going to be this external
41:52
innovation that's going to just keep moving ahead.
41:55
One of the things that are fascinating to see, by the way,
41:57
since we did talk, we're talking about building buy is also this
42:01
wave in AI based software systems or AI systems that are
42:07
empowering and enabling this texture and software to be more
42:10
precise is how much more you can now do custom software
42:16
development better. And the, the reason I just
42:18
mentioned this is that before you had to almost rely on your
42:21
package software provider to do the things that you needed them
42:25
to do in the product and you're waiting for those releases to
42:28
happen. And then you were implementing
42:30
that package software. But we're at a world where you
42:34
can build taking some of the componentry that we're talking
42:38
about and the componentry, the data, the pipelines, the
42:45
building blocks can continue to evolve.
42:47
But then how you harness that can be very much highly
42:50
customized software experiences that are for all intensive
42:55
purposes custom software with a number of really upgradable
42:59
underpinnings like the life cycle management of it becomes
43:01
much more straightforward. So I do think there is a moment
43:04
here that you can just deliver highly specified custom
43:09
software, but the advancement of all the underlying componentry
43:14
to do it continues to have its own upgradeability life cycle.
43:18
That's not where you have to build bespoke stuff.
43:22
It's actually in the layer, the abstraction layer above where
43:25
you can continue to iterate that faster because you have these
43:29
building box you're working from.
43:31
And it's it is really fascinating to see how much you
43:33
can get done today from a custom software development than ever
43:36
before. Is this because you're thinking
43:40
about it as I can prompt my software or my software, I'm not
43:47
interacting it always or I'm not interacting with it always
43:52
through a GUI, It's also just through language there's.
43:57
That or just the way that the the UI falls out in this
44:00
experience, but it's also just a when you have to now take your
44:06
existing software packages and almost bolt on the AI
44:10
capabilities for versus like natively building it up from the
44:15
bottom up. And then natively having a very
44:20
enterprise specific functional experience with your software
44:25
that you're not waiting for someone to develop a feature
44:28
that you need versus you're capable of building those higher
44:32
level features yourself. I think you're getting much more
44:35
specific curated software built for your business, right?
44:39
And that, that and, and by the way, like there's moments you
44:42
don't want to do that, right? When everything's settled and
44:44
there isn't a lot of innovation that's going on and you can just
44:48
take something off the shelf and it's as good or better than
44:50
anything else that you can build.
44:51
But we're at a moment where you can build stuff better than you
44:55
can take anything off the shelf. And so that's why I think this
44:58
would be compelling moment for custom software development in
45:02
the middle of all this. Yeah.
45:04
And it is looking at that abstraction layer and saying
45:07
that we're going to benefit from things as it continuously moves
45:12
underneath us, but we can have what is needed above for the the
45:18
users. And yeah, it is interesting to
45:22
think about because it does feel like it is more Lego blocks now.
45:28
And since it is so new and there isn't really much settled and
45:34
every day you get something new, you are more inclined to build
45:44
quicker or build on your own. And I imagine, I'm just thinking
45:49
back to the Pinterest story and how I bet in those days there,
45:53
they went out and they were like, look, there's not really
45:56
anything out there that's good enough.
45:58
We can create the abstraction layer we need and then we'll
46:01
have our custom stuff underneath.
46:04
And that is getting us to where we want to go.
46:09
But inevitably, like you said, it changes.
46:14
There's that. Right now we're in that phase of
46:18
it's easier to build for this type of stuff and create our own
46:22
custom software. And then you probably will get
46:25
over the hump and you'll start to realize, all right now
46:28
there's offerings that are pretty nifty and we're going to
46:31
want to go with those instead. Yep.
46:33
And there's another nuance to add in since you mentioned
46:37
Pinterest. When you look at some of the
46:40
more compelling open source that has actually emerged this past
46:44
decade, it actually started as custom functional work that was
46:49
needed inside a civic organization like, oh, other
46:52
people might actually need. Airflow came from Airbnb, yeah,
46:55
right. Kafka came from LinkedIn.
46:59
And so you look at these like now modern OSS components, they
47:04
actually emerge from a very domain specific need that
47:08
happened to be in a very specific domain and then
47:12
generalized into the rest of the world by just releasing it as
47:16
open source. So that there is some
47:17
interesting benefit to even building these things in a way
47:21
that you're getting what you need out of it through that
47:24
customer experience. And then when you realize there
47:27
is some things that other people will also need that are is
47:30
reusable and there's a benefit to then open sourcing it.
47:34
And we certainly saw that with the number of open source
47:37
capabilities, particularly this last second.
47:39
Yeah. What a great example.
47:42
Speaking of open source, there's a little open source project
47:47
called Spark. Oh, that one.
47:49
Is that still around? Is that still doing anything out
47:51
there? People are.
47:51
Trying to kill it, yeah, but it's still.
47:55
Around yeah, I got a fun story about that by the way, about
47:59
Spark yeah I was at Spark Summit 2015 and my team at Altrix
48:08
including myself worked on shipping Spark R and our Co
48:14
developers on it was a very clever VP of engineering at data
48:22
bricks name Oligarchy and his team and that's how we got Spark
48:27
R into the market. No way yeah.
48:30
So you were working. With him back in the day and now
48:32
you're working with him again because I saw recently you
48:36
posted you were part of their. Series J Yeah, one would not
48:43
have imagined that data bricks back then would have arrived at
48:46
its Series J, but here we are a decade later.
48:51
Yeah, how is it even possible that a series J exists?
48:55
That's one thing. And then why now?
48:59
What is it about you? And you've obviously been in the
49:06
data break slash Spark ecosystem for a decade plus.
49:13
Why all of a sudden you do you want to become a shareholder?
49:18
Yeah. Sometimes it is OK to be late to
49:19
a party, right? Yeah.
49:21
But in this case, it's a little bit more methodical than just
49:25
being late to the party. We were investors about two
49:27
rounds ago with a much smaller position and we're continuously
49:33
excited by how much the evolution from Spark to
49:38
Databricks to really what Databricks stands for today as a
49:43
unified data management platform using the lake house as the, you
49:50
know, construct for all the preprocessing, all the data prep
49:56
at petabyte scale. And then this little capability
50:01
that they've introduced called data warehousing inside of
50:04
Databricks, which is now the data warehouse product is a $600
50:08
million business for data bricks, which is incredible,
50:11
right? When you think about it, as you
50:13
sum all of these things up, the Mosaic acquisition, the in house
50:19
built capability around data warehousing, the Unity catalogue
50:23
itself, that's how it becomes a very unifying layer for the
50:27
enterprise, the tabular acquisition and now owning the
50:31
iceberg format. And that's just like the product
50:34
strategy. We're not even talking about how
50:35
much to go to market and industry strategies have really
50:40
evolved from even even its early inception.
50:43
It it just feels is a generational company.
50:46
I don't think I'm saying anything out of turn by by
50:49
saying that out loud. And it does feel like there's
50:52
many more miles to go. And so that was our thought
50:55
process there. The opportunity came out to be
50:58
part of a very large $10 billion Series J, which is I think the
51:05
largest venture capital deal ever.
51:08
Our friends at Thrive and Josh Kirschner LED it.
51:11
We Co LED it with a pretty sizable check that you know was
51:15
really helping data bricks continue its journey.
51:20
And for them, it was just making sure that there was tremendous
51:23
amount of just liquidity for everyone who had been on this
51:28
journey for a long period of time to consider their needs as
51:34
they continue to grow and scale the company all the way to the
51:37
next levels of stages of growth. Yeah, we couldn't have been more
51:41
excited. And it certainly was a was a big
51:44
investment for Insight. It was a very large round for
51:48
everyone that was involved and it does appear like there's many
51:53
more miles for data books to go in the market and we're just
51:56
delighted to be a part of that journey.
51:59
And I guess one would say it's never too late to come to a
52:03
company like that and just help them along their journey.
52:07
Yeah, you say like generational and it really does feel like
52:14
that. And I instantly thought, do you
52:18
feel like there are any other generational companies in this
52:21
space as you look around in the data AI space?
52:28
Because for me, it's like Databricks is just performing
52:36
top notch right now. But I don't know, maybe you're
52:40
looking at other companies and thinking, yeah, they're
52:45
generational companies too. And or maybe it's if some things
52:49
go right, there's going to be another one of those.
52:53
Yeah. Look, there, there should be, by
52:56
definition, the next generation of companies that emerge.
53:00
And so they're coming. And I just want to be clear
53:03
about that. But it does feel like there's
53:05
something special this moment, certainly with data bricks, you
53:08
know, we shouldn't, you know, not mention what Snowflake has
53:11
been able to accomplish as a company in that same cohort and
53:15
period. If you look at Palantir as
53:18
another great example of a company that has in a very
53:22
different way, build something generational in nature.
53:27
Yeah, it's it is exciting to see those folks who started 1015
53:31
years ago build the companies that are now the sort of the
53:36
heme. It's the foundations of how at
53:38
least the data AIML space continues to to progress.
53:42
I think there is a clearview of what's next, right?
53:45
Without stretching too much, you can certainly say that the
53:48
foundation model builders and the growth and scale that
53:50
they've respectively had have been quite impressive.
53:53
The enablers of that growth and scale like a company like Scale
53:57
AI for instance, is pretty impressive as well.
54:00
And just being able to initially provide a lot of the necessary
54:06
efforts to be able to enrich data that's necessary to build a
54:10
model and now providing things like our LHF for the model
54:14
providers themselves as they're coming to to market.
54:17
So it does feel like there is something surrounding, you know,
54:21
these current model builders and the you know, capabilities
54:25
underlying the model builders that have gotten pretty big.
54:29
It is a little trickier, right? Because when you look at
54:33
something like data bricks, it hugely is benefited by DeepSeek,
54:38
right? And the reason I mentioned that
54:41
is if you think about what DeepSeek is a manifestation of
54:45
is, and I think everyone's been talking about this past week,
54:47
it's like this idea of, of Jevin's paradox, right?
54:50
As the cost of a resource comes down, you consume that resource
54:54
more and it takes up all the capacity that's available to you
54:57
at that moment. So I think that should be the
55:00
case for everything related to the GP, US, the computer
55:03
infrastructure that we're talking about over time.
55:06
But we have a moment here where no matter what happens, you just
55:12
need high quality data to be able to build this next
55:16
generation of models, whether they are hyper efficient domain
55:21
specific ones that you can see emerge from like the likes of
55:25
DeepSeek and it's cousins that will literally come out within
55:28
the next weeks and months, as well as some of the largest
55:31
foundation models in the world. So we feel pretty good that
55:34
we're sitting in a good place with the data bricks investment.
55:36
It feels like it's going to be a bit of a roller coaster for a
55:41
little while for the foundation models themselves.
55:44
And look, the next generation of companies are coming about and
55:49
getting to to scale. I mentioned the weights of
55:51
biases example, right? There's 900,000 machine learning
55:54
practitioners that are now standardized some weights of
55:56
biases. There's a lot more miles to go
56:00
for a company like weights of biases now.
56:02
So I do think it comes in these waves, right?
56:05
And so the wave of like full scale and maturity is out there
56:08
and we see them as public and private companies.
56:10
Now this next wave of particularly around the
56:14
foundation models and the support infrastructure around
56:17
those foundation models have come to fruition as we speak.
56:21
And then you're now starting to see this next, you know much
56:25
smaller companies, incredible founders starting to build that
56:29
will get to the next level of scale in the 510 year time
56:34
frame. Yeah, that is one thing that you
56:38
have a nice position of being in is the almost like round
56:43
agnostic ability that I know you don't do super early precede,
56:48
but you still see really early companies and you get to see and
56:51
participate in the later companies too.
56:54
We've been very specific and inside about this.
56:56
So we've even said this out loud.
56:58
We don't think that stage is a strategy.
57:02
So we don't believe that this idea as investor stage is a
57:05
strategy. We should be able to find the
57:09
generational growth opportunities at whatever stage
57:12
that they might exist. And we tend to be a little bit
57:15
more focused on that middle stage of growth and certainly
57:19
have done quite a bit of work, particularly in our teams that,
57:22
that, that do a lot of the scale that work in, in terms of
57:25
buyouts and some of the larger checks that we, we've written.
57:28
But that's not a strategy. It is that we look for growth
57:31
everywhere. And one of the, one of the
57:33
wonderful things I had a chance to do working with my partners
57:36
for the last four years is really build out our early stage
57:39
right where we are now very capable at not only the series
57:45
age check, but occasionally even the seat checks, right.
57:47
Because we see these opportunities and we see that
57:50
profound growth happening with a founder that just really has a
57:57
maniacal view of what they want to deliver and build in the
58:00
future and automatically want to be part of those conversations.
58:04
So that certainly has been part of our our strategy.
58:06
So work with founders whenever and wherever that growth is
58:11
existing, independent of Sage. How are you thinking about when
58:18
it comes to the way that we are interacting with computers right
58:21
now? We have very compartmentalized
58:24
applications, but it almost feels like some of the stuff
58:28
that you're insinuating at it would be better to just have the
58:32
one AI application that can know and be with us, right.
58:38
And you mentioned it about the sci-fi series, how it's called
58:45
culture. And in order to enable something
58:51
like that, we need this. What I, I think Microsoft is
58:55
trying to do it right now with recall where it's like this
58:59
always on AI that can come and it can help you and be there
59:03
with you when you need it and then just be in the background
59:06
when you don't. And for me, and the way that I'm
59:09
trying to wrap my head around it is, yeah, it's much better if
59:16
you have all of the context to give to the language model.
59:21
And then it can reason about it and it can understand when we go
59:25
back to that Jira example, if it doesn't have the context, then
59:28
it gets lost. And so in striving to get to
59:33
that point where you want to give all of the context, I think
59:37
what you end up doing is just saying, come with me everywhere,
59:40
be with, be like my shadow in a way.
59:43
And so have you thought about that and how that could
59:47
potentially play out in the future?
59:49
I have a little bit, Yeah, I have.
59:51
And by the way, interestingly, that description of a almost
59:57
ever present in how how do a moment say it like just like
1:00:03
completely in the moment available AI system that's doing
1:00:09
things that it needs you to you needed to do.
1:00:12
It doesn't have to be single and monolithic either when you think
1:00:15
about it, right? There could be a view that it's
1:00:18
like just hundreds of agents that are just continuing to
1:00:22
Marshall out to do many things and then converge into your
1:00:26
needs. So I don't know if I necessarily
1:00:29
see this sort of a singular thing that just is always
1:00:33
constantly doing the bidding of what you want it or other folks
1:00:40
would want it to do. I, I actually see a lot more, at
1:00:44
least in the near term, this emergence of massive amounts of
1:00:50
personal enterprise and almost hybrid agents that are
1:00:58
continuously doing little things, medium sized things, big
1:01:02
things on our behalf. While we're talking here,
1:01:06
wouldn't have been nice for an agent to go figure out how to
1:01:10
book that trip that I need to make to the Middle East and give
1:01:15
me all the optimum answers to do that and book the ticket.
1:01:21
Like that's the kind of things that we're starting to see with
1:01:25
these automations that are embedded into the foundation
1:01:27
models, but also just in generalized sort of agentic
1:01:31
systems that are coming about. So I do think that we're going
1:01:35
to end up in a world, at least in the near future that it's
1:01:40
about these really well meaning agents that are doing smaller
1:01:48
things, tasks that are wrote repeatable that should not have
1:01:54
to be done by any other human on our behalf.
1:01:57
And I think that should be how at least the near term futures
1:02:00
and then we'll see where that or science fictiony future of a
1:02:04
complete system that's always with you, which is part of your
1:02:09
being it. It seems like it's a little
1:02:12
further out, but certainly that is a possibility that could
1:02:15
emerge when we're all set and done.
1:02:18
Yeah. And I've thought about it as the
1:02:22
what you're talking about there is like where is it going to
1:02:25
live? Is it going to be in on the web
1:02:31
and or is it going to be something that you're starting
1:02:34
to see more is these ways that the products are being created
1:02:38
is directly in Slack, like a Devon, for example.
1:02:42
And then you just deal with these agents in Slack.
1:02:45
And I think it would be pretty funny if all of a sudden Slack
1:02:49
becomes your Control Center too and you just send out agents
1:02:52
from Slack to go do different things that you needed to do.
1:02:55
Yeah. It's actually I was searching
1:02:56
for the word because it was escaping me.
1:02:58
But it's the nature of how AI systems are gonna be ambient or
1:03:03
not, right? I think a lot about that, like
1:03:06
the ambience of our AI systems of now in the future and how
1:03:12
could they be so embedded, so ambient, so just an extension of
1:03:17
how our minds work that like it's just like we think about it
1:03:22
and things get done that that does feel like not a science
1:03:26
fictiony future. I think actually that that
1:03:28
we're, we're not that far away from getting to more of these
1:03:31
sort of ambient AI systems that are doing important work at home
1:03:37
and at in the enterprise and everywhere in between.
1:03:41
Yeah, because how many times, I don't know if it's happened to
1:03:44
you, you might have more willpower and focus than me, but
1:03:48
I have something in my mind that needs to get done.
1:03:52
And then I pick up my phone with the intention to do that thing
1:03:56
in the moment. And some kind of a notification
1:04:00
distracts me, of course. And 30 minutes later I'm like,
1:04:04
oh, what was I? What did I pick up my phone for?
1:04:06
I can't even remember anymore. And that task that I wanted to
1:04:10
do is in the ether. And maybe it comes back to me,
1:04:14
maybe it doesn't. But what if he had an ambient
1:04:17
agent that the moment that task came up, fires off and goes and
1:04:22
does that task in a highly reasoning long form way 20-30
1:04:26
minutes later, or even longer if needed?
1:04:28
Yeah, I'm just excited for a Siri that actually works and
1:04:32
then I can give like tasks to because right now the Siri that
1:04:38
I use, it's very hit or miss. And so you do see the world
1:04:43
where, yeah, typing into a Slack chat is great, but I also have a
1:04:50
million notifications on Slack, so it would just be cool if I
1:04:54
could talk to the phone directly or.
1:04:56
It does seem like a a more obvious opportunity for Apple
1:05:01
intelligence, right, just to be that conversational layer, but
1:05:06
doesn't seem like it's emerged out.
1:05:09
We'll see. Something's going on there.
1:05:10
I don't know what it is, but something, yeah.