YT YouTube2Text

YouTube2Text → Transcripts → AI Agents Are Here: What They Can Already Do—and What’s Next

AI Agents Are Here: What They Can Already Do—and What’s Next — Transcript

by The Generalist · 16,645 words · 2,273 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00The big question we see in the market
0:01today which is interesting is agents
0:03versus AI workflows. We do foresee a
0:06world where those agents will be actual
0:07co-workers and I don't think you can
0:09really encode a co-orker with a
0:10workflow. We're really bullish on trying
0:12to help [music] people create agents,
0:14not workflows. What is the right way to
0:16think about what an AI agent truly is
0:18and what it isn't? You can often do the
0:20same things with workflows and [music]
0:22agents. It's just the ease of how you
0:24describe it. In an agent, it would all
0:25be in natural language, right? Like you
0:26could have a recipe that you just put in
0:28natural [music] language. say, "Hey, do
0:29A, then do B, then do C." And it's not
0:31as deterministic, so it's not as safe,
0:33but it's way easier.
0:34>> I'd love to zoom out for a moment and
0:36talk also just about what it means and
0:38what it's like to be building in AI at
0:41the moment and some of the specific
0:42dynamics [music] that founders have to
0:44face.
0:44>> The fog of AI is the fact that the
0:46foundations are moving very quickly and
0:48[music] so you have to have a vision of
0:50where you're going, but you cannot paint
0:51it because if you paint beyond 6 months,
0:54whatever you wear painting will probably
0:56be not true.
1:01Hey, I'm Mario and this is the
1:03Generalist podcast. As the saying goes,
1:06the future is already here. It's just
1:07not evenly distributed. Each week, we
1:10sit down with the founders, investors,
1:11and visionaries living in these pockets
1:13of the future to help you see what's
1:15coming next. Today, I'm speaking with
1:18Stan Pulu and Harrison Chase about the
1:20future of AI agents. Stan is one of the
1:23founders of Dust, a Sequoiabacked
1:24platform that makes it easy for
1:26enterprises to build and deploy custom
1:28agents across their workforce. Before
1:31that, Stan was at OpenAI, researching
1:33the mathematical reasoning capabilities
1:35of large language models. It was during
1:37his time at OpenAI that he first met
1:39Harrison Chase, who went on to found
1:41Langchain, a popular developer framework
1:43that has become essential infrastructure
1:45for building agentic applications.
1:48Langchain has also become one of Silicon
1:50Valley's hottest companies, raising from
1:52Benchmark, Sequoia, and IVP. Both Stan
1:54and Harrison spend their days thinking
1:56deeply about AI in general and agents in
1:59particular. In our conversation, we
2:01explore what comes after chat, why there
2:03probably won't be one agent to rule them
2:06all, and the unique challenges of
2:08building a business in the fog of AI. I
2:11walked away from our conversation with
2:13practical insights about where agents
2:15are headed and how they might change the
2:17way we work, think, and build over the
2:19coming decades. This is a new podcast,
2:22so if you enjoyed today's episode, I
2:24hope you'll consider subscribing and
2:25joining us for the incredible
2:26conversations we have coming up. Now,
2:29here's my conversation with Stan and
2:31Harrison.
2:33Thank you both so much for for being
2:35with me today. I'm excited to to chat
2:37about AI in general and and sort of
2:39agents in particular with two people who
2:40are really spending their lives
2:42dedicated to this, you know, field and
2:44and and this part of the world. Uh maybe
2:47to start we could just begin with a
2:49little bit about both of your your
2:50companies and and what they do. So Stan,
2:53why don't you tell us about Dust?
2:54>> Yeah, well Dust is the place where you
2:56build agents for for Walker. So, it's a
2:59product that lets you create, manage,
3:02and uh operate agents uh with access to
3:04your company context and tools.
3:06>> Amazing. Uh and we'll get into the the
3:08details of what that means and and why
3:10it's become so powerful uh more, I'm
3:12sure, over the course of of this next
3:14hour or so. Harrison, uh yeah, I'd love
3:16to to have folks know about Langchain.
3:19>> Absolutely. At Lingchain, we build
3:21developer tools to try to make it as
3:23easy as possible to build a genfic
3:26applications. That's the oneliner, but
3:28I'm sure we'll get deeper into it over
3:30the course of the conversation.
3:31>> Amazing. Well, uh I think something that
3:33you both have in common is that you've
3:34been in AI, uh at least by the industry
3:37standards for a long time. Uh you know,
3:40predating sort of the chat GPT moment in
3:42in November of 2022. I started in VC in
3:452016. That was my first venture job. And
3:47I remember there was a big sort of
3:49excitement around those first phase of
3:51chat bots. But, you know, honestly, lots
3:53of the excitement sort of withered for
3:55the years that followed. And I think
3:57there was a question mark about when
3:59exactly AI in the way that we see it
4:01today was really going to come to the
4:03four. And so I'm curious for you both,
4:05you know, what was it that you were
4:06seeing in the industry before, you know,
4:08it became so obvious to everyone else
4:10when chat GPT came out that excited you
4:13to sort of like build your career around
4:14this? and and maybe let's start with you
4:16Stan because I know you spent some time
4:17at OpenAI so you probably had you know a
4:19front row seat to a lot of of of these
4:21questions. So actually the fun fact is
4:22that uh we we chatted together with
4:24Harrison. I think it was uh maybe
4:26September 2022 or maybe October 2022. It
4:30was funny because it was a pre-hat GPD
4:32and we we naturally met together and
4:36spent some time chatting together
4:38because there was so few people like
4:40interesting themsel in the uh use of uh
4:43LMS for development or use of LLMs for
4:46doing anything else in product. And so
4:49that's for the the fun fact to to to
4:51give everybody the kind of a context of
4:53how small the community was uh during
4:55those few months that were the end of
4:58the summer 2022 all the way to Chaduki
5:00release which was in October or yeah I
5:03guess October. Uh so it's a fun fact as
5:05far as I'm concerned had the the chance
5:07to work at OpenAI. So uh I had been
5:09working on the uh studying the
5:11mathematical reasoning capabilities of
5:13LLMs but obviously was also exposed with
5:16GB4 which had under training uh early
5:19summer 2022. I mean it was pretty
5:21obvious to me that there was a massive
5:24disconnect between the capability of the
5:26technology and the actual impact it was
5:28having on the world. uh the revenue of
5:31of OpenAI uh probably not public and I
5:33don't even remember the right numbers
5:35but was like a drop compared to what it
5:38is today. it was just a few tens of
5:40millions of dollars. Uh and so compared
5:42to the power of GP4, there had been the
5:44luck of playing with internally. I think
5:46it was obvious to me that there would be
5:48there was something missing at the
5:50product layer uh to really unlock the
5:52use of LMS everywhere. And so that's why
5:55I've kind of motivated me to start
5:57building in the product space rather
5:58than doing research.
6:00>> And and just to go back to that moment
6:02in the summer of 2022, Harrison, do you
6:05remember like what you guys were sort of
6:06talking about at that point? like what
6:08was the tenor of those conversations or
6:10sort of the the contours of what you
6:11were thinking might be possible or might
6:14be interesting.
6:15>> So I think and Stan you should correct
6:17me on this but I'm pretty sure I just
6:19cold emailed you or cold DM'd you on
6:22Twitter or something. So, so I can share
6:23a bit of my background as well because I
6:25think it leads up to this. But my my
6:26background, so I studied stats and
6:28computer science, then worked at a
6:30fintech company doing more like time
6:32series stuff, but a little bit with some
6:34of the early BERT models on entity uh
6:36linking. And then I went to an MLOps
6:38startup where I was doing kind of like
6:39tooling for ML. Um and and then I
6:42remember in in like August, September, I
6:45was going to a bunch of meetups in SF.
6:47Um a lot of them were on generative UI,
6:49but they were more on like stable
6:50diffusion and image things like that.
6:51But there were a few people doing cool
6:53things with language models. And I just
6:54remember being like, "Holy crap, these
6:56language models are fantastic. They're
6:58just so different than like the
6:59traditional kind of like ML models that
7:01I'd worked with before." And so then I
7:04think I was just paying attention online
7:05and Stan I think was tweeting a bunch
7:08about stuff that he was working on
7:09around early versions of Dust. And I
7:11think I just like either DM'd or emailed
7:14um and uh he was very gracious to kind
7:16of like respond to that and and hop on a
7:18call. And I think uh I I think we had
7:22one or two calls and just kind of jammed
7:24on that and then kept on working. Yeah,
7:26I that was the I think that was the
7:28start.
7:29>> Yeah. Yeah, exactly.
7:30>> That's awesome. Um, yeah. So, you talked
7:33about this sort of discrepancy between,
7:35you know, the the power of the
7:37technology and the impact it was having
7:38in the world and and you know, it sounds
7:40like that was sort of the maybe the
7:42spark that led you to Dust, but like
7:44more specifically, what was it that you
7:46were saying here's the gap or here's the
7:48real thing that needs to be solved and
7:49like here's how we we go about doing it?
7:52>> To be very candid, at the beginning it
7:54was not necessarily perfectly
7:56identified. was really this thing is
7:58great nobody uses it and you're like
8:01what what is happening where where is
8:03the uh heat being dissipated it must be
8:06dissipated somewhere and uh it's unclear
8:08where uh and I think it was at the
8:10product layer and chibdity kind of was a
8:11confirmation of that hypothesis
8:13chibility in a sense is mostly a product
8:15as a bit of a model work compared to
8:17what was available off the shelf at the
8:19time but not that much it's just a very
8:21nice UI and the fact that it was free
8:24super interesting to think back But uh
8:27uh and I don't want to like sidet track
8:28you too much but character I was
8:31incredibly well positioned at the time.
8:33>> Character I at the product at the
8:35research team at the model. Why is it
8:38chipd that takes uh that takes it all?
8:40Why character I didn't explode at the
8:42time? Uh that's a very interesting and
8:45and fascinating question that we'll be
8:46able to study when we start doing
8:48history of that period I guess.
8:49>> Oh interesting. Uh I I appreciate the
8:52desire not to get too sidetracked but I
8:53do think that's interesting. like what
8:55what was what's your sort of working
8:57hypothesis of why character AI maybe
9:00wasn't the one to to capitalize in that
9:02moment?
9:03>> So character I I think was speaking to a
9:05to a specific population. It was kind of
9:07weird in many ways. You would like start
9:09by create a fake character or you would
9:12create Alen Musk clones or clones of
9:14Albertstein or whatnot. And I think it's
9:17a it's a it created a a complexity uh
9:21level even for the for the like the the
9:24pure B2C audience. It was maybe a bit
9:26too high. It had a very strong community
9:29which is the interesting bit. It was
9:31going very well. It's just that it
9:33didn't just catch up the same way that
9:35CHP did. Did you need the OpenAI brand
9:38gravitas for CHP to emerge? Maybe
9:41because Open AI was not completely
9:43unknown. I mean most people wouldn't
9:44know OpenAI but in the text fair and the
9:47journalistic fair I guess people did
9:49know it. It's only hypothesis. I'm not
9:51sure but um character was still a
9:54slightly kind of a bit complex a bit
9:58gimmicky a bit geeky products and maybe
10:00they missed on on that opportunity
10:02because of that. I don't know. What do
10:03you think Harrison?
10:04>> I think there's some level of kind of
10:05like simplicity that just chat GPT kind
10:08of like brought. It was just one chat
10:10box. You didn't have to Yeah. you didn't
10:11have to select who you were who you're
10:13kind of talking to even now like I mean
10:15this is uh this is again a bit of a
10:16segue but and I'm actually curious how
10:18you guys handle this at Dust but in a
10:20lot of consumer products you choose kind
10:22of like what model you want or or things
10:24like that and I think there is a
10:25simplicity in just like you know just
10:26one chat box type in and you know now
10:28OpenAI has this but at the time when
10:29they launched I think it was probably
10:30just one model that they gave you kind
10:32of like access to so you don't have to
10:34think about that and then I had some
10:36friends at character as well and I
10:38remember when chat GPT came out they
10:39were very worried by that um and and
10:42obviously chat GPT exploded more than
10:44character but it did kind of it was a
10:46little bit of like rising tide lifts all
10:48boats like there was just more people
10:49interested in kind of like chat bots and
10:51seeing what was going on. So I feel like
10:53to some extent it's a consumer product
10:54and so who can really kind of like you
10:56know know exactly why the cultural
10:58zeitgeist catches on to one thing but
11:00once it does it's just you know
11:02explosion.
11:03And so for for you Harrison what were
11:05sort of the precipitating events to
11:07going from the ML ops company you were
11:10at to really deciding hey there's
11:11something to be built here and the thing
11:13that I want to work on is is lang chain.
11:15>> Part of it was my background is just in
11:17kind of like the bu building developer
11:19tools. Part of it is my backgrounds in
11:22ML and I got really excited when I saw
11:24these models and was really like oh like
11:26these are these are kind of like
11:27amazing. So I was still at my my
11:29previous company at the time and I I
11:32knew I was going to leave. I didn't know
11:33what I was going to do but I basically
11:35wanted to explore this area to kind of
11:36like figure out what what I would do
11:37next and my plan was to basically leave
11:39take one to three months to just like
11:40figure out what to do next and then
11:42start working on it. But I was I was
11:43sticking around this asked me to stick
11:45around for a few months to help with the
11:46transition and so in that time period
11:48was going to meetups. That's when I ran
11:49reached out to Stan and uh wanted to
11:53just build some things kind of like to
11:54get my hands kind of like dirty with the
11:56tech. And I remember chatting with uh
11:58Anton who's one of the co-founders of
12:00Chroma, another kind of like developer
12:02tool in the space. He actually remembers
12:04the story better than I do. Uh but
12:06apparently I had kind of like four ideas
12:07which I was talking to him about and and
12:09one of them was Langchain. Uh and then
12:11one of them actually would end up being
12:13Lang Smith which is our commercial
12:14product that we that we build. Um the
12:16difference is lingchain you can build
12:18kind of like nights and weekends. It's
12:19just an open source kind of like Python
12:21package and and so I built that.
12:23Langmith is more of an actual product
12:24and so you know that that came later but
12:26basically just wanted to build stuff to
12:28kind of like get get my hands dirty
12:30released it tweeted about it uh kept on
12:33adding to stuff. a month later chat GPT
12:35comes out keep on doing the same and by
12:37the time I get end up leaving uh my
12:38company in in early January it's pretty
12:41clear that there's you know LMS are
12:43going to transform how applications are
12:45are built and we we think there's a lot
12:48of tooling that needs to be built around
12:49them to make it easy to kind of like
12:50plug them into applications and so kind
12:52of like those two things like LMS are
12:54great but then also we we saw this with
12:56ML stuff and ML ops like there there is
12:58tooling that needs to be built and a lot
13:00of that manifested in in in Langmith and
13:02other tools that we're
13:04It sounds like you had a lot of clarity
13:05around sort of like the early vision
13:07both with with Lang Chain and Lang
13:09Smith. Are there parts of that vision
13:11that have surprised you with how true
13:13they sort of have turned out to be or
13:15parts that maybe you were you know
13:17making a bet on that you're like oh
13:18actually you know the the way the the
13:20industry is developed is actually quite
13:22a bit different than I initially
13:23foresaw. I think like high level like
13:26you know maybe we had some clarity that
13:28we were right on like high level like we
13:30thought LMS would be great and transform
13:32how applications are built and high
13:34level we thought there'd be tooling that
13:35would be needed to build around them but
13:37like that's very high level I think a
13:38lot of the lower level stuff like we
13:40were we were figuring out as as we went
13:42along maybe one thing to that effect
13:44like Lang Smith which is our tooling for
13:45kind of like debugging and and testing
13:47these types of applications we we see it
13:50being used by a whole set of kind of
13:53like uh uh builders. So not just
13:56engineers and not just AI engineers but
13:58product folks and subject matter experts
14:00and everyone coming together and and
14:02when we were building it especially
14:03because of my background in MLOps where
14:05in MLOps the people using these tools
14:06were ML engineers the small segment of
14:09the population you know it it we didn't
14:11I don't think we fully appreciated or
14:12anticipated how many different types of
14:15folks would be involved in the creation
14:16of these types of applications and how
14:18therefore our tool would need to respond
14:20to these different folks. So I think
14:22like you know high level I' I'd say we
14:25were we were you know I think we had
14:27some good insights but like low level
14:29all the details I mean that was like no
14:31way we could have seen all of that that
14:32was just kind of like you know peeling
14:34back layers of the onion as as we went
14:35on.
14:36>> Stan it sounded like you sort of maybe
14:38started with an acknowledgement that
14:39this was you know a very exciting space
14:41but the exact sort of dimensions of dust
14:43have been more uh emergent like what
14:46does that journey look like for you?
14:47like what were the early bets you were
14:48making that yeah have been proven true
14:50and and maybe have been a little
14:51different. Yeah, I think the uh the uh
14:54the early days of Dust is mostly me uh
14:56as as Arison did, mostly me messing
14:59around with a with a product and I think
15:01we uh we really started building the
15:03what Dust is today much later actually.
15:05It's more like early 2023 when my
15:08co-founder Gabriel joins me and there at
15:10this moment we decide to really focus on
15:12applying LMS to internal productivity
15:14and so not building dust was at the
15:18beginning something uh addressing the
15:20same kind of space as lang chain because
15:21the developers was the only person you
15:24could talk to basically in late 2022. uh
15:28but as chibd comes out the uh it uh it
15:30kind of educates the entire market and
15:32know it opens up the opportunity to talk
15:34about LLMs and agent or not agents yet
15:36but LLM and how it can impact work much
15:40more broadly and so as Gabriel Moffer
15:43joins we really decide to focus on
15:45enterprise internal productivity and uh
15:48applying that technology to how people
15:50walk. This is a very long journey that
15:52we are only starting. Uh I think the uh
15:56the the great thing is that there is a
15:58there is so much to be done and the uh
16:01the form factor of what it's going to
16:02mean to work with agents is still pretty
16:05much unknown and so it's been as well a
16:08very long iteration on many stuff. We
16:11started with the idea of with two main
16:14hypothesis at that time. So it's kind of
16:16the rebirth of dust as we uh we
16:18incorporate it and start working
16:20together and I think we start with two
16:21main product hypothesis was first we
16:24need to have the company context because
16:25if we don't have the context we won't be
16:28very useful for doing any actual work
16:30there is a lot of work that can be done
16:31wizard context translating creating
16:33ideas doing research on the internet etc
16:36but for doing actual work at scale
16:37within the company you really want to
16:39have access to the company information
16:42and the second thing was pretty simple
16:43is that I mean the company context is
16:45kind of big. So if you give it to an
16:47agent or to an LLM, it's going to be
16:49lost in it because the retrieval is not
16:50perfect. The models are not perfect. And
16:52so the second hypothesis was really the
16:53ability to create custom agents so that
16:55you could uh pinpoint what context you
16:58really need for a specific task, which
17:00mechanically gives you much better
17:01results. And those two hypothesis were
17:03somewhat early in the market, meaning
17:06that when we were pitching dust in 2023,
17:09everybody was kind of looking us with
17:10weird highs. uh and we've seen the
17:13market open up to those ideas which has
17:15been super exciting. I think the latest
17:17step function in there is probably the
17:19uh Toby Lutker memo which really made us
17:22to some extent go from the early
17:24minority to the uh to the uh to the
17:26early majority. I don't know remember
17:28the the exact terms but we we we went
17:31one step in the adoption curve. So we we
17:34kind of nicely positioned at a place
17:35where the market is going now. But I
17:36think the the one thing is that it's a
17:38constant reinvention of what we are
17:40trying to achieve because uh it's for
17:42sure not the end states and we can dive
17:44into that. We have many uh many
17:46opportunity to see that through
17:48different prisms. I'm sure we will dive
17:50into that and and just for listeners,
17:51you know, I'm sure most folks saw it,
17:53but the Toby memo was, you know, Toby
17:55from Shopify basically saying, you know,
17:57if you're not using AI for your job at
18:00Shopify, you're in big trouble. And so,
18:02you need to really be adopting it in a
18:05very aggressive way.
18:06>> Yeah. Yeah. I think it was it was
18:08slightly more gentle if I'm maybe I'm
18:10wrong, but I think it was don't come at
18:13me for a headcount if you haven't tried
18:14using AI first.
18:16>> Yes. When I say in big trouble, I mean
18:18less uh I'm going to come get you, but
18:20more like you're going to be left
18:21behind. And uh anyway, it was, you know,
18:24I think it was funny. We saw Toby write
18:26that memo and then I feel like there was
18:27this mass nemesis of like every CEO felt
18:31like they wanted to get their letter out
18:32into the public to say like, "Hey, we
18:34also really care about this and we're
18:36hardcore." It was very funny. But um
18:39Toby Toby was early, I think. Um and
18:41yeah, I imagine that was a very
18:42validating moment. you know, we we're
18:44sort of talking about this the framing
18:46of of Langchain and and Dusk, but just
18:49to bring as many people along in this
18:51conversation as possible, like very
18:53crisply, what is the right way to think
18:55about what an AI agent like truly is and
18:58and what it isn't? Uh just so that
19:00folks, you know, have maybe a crisp
19:02wrapper for that that thinking.
19:03Harrison,
19:04>> an agent is an application where an LLM
19:06decides the control flow of of the
19:08application. And so I think that that's
19:12a bit vague. If if you want to get more
19:14technical with it, I think for all
19:16intents and purposes, the way that a lot
19:18of developers think about an agent is
19:20running kind of like a while or for
19:22loop, calling an LLM to decide what to
19:24do and then and then taking those
19:26actions if they're actions and until it
19:28basically decides that it's finished.
19:30And I would say that, you know, like
19:31there the LLM is very clearly deciding
19:34the control flow of the application like
19:36every single step is is decided. I think
19:38you also have applications that are, you
19:41know, not as kind of like maybe agentic
19:45in in some sense where the LLM decides
19:47maybe a few steps, but there's some
19:48steps that are hardcoded like maybe
19:50after A you always do B. Or even when
19:52you talk about multi- aent applications,
19:54maybe you you run one of these loops and
19:56then after that it immediately goes to
19:57this other application which runs a
20:00check or or or another agent even and
20:02then and then goes back. So, I feel like
20:04there's there's this spectrum and and
20:05Andrew Wing has a has a good way of
20:06talking about it that I like, which is
20:08like rather than talking about whether
20:09something is an agent, let's talk about
20:11how agentic it is. And so, that allows
20:14for this kind of like spectrum of of
20:15agenticness. And I would say, yeah, the
20:17more an LLM's deciding what to do, the
20:20more agentic it is. And then probably at
20:22some point there's some threshold where
20:24it crosses from, you know, a non-aggent
20:27to an agent. And to be honest, I don't
20:29100% know exactly where that is, but but
20:33I I I like the idea of agenticness. But
20:35again, for for all intents and purposes,
20:37if you're talking to a developer, I
20:38would largely say they think of it as
20:40running an LM in a four while loop and
20:43having it call tools until it decides
20:45that that's kind of done. Stan, is there
20:47anything you you'd want to add to that?
20:48>> No. Yeah, I think the uh the the
20:50division is is is perfectly on point. I
20:52do like the the agenticness framing.
20:54That's makes a ton of sense. Uh the big
20:56question we see in the market today
20:58which is interesting is agents versus AI
21:00workflows. It's a it's a different way
21:02of looking at the same question. Uh
21:04there's a a lot of there's a lot of
21:06traction on companies that are doing AI
21:08workflows. We all know have heard about
21:10NAN and those those kind of of companies
21:13and I think there's a there's there's a
21:14lot of value in AI workflows. And I used
21:17to use a definition of agents that I
21:18think apply to AI workflows which is any
21:21program where one uh conditional is
21:24driven by an agent. But to some extent
21:27that that agenticness framing is very
21:29interesting. Workflows versus agents is
21:31a massively interesting question uh
21:34which is not answering your question and
21:36that we can revisit but it's a it's a
21:40there's a lot of value in having
21:41workflows because it gives you much more
21:42control on what's happening. uh but uh
21:45to give you a sense I think that
21:47long-term they they don't really make a
21:50ton of sense at the end of the day
21:52because uh we do foresee a world where
21:56uh those agents will be actual
21:57co-workers and I don't think you can
21:59really encode a coworker with a workflow
22:01uh in the same way you cannot encode
22:03code with a workflow we're really
22:05bullish on trying to help people create
22:08agents not workflows very interestingly
22:11which is the weird part the agents is
22:13richer but is more risky in a sense but
22:15it's also easier to build which has a
22:17ton of value in the kind of a work
22:20setting that means that anybody can
22:21build an agent not anybody can build a
22:23workflow uh workflow is like typical
22:26make n uh zap year and stuff it's pretty
22:30easy most people will be able to
22:31interact with those products but not
22:33everyone and it requires a little bit of
22:35a kind of a learning and and and and
22:38learning curve uh compared to that
22:40building an agent is actually just
22:42describing in plain English what you
22:43want to do and giving being clicking the
22:44capabilities of that the agent should
22:46have and it's it it's actually
22:48accessible by a much broader uh audience
22:51which is which makes it also very
22:52exciting and so that's weird that the
22:55long-term most powerful thing is that so
22:57is actually at the same time the easiest
22:59to build to some extent I'm I'm
23:01obviously exaggerating because to build
23:03a great agent you need evals and stuff
23:05like that but to build a proto version
23:06of an agent and it's actually extremely
23:09easy
23:09>> is it is it a fair synthesis and you
23:12guys can can push back on this But, you
23:14know, if we're talking about sort of the
23:15workflow versus agent idea, a workflow
23:17might be more like, you know, writing
23:19out a recipe step by step that you're
23:21sort of, you know, saying, "Here's what
23:23I want you to do." And an agent is more
23:25like creating a chef and saying like,
23:27"Please go cook me something." Is that,
23:29you know, roughly a way that someone
23:30might think about it? Yeah. I mean, it's
23:32it's the difference between McDonald's
23:34and Michelin star restaurants, right?
23:36You know what you're going to get at
23:37McDonald's. It's very well uh
23:39streamlined. And when you're going to a
23:41Michelin star restaurants, you don't
23:42know you don't know what you're going to
23:44get because you'll be at the the chef
23:46will be improvising with the situation
23:48and the the food that has been
23:50available. It's it's a again a funny way
23:53to frame that. Harrison, I saw your eyes
23:56Yeah. sort of light up with a with a
23:58question there. I mean, yeah, maybe
24:00pushing back on this a little bit. Like
24:02I feel like I to to
24:05your point, Stan, I feel like you can
24:07you can often do the same things with
24:09workflows and agents. It's just the ease
24:11of how you describe it. Like it in an
24:13agent, it's it would all be in natural
24:14language, right? Like you could have a
24:15recipe that you just put in natural
24:16language. You say, "Hey, do A, then do
24:18B, then do C." And, you know, it's not
24:20as deterministic, so it's not kind of
24:22like as safe, but it's way easier. And
24:24so I I I I'm thinking out loud of this
24:27analogy for the first time. So, so I
24:29don't like I don't have super strong
24:30opinions on, but I do think I don't know
24:32if it's so much different things. It's
24:33like different ways of accomplishing the
24:35same thing. And so maybe it's uh I don't
24:38know the difference of a person in
24:41McDonald's who's doing it all by hand
24:44versus use I using some of the
24:46pre-portioned things. I I I don't know
24:48exactly where I'm going with this, but I
24:49do think like that I I just I really
24:51agree with what Stan said where like
24:53there is a beautiful kind of like
24:54simplicity and how easy it is to to
24:57build the agents. It's it's usually
24:58natural language and then you choose
25:00some tools and I remember one of my
25:02favorite releases we ever did in
25:04Langchain. It must have been like the
25:0513th release or something like that
25:06where we took this idea of like the the
25:09React agent which is this great paper by
25:11by Shenu and like was a little bit
25:13focused and the examples they had in the
25:15paper were very focused on some like uh
25:17hot pot QA some like Wikipedia question
25:19answering like a narrow task. But I
25:21remember we took this and we made a we
25:22made an abstraction around it where you
25:24just did exactly what Stan said. you
25:26gave it some tools, you gave it some you
25:28gave it up a a system prompt and it
25:30would just do things and I was like holy
25:31crap this is amazing and I think like
25:33that is like that there is this
25:34beautiful simplicity in agents and so
25:37yeah
25:38>> and very interesting data point here is
25:40that the uh the react paper which I
25:42invite everybody to read and everybody
25:45who will read that paper today will look
25:46at it and say what the is that it's
25:48so completely obvious what is it even a
25:51paper seriously and it's funny how at
25:54the time it was kind a mindopening and
25:57people we were we were so early in that
25:59technology that it was kind of a really
26:02uh interestingly mind-opening paper
26:04despite retrospectively it feeling
26:06extremely trivial and obvious with
26:09everything that's been built since then
26:10and so that's really funny very funny
26:12exercise to all listeners to open the
26:15react paper and skim through it quickly
26:17that will show you and it was a great
26:19paper kind of a really engaging paper uh
26:22in late 2022 and that's how early we
26:24were that's So I said so no clue we had
26:27about what we can do with
26:28>> and maybe to that point like when when
26:30we launched it and even right now in
26:31lingchain I think we still call it like
26:33the react agent but we're going to stop
26:36doing that because it's just confusing
26:37to people. They're just like what do you
26:38like what is react like this is just so
26:40obvious it's just run it's like you know
26:42just taken for granted now.
26:43>> That's amazing. um you know to to maybe
26:46get a little bit more tactical with it
26:48like what are the the different use
26:50cases you guys see for agents at the
26:52moment that are sort of most productive
26:54and you know Dust is obviously creating
26:57uh a lot of those sort of for companies
26:59you know deploying them uh but yeah
27:01curious like even within that where
27:03you're seeing the most leverage for for
27:05businesses
27:06>> the list is long uh we we we're taking
27:09the we have a pretty horizontal I mean
27:11we have a completely horizontal product
27:13because we believe that it can be
27:14applied in so many places and that there
27:15is value of having one platform for
27:17everybody to share creating and sharing
27:18agents having agents interacting with
27:20others in a business setup. I think that
27:23I can only give you uh examples. It goes
27:26from um extremely simple slack thread to
27:30issue creation and that's not completely
27:32trivial because at dust we have
27:35something like three or four different
27:36type of issues which goes in different
27:38types of projects given the different
27:41shape or or type of discussion that
27:44needs to happen and having an agent that
27:46help you go from a thread a thread slack
27:48where there's a discussion where maybe
27:50there's a bug that is emerged there's a
27:51place obvious where to put it when when
27:53that's a bug but when it's a decision
27:55that needs to be taken, it needs to be
27:56moved into a different place and and
27:58taking a few actions around that. And so
28:00streamlining that process of having
28:03something happening organically on Slack
28:05as an example to following a workflow
28:08that makes it uh represented in the
28:11canonical way that a company represents
28:13that artifact that is being discussed on
28:15Slack is a very general use case that I
28:19covered for issues on GitHub, but that
28:20it can be done for many other stuff.
28:22Whenever there's a sales transcript, uh
28:26there's a an army of agents that gets
28:28kicked in uh for providing feedback, for
28:31auto fire, filling in uh Salesforce, for
28:34extracting product interest and going to
28:37create comments on notion pages related
28:39to the product that has been discussed
28:41with during that transcript. And so
28:43there is so much that uh uh would requ
28:46so much stuff that uh would require
28:48human work that is now being uh uh
28:51doable by agents that it's really
28:53interesting. The most interesting places
28:54is for the things that no human would
28:57ever do uh because it's just too much
28:59work. So for any sales transcript being
29:02able to put a comment in the right uh
29:05product document uh on notion about the
29:08fact that it's been discussed the link
29:09to the transcript and the kind of a one
29:11sentence summary of what has been said
29:13is something that never happened before
29:15because nobody was there to make it
29:17happen nobody had time to review all the
29:18transcripts and so those new use cases
29:20are almost the most the most exciting as
29:22well to me. Harrison, I curious, you
29:25know, maybe how you guys use agents at
29:27Langchain internally and maybe some of
29:30the use cases you've seen that you found
29:31particularly powerful.
29:32>> Yeah, so internally um we use agents in
29:39a few ways. So we so and they kind of
29:43map with like the the big use cases that
29:45we see out there as well. So like we see
29:47customer support being a big use case.
29:49Uh we've built slash are building an
29:51internal kind of like customer support
29:53agent to help us with a lot of those
29:54inquiries and responses. Um coding is a
29:58massive use case. We both use and build
30:01some kind of like internal uh coding or
30:03coding adjacent agents for kind of like
30:05responding to issues, things like that,
30:07managing discussions. I personally use
30:10an agent to kind of like monitor my
30:11email and uh draft responses and flag
30:14things and and and so that's that's
30:16probably the one that I use the most. I
30:18mean, we use uh off-the-shelf and
30:21internal kind of like versions of deep
30:22research agents. Like, that's been a
30:24huge kind of like style of things that
30:25we've seen pop up. Oh, we use some for
30:27marketing as well. Like marketing is a
30:29fantastic use case. And so, we use some
30:30for translating some of the uh blogs or
30:33things we do into tweets or LinkedIn
30:35posts. And so, I think I think those are
30:37a lot of the big use cases that that we
30:38use internally. And I think they
30:40generally match up with what we see
30:41people building in in the industry.
30:43you've written a I think it was a blog
30:45post where you've talked about you know
30:46how we interact with agents and how that
30:49might change from sort of uh typing uh
30:52you know prompting these agents uh with
30:54voice or text uh and shifting towards
30:57more of what you call like ambient
30:58agents that you know uh require less of
31:01that. Why do you see things going in
31:03that direction and and maybe you can
31:04share a little bit more about you know
31:05what an ambient agent really looks like?
31:07>> Yeah, absolutely. and and and I'm I'm
31:09super curious to hear what Sam has to
31:11say on this as well because he I think
31:12he mentioned something about uh you know
31:14thinking about how we interact with
31:16agents as part of the core mission of
31:17Dustin. That's one of the things that
31:18I'm a little bit jealous that we don't
31:20get to do a lot of link chain because we
31:22are more developer facing so we do think
31:24about it um and it does absolutely
31:25inform like you know what tools we build
31:27but but not nearly as much as folks
31:29building um products in the space do.
31:31Um, but I mean so far the dominant UX
31:34for agents has kind of like been chat.
31:36And I think and I think and I think if
31:38you think about it from first principles
31:39that actually does make a lot of sense
31:41like it puts the human in control. It's
31:42very human in the loop. Uh you can you
31:45can not only does the human initiate it
31:47but the human can see what's going on
31:48because you can stream back results. If
31:50the agent wants to do an action, you can
31:52have the human kind of like immediately
31:53approve it there. So you can kind of
31:55like have uh some sort of like approval
31:57for for dangerous actions. and and it's
32:00relatively fast for the most part, but I
32:03think like you know people have been
32:05saying that that chat won't be the only
32:07UX or or you know the the forever UX and
32:10and while it has actually lasted a lot
32:13longer than maybe people saying that
32:14would have initially thought I I do I I
32:16do think it's interesting to think of
32:17what what besides chat is out there and
32:20and also like what some of the downsides
32:21of chat are. Um, and I think some of the
32:24downsides are also some of the things
32:25that that that make it good in some
32:27cases. Namely, like you have to kick off
32:29all conversations. So, if you want to
32:31run it over kind of like a thousand or
32:3310,000 things, like that's that's a
32:35little bit tedious to kind of do. And
32:37and also because you generally expect to
32:39be in the moment, they can't really take
32:41that long. Um, otherwise you get a
32:43little bit bored and maybe switch. And I
32:45actually want to come back to that point
32:46because I think with like deep research
32:47and some of the coding agents, we're
32:48starting to see some that like that
32:50starting to happen. But it's still
32:51anyways I'll I'll come back to that. And
32:52then the other thing is like yeah I mean
32:55ju just based on inside an enterprise or
32:57inside a company you have all these
32:59events happening and rather than like
33:00copy paste an email and then take that
33:02email and go and and you know put it in
33:04chat. Wouldn't it be nice if that would
33:06just kick something off automatically?
33:07And so I think I think the email
33:08assistant that I use is actually a great
33:09example of this. It just monitors my
33:11email inbox. It just gets triggered by
33:12these events. Um, and then it goes and
33:15does something and if it wants to take a
33:16an action that I that I deem kind of
33:18like dangerous enough where I want to
33:20approve it, which right now is is
33:21scheduling a calendar invite or or
33:23responding to the email, then it
33:24presents it to me in some way. And I
33:26think that like what is that UX? What
33:27does that look like? Is that just a
33:28draft in my email inbox? Is I I we have
33:31a concept of like an agent inbox, which
33:33is a dedicated view for this. Back to
33:34the original question like ambient
33:36agents we define as agents that that
33:37listen to a stream of events and then
33:39act on one or multiple at the same time.
33:41And and I think crucially these are not
33:42necessarily autonomous agents. They're
33:44still kind of like have some human in
33:46the loop at some component because I
33:47think that's still necessary for for
33:48enterprise adoption. I I have some other
33:51thoughts on on kind of like the deep
33:53research and coding style agents, but
33:54I'd actually be curious to hear Stan's
33:56thoughts first because you work on
33:58actually delivering these to a lot of
34:00users.
34:00>> Uh totally. I think uh basically the the
34:03way we we see it indeed is that the
34:05conversation interface has been the the
34:06the interface. um we we we hypothesize
34:11and uh maybe that's never been the case
34:12because indeed it survived for a long
34:14time that there's going to be a fork
34:15probably in the typical UX UI that means
34:19working with agents as deep research
34:21agents take longer and longer as agents
34:24are being triggered with human out of
34:26the loop uh you probably want something
34:28that looks like more like a command
34:30center than a list of of conversation
34:32the conversation paradigm will probably
34:34make sense on the in the B2C setup for a
34:36much longer time Because in the middle
34:38SE, the the truth is that your agent is
34:40your executive assistant. And so you
34:42have one stream of conversation with
34:44them or a couple different streams. But
34:47when you think about the enterprise,
34:48there's going to be agents that take a
34:50lot of time. Agents uh conversation with
34:52agents that involve multiple people.
34:55That's something that the market hasn't
34:57even started exploring much, right? I
34:59mean we know we don't explore it much
35:01yet because I think we consider them as
35:04toy and but the more powerful the use
35:07case will be the more meaningful it will
35:09it will be for people to actually
35:10interact interact multiple people
35:12interacting with the same with an agent
35:14or with multiple agents completely align
35:16with your vision uh Harrison of the
35:18ambient agents. Uh I think the the first
35:21step as you described it is really to uh
35:24have more of a ninbox paradigm when you
35:27interact with those agents having agents
35:29that are being triggered by external
35:30events. Uh, and I think the uh the crazy
35:34idea is agents that are not necessarily
35:37mechanically triggered in the sense that
35:38when these do that or when these
35:41actually execute, but having agents that
35:43are just skimming through what's
35:44happening inside of the company and
35:46maybe ping you with offers to give you
35:48to provide you some value, which uh is
35:50uh obviously what's the end state should
35:53be. Um when you think about all multi-
35:55aent systems I mean today we obviously
35:57don't see that in the enterprise but you
35:59could imagine giving a giving a very
36:01high level project to a set of agents
36:03and just let them walk and and organize
36:08for delivering the project on their own
36:10and give you a report multiple days
36:12after. I think all of that is completely
36:14uncharted territory. uh and so but yet
36:17it's obviously the end state and so
36:19that's why it's so important and so
36:20exciting to be uh working in that uh in
36:23that space.
36:23>> You mentioned a command center like is
36:26that in the product now or is that a
36:27future kind of like direction because I
36:29love that idea but I would love to see
36:31like yeah I would love to I would love
36:32to see what that I don't know what that
36:33looks like. I would love to see what
36:34that looks like.
36:35>> No, I don't know what that looks like
36:36either. I think the the first step will
36:38obviously be like what you've been
36:39building uh internally and I think uh
36:41you've you've shared on some of your
36:42blog post but then what you just
36:44mentioned a form of inbox is just is
36:46already a first step in that direction.
36:48Obviously, the weird thing about the
36:49agents is that there they also I mean
36:51the APIs are so uh uh biased toward a
36:55conversation
36:57that often you're like do something for
37:01me and you have that agentic loop that
37:02can last for a very long time but at the
37:04end of the day you still have an agent
37:06message which is kind of weird because
37:08that means that you somewhat have a cap
37:10on the interaction through the agentic
37:12loop. You can make very them very long,
37:14but you're also uh you're going to be
37:16exhausting your context at some point,
37:17but that's not necessarily completely an
37:19issue. But at the end, the whole system
37:21is kind of is kind of a post train to
37:23give you an answer. Yet, when you think
37:27about agents walking, you just want to
37:29say go do the work for a day and ask me
37:32questions if you have any, but don't
37:35give me an answer in 30 minutes. I just
37:37want something delivered in one day and
37:39ask me questions if you have anything.
37:40Yet there's no good system for
37:43interacting with the current shape of
37:45those agents for doing that. You could
37:47think about multi- aent stuff etc. But
37:49the ecosystem is not there yet I guess
37:51and so we still even at the API and post
37:54training level a little bit bound to be
37:56staying close to the conversational
37:58interface but I'm sure that we'll see
38:00kind of stuff emerge around that.
38:02>> When we were building at the start they
38:03like messages weren't a thing. it was
38:05just text in text out and then like
38:07OpenAI uh I think it was it might have
38:10been 3.5
38:12or maybe four that they released and it
38:14was only the chat message API and I
38:16remember talking with someone from
38:17OpenAI it's like so are you going to
38:19release like the non- chat message thing
38:20and they're like h we don't know and
38:22they ended up not and now everything is
38:23just like that and that kind of happens
38:25two what's also really annoying about
38:26this is there isn't really like a
38:29official schema for what messages is
38:31like open AI has their kind of like
38:32input output schema but that's different
38:34from anthropic which is different from
38:35Google's and like you'd think if this
38:37was like this is like you know the base
38:39thing which has how we interact with
38:41these I you know I wish it was a little
38:43bit more standardized what that schema
38:45is although it's constantly evolving as
38:47well so tough to always tough to do that
38:49um and then the third point is maybe
38:51like so all the chat agents to date have
38:53been kind of just like synchronous
38:55agents and that like you just chat in
38:56the moment now with deep research you
38:58have things that start sync you start
39:00with a chat but then you go to this deep
39:02research and then maybe you actually
39:03come back to synchronous at the end. And
39:05I think for some of these ambient
39:06agents, you could almost view them as
39:07async running in the background, but
39:09then at some point, like you said, Sam,
39:11they ping you with something and then
39:12they become synchronous. And so I I
39:14think chat is a pretty good like form of
39:16synchronous communication. And then like
39:18you know what what async means is like
39:20maybe that's hidden a little bit through
39:22some like context or prompt engineering
39:23and it's just like by the time it
39:25surfaces to the user, it's just all a
39:27message because that's the dominant form
39:28for synchronous communication at least.
39:31This episode is brought to you by
39:33[music] Brex. Fred Adler, the
39:35influential venture capitalist of the
39:371970s, was known for displaying
39:39decorative pillows in his office that
39:41featured a signature business
39:43philosophy. Corporate happiness is
39:46positive [music] cash flow. In today's
39:48poster environment, Adler's wisdom feels
39:51particularly relevant as founders need
39:53to make every dollar [music] work
39:54harder. That's exactly what Brex
39:57delivers. Their modern finance platform
39:59was built specifically for startups
40:01[music] like yours and designed to help
40:03extend your runway when capital
40:05efficiency matters most. With Brex, you
40:08get global corporate cards with up to
40:1020x higher credit limits and no personal
40:12guarantee [music] required. Their
40:14banking solution has no minimums and no
40:17transaction fees while letting you earn
40:19high yield from day one with sameday
40:21liquidity. Best of all, Brex knows you
40:24were born to build, not juggle
40:26spreadsheets and finance tools. Their AI
40:29powered platform [music] brings cards,
40:31banking, expense management, and travel
40:33all in one place. It's simple, scalable,
40:37and designed to get you back to what you
40:39do [music] best, building. More than
40:4230,000 companies, including one in three
40:45US venture-backed startups, trust BS to
40:48help make every dollar count toward
40:49[music] their mission. join them at
40:52brex.com/mario.
40:54also that it sounds like there's just
40:56this layer of pro uh proactivity that
40:59you know you're suggesting might be
41:00different that you're sort of saying
41:01here are the the goals that we have as a
41:03business and actually as long as you're
41:05sort of have the sufficient context and
41:07you know enough power you can start to
41:09say hey by the way uh you should
41:11consider doing something as little as
41:13writing this tweet to boost the numbers
41:16that you want to do or to something as
41:18big as you should consider you know this
41:20new product that might you know be be
41:22really important over the next few
41:23years. What are the sort of major
41:25limitations to a more ambient model
41:28today? Like how far do we are we from a
41:32true command center world where you know
41:34maybe you're really ushering out a swarm
41:36of agents per person uh and having them
41:39sort of monitor and think for you and
41:41and do this deep asynchronous work on a
41:43regular basis. I
41:45>> I think reliability is obviously a
41:47limitation. It's it's still mindblowing
41:49to me how dumb those agents can be in
41:52pretty obvious situation and yet
41:56f star get a nao gold medal at the same
42:00time. It tells a long story about about
42:03the importance of data, the importance
42:05of pre-training, the importance of post
42:07training and how there's been focus on
42:09code on mass a lot and yet on different
42:12places there is some obviously some some
42:15uh some gains as well but so many cases
42:18where they're like damn you're being so
42:21silly there you can solve very complex
42:22math problem and yet you don't
42:24understand from the context that it two
42:26women speaking together what not anyway
42:28so I think this it is the main blocker
42:30And so as an example like I wanted to
42:32share that I think a way to a different
42:34way to think about working with agents
42:36is that could be a transition towards
42:38the very long ambient agents is the
42:42concept that I really like which is the
42:43concept of a walk plan and if you ask me
42:45I don't understand why linear and all
42:47the kind of a stuff isn't doing that
42:49aggressively but you can you can imagine
42:51a work plan you have a very high level
42:52task and you start splitting it in
42:54smaller task and you start splitting the
42:55st in smallest task and once you start
42:57doing that work you can do it assisted
42:58by an agent you can do it yourself and
43:00then you can start delegating or
43:02discharging those tasks to agents and
43:04those agents start working on the task
43:06come back to you and like no not quite
43:07yet and all eventually you start
43:10clicking the task that being done maybe
43:11it's you maybe it's another human maybe
43:13it's an agent maybe it's a bunch of
43:14agents and so there you kind of have a
43:17nice mesh between the conversation and
43:19the kind of ambient agent it's not at
43:21all ambient to begin with because you
43:24kind of de develop the work plan and
43:26discharge to agents as you as you go but
43:28the better the agents are the more
43:31they'll be taking of that walk plan task
43:33all the way to eventually maybe someday
43:35defining the work plan speaking the task
43:37discharging to other agents etc and so I
43:40think the um there's there's even in in
43:42the current world where we have very
43:44deep limitation in the reliability of
43:46agents on some task which make uh the
43:49presence of a human uh to monitor what's
43:51what's going on kind of very important I
43:54think there's many products surface we
43:55can imagine that will start to you know
43:58mesh between the sync uh interactions
44:02all the way to more async uh through the
44:05ability to probably introspect what has
44:07been happening. Does that sh with you
44:09what you think Harrison is the major
44:10blockers?
44:11>> Yeah, I I I would agree with that. I
44:12mean I think like reliability of
44:14individual agents. I then think there's
44:16a lot of work to be done at kind of like
44:18the UX layer. Um and then I'd also say
44:21like and this kind of gets to the
44:22reliability aspect but just like
44:24learning slashmemory is also interesting
44:26as well. like there needs to be some
44:28like that's what we as humans do. And so
44:30that's maybe a little bit further out.
44:32But I do think that's a component. I
44:34also think like I think code often leads
44:36this space just because the models are
44:38really good at it. And so I think if you
44:40look at like claude code like that's a
44:42great example where the model got good
44:44enough. Okay. So reliability is a little
44:47bit better. They did a good job of
44:48writing up a a CLI and giving it access
44:51to some tools. So some great context
44:52engineering there. And then you start to
44:54see like some interesting UX things
44:55happen. So I think I think there's an
44:57open source project called like
44:58taskmaster or something that like you
45:00know keeps an eye on like five or six
45:01cloud code things that that happen kind
45:03of like in the background. I think they
45:05released a view to kind of like see kind
45:07of like usage of cloud code and and and
45:08and Chip Huan released something as a
45:10way to debug like the errors that cloud
45:12code kind of like made. So like now that
45:14we get these like more like I think
45:16that's the first example or that's a
45:18yeah one of the first examples of these
45:20really like longunning kind of like more
45:22autonomous agents and now you're
45:23starting to see a bunch of like
45:25interesting kind of like command centery
45:27type vibe things coming out for how to
45:29interact with them but it's still it's
45:32still really early on but I I I like to
45:34look to uh to code for an example of
45:37like where where the space is generally
45:39headed just because I think it's ahead
45:40of of the other verticals given how Dust
45:43operates. I imagine you have a a a
45:46specific opinion on this, but when you
45:48think about how agents play out over the
45:49next few years, do you think that like
45:51it's unequivocal that there's going to
45:53be really many many specialized agents
45:56or over time do we just sort of start to
45:58converge into, you know, a super agent
46:01that has enough context on work and life
46:04or whatever it is? Yeah. Are we are we
46:06heading towards a a true multitude or
46:10you know an oligarchy or or or or one
46:12true true ruler?
46:13>> That's the big question. Uh we don't
46:16have a clear answer on that and we're
46:17trying to uh stay very uh uh uh humble
46:21with respect to that question. We
46:23started with many custom agents and it
46:25was a clear good decision at the time
46:27given the state of the models. As the
46:30models are getting better, there is an
46:31undeniable force towards uh higher level
46:35agents until the agent is is is doesn't
46:38have a really functional memory uh so
46:41that it can interact with humans, learn
46:43from them, being coached and
46:44understanding how the company operates,
46:47I think you're still going to have the
46:48need for custom agents because if the
46:51agent doesn't have a good episodic
46:53memory in a sense, it's going to be very
46:55hard from them to learn that this data
46:58is routed And this data is fresh. I mean
47:01uh in every company you have data that
47:04is not up to date and data data that is
47:06good and data that is bad and you have
47:08ways of doing stuff etc. And so today
47:12having custom agents lets you point to
47:14the right data explain the right process
47:17so that you don't have to do it each
47:18time. I think the state of memory of
47:21agents doesn't scale us to a point where
47:24you could have just one agent and it's
47:25going to learn it all. And also kind of
47:29feels weird that you would have to teach
47:32your agents.
47:33And there's also weird stuff in a in a
47:37when a company gets somewhat big, you
47:39even have contradictory ways of doing
47:41stuff within the company. Team A will do
47:45stuff this way and team B will do stuff
47:48this way. And so now it begs the
47:49question, where's the memory? because
47:51teams will be competing for the same
47:53memory slots in the sense of doing the
47:56thing the right way. So there's still
47:57many questions even if you assume a
47:59really great perfect agents there's
48:01still a ton of a ton of questions to be
48:04answer your question I don't know if
48:06it's the end state I think the uh level
48:08of abstraction of the agents uh in
48:10general will increase and so the number
48:12of agents being necessary to walk and to
48:15do walk will probably decrease there
48:18will be probably a convergence toward
48:19one but it's very unclear when that's
48:21going to happen and I think we're trying
48:22to keep our our finger on on that that
48:25trend
48:26>> so you do think eventually there'll be a
48:27convergence towards one workplace agent.
48:30>> No, I'm sorry. I'm saying that there's
48:32going to be a an increase in the
48:34abstraction level of agents and so a
48:36decrease in their number. I don't know
48:37if it's going to converge towards one. I
48:39guess I guess we'll converging, but
48:42maybe there's going to be a sailing
48:43>> 10 versus 100 or whatever it might be.
48:46>> Exactly. Yeah.
48:46>> Um does do you take the same position,
48:49Harrison, or do you see things a little
48:50differently?
48:52>> No, I think I largely agree. Maybe like
48:55a few kind of like you know thoughts as
48:58well like one like generally like what
49:00does it even mean to have like mult like
49:02h what are like what does it mean to
49:04have multiple agents like how are they
49:05different and generally it's it's the
49:07prompt it's maybe the model but mostly
49:09like the prompt and the tools it kind of
49:10like has access to and so sure in the in
49:13the limit you could maybe have like you
49:15know one agent with every single
49:17instruction for how to do everything at
49:18the company and the system prompt and
49:20all the tools there under under the sun.
49:23That's definitely not what we see right
49:25now. Maybe it will go towards that
49:26direction or towards a smaller number of
49:28agents. I think what we see more now and
49:31is maybe an alternate view is like there
49:33will be one agent that a user at a
49:36company interacts with but under the
49:37hood there will be many sub agents that
49:39it can call out to or route to or use to
49:41and and those have like the specific
49:43instruction that you know like when we
49:45talk about people for building agents
49:47it's like write down a standard
49:48operating procedure and figure out what
49:49tools it needs and then that's your
49:50agent. So maybe there'll be like you
49:52know one kind of like central supervisor
49:54agent that can interact with all these
49:56other agents either by and and now we
49:58get start to get into multi-agent stuff
50:00and that's very very kind of like early
50:02on I would say but there are some
50:03initial ideas of how to do that. I mean
50:04even if you look at some of the coding
50:05agents going back to kind of like
50:06looking at code uh Google jewels is kind
50:09of interesting. It has kind of like this
50:11chatbased synchronous agent that kicks
50:13off other async kind of like background
50:14agents. And I'm assuming there's some
50:16difference in the system prompts and
50:17tools that has access to something like
50:19that. I think is very very reasonable
50:20and most people even right now are kind
50:22of building choice because people don't
50:23want to have to choose like oh I have
50:25hundreds of agents. I no they just want
50:27a chatbot. It's this it's a simple kind
50:29of like you know approach to that. But
50:30under the under the hood at least right
50:32now and and even for the foreseeable
50:34future I think they'll still be like
50:36relatively specialized. It's like
50:38having, you know, your agent that is
50:39your VP of marketing who's also managing
50:42the agent for your social media
50:43marketing, your performance marketing,
50:44your brand marketing, and you don't have
50:46to worry about, hey, I'm trying to
50:48figure out, you know, how to do my
50:49performance marketing. Here's the exact
50:51one I have to go to sort of thing.
50:52>> I think that's exactly right. I think
50:55like um one thing that we do, and I
50:58genuinely don't know if this is good or
51:00bad, but I feel like we often like
51:02anthropomorphize how we interact with
51:04these agents. On one hand it might be
51:06good because like yeah that's how we are
51:08used to communicating and that maps to
51:09our mental model and that's a good like
51:11all these like context engineering which
51:13is like you know the the topic of the
51:14month is just communication right but on
51:16the other hand like these things are
51:19different than us so like why should the
51:20way that we communicate be the way to so
51:22like I genuinely don't know if it's good
51:23or bad but like the analogy you just
51:25made like I think that's what we often
51:27do and other builders often do to try to
51:29figure out what the best way to organize
51:30and communicate across these agents are
51:32again for better or worse. Here's
51:35a question I have, you know, that that
51:37maybe goes beyond the paradigm of the
51:38agent, which is how can we make sure uh
51:42the agents and AI in general is doing
51:44properly useful work when we see like so
51:47much of this sophantic posture from a
51:50lot of the responses like you know
51:52something I worry about when we see you
51:54know an a company full of agents is like
51:57will you really have a VP marketing VPN
52:00you know whoever who's really able to
52:02think critically about this when there
52:04is just this you know reflexiveness that
52:06that is so so pleasing like do you see a
52:09solution to that that question anytime
52:11soon? I I don't I don't have a solution
52:14to to offer like that. But I feel I mean
52:17my one of the
52:20small side research project I'd love to
52:22work on. I don't have the time but if if
52:24I if I had time I would play with that
52:26is probably to um know try to uh have
52:30agents debating against each other
52:32towards a goal like adversarial not
52:34necessarily adversarial maybe more like
52:36a research community. they they share
52:38results and then you have something like
52:40a hacker news hacker news system where
52:43the things that are the most cited go up
52:44and it's a clear objective of the agents
52:46to get ranked high and they try to push
52:49towards getting some answers with that
52:51by trying to follow some form of notion
52:53of truth which is obviously a whole I
52:55mean you have no no guarantee that they
52:57they would do it or or whatn not but I
52:59think there's in the multi- aent setup
53:01there's probably a new dimension that it
53:03creates that can probably uh alleviate
53:06that that that that problem because
53:08you'll have an opportunity to prompt
53:10agents to be actually a little bit
53:12adversarial to other agents providing a
53:14very as you as you mentioned very
53:16reflexive response that try to please
53:17the user and so I think I think there's
53:20there's a lot of stuff to be explored
53:22there obviously we are light years away
53:25from productizing this kind of stuff
53:27we're light years away from producting
53:28this kind of stuff because uh the the
53:31state of the market is light years away
53:33from uh even those kind of questions
53:36which is which is interesting but I
53:37think there's a lot of stuff to explore
53:38in that direction.
53:40>> Harrison, anything that comes to mind
53:41for you there?
53:42>> Yeah, prompting these agents to have
53:43different points of view is practically
53:44speaking and what I think is is feasible
53:46now and then I also imagine some of kind
53:48of like these, you know, uh issues will
53:51get handled through better models that
53:53just come out from the foundation
53:54models. Are you seeing folks do some of
53:56that prompting to do some of those sort
53:58of like uh prompting different views and
54:00and sort of you know strapping that
54:02together to have like a the hacker news
54:04style ranking or whatever it might be?
54:06>> I mean I I don't know there's probably a
54:07ton of teams all over the world working
54:09on those kind of ideas. Open a obviously
54:12has a a multi- aent team. It seems like
54:14the IMO result comes from the multi-
54:16aent teams. They say they have a special
54:17model. If you ask me I would say it's
54:19probably a multi- aent setup but they do
54:20exactly those kind of And so uh I
54:23think many people are exploring for sure
54:25and that that makes sense but it's still
54:27definitely in the realm of research at
54:29this stage. I I would say
54:31>> I think we see like very like um simple
54:35and naive versions of that or a simple
54:37version of this is just like reflection
54:39or critique on an initial thing. And so
54:40like I think like a a pattern that we
54:44sometimes see is like yeah have one
54:45agent or one LLM generate something and
54:47then and then give some feedback on that
54:49through whatever. I mean this is this
54:50kind of gets into like some of the
54:52reward systems that actually go into RL
54:54and so like for code it's kind of easy
54:55you can run the code so you actually
54:57don't even need another agent to provide
54:58this kind of like other point of view
55:00right you could almost view it as like
55:02hey this is the systems point of view
55:03where your code doesn't compile that's
55:05just like a fact right um but I think
55:07you can imagine doing stuff like this
55:09for you know essay writing or something
55:11like that where uh you you have you have
55:13kind of like one agent that you know
55:15reviews it or or or gives some feedback.
55:18I think right now it's a little bit more
55:21researchy unless you have kind of like
55:23these verifiable kind of like rewards
55:26almost that you can feed back into the
55:27agent as it's running like eval in the
55:29loop is kind of like what what what we
55:31call them and so you can add these
55:32checks from yeah like running code is
55:34kind of like the most obvious example.
55:36We're working on an internal uh coding
55:39agent and as part of that we're
55:41experimenting with having kind of like a
55:42separate agent that kind of like decides
55:45whether it's you know done with a loop
55:47and that's a little bit different. It's
55:48not like as adversarial but it's kind of
55:50just like delegations of concerns almost
55:53or separations of concerns. Um, and so I
55:57I think you could I think you can view
55:58some of the some of the some of the
56:01stuff that people do in this vein, but
56:03it's very kind of like brute force in
56:05some way or like simplistic. I'd love
56:08to, you know, zoom out for a moment and
56:10talk also just about what it means and,
56:13you know, what it's like to be building
56:14in AI at the moment and some of the
56:16specific dynamics that founders have to
56:18face. And I think one of them is really
56:21uh just how fast the fast following is
56:24happening. You know, you see any good
56:26idea, uh, there will be three, four,
56:29five, however many, uh, folks, you know,
56:32chasing that that very, very very, very
56:35quickly and able to raise, you know,
56:37considerable amounts of money as you've
56:39gone about building your businesses like
56:41how do you think about, you know,
56:43protecting against that, building in
56:44defensibility where you, you know, you
56:46think you have a real chance to build a
56:48moat. Yeah. Stan, how have you thought
56:50through that at dust? I mean so building
56:52an AI is a cluster of that's for
56:54sure. [laughter]
56:57So basically for the past uh many
56:59decades the technological substrate has
57:02been extremely stable. Uh for the SAS
57:04let's say for the SAS uh decades that
57:07were behind us it was JavaScript and
57:09posgress. I'm exaggerating a bit again
57:11but uh extremely stable technological
57:14substrate. So when you could when you
57:15were building something, you knew that
57:18the foundations were not moving. So you
57:19could describe where you were going. You
57:21could build an image of your vision. So
57:23you had the vision and you had what you
57:24wanted to build to to to to realize that
57:27vision. And today we one one very
57:30specific thing that you faced as a
57:32founder building an AI is that you have
57:35what I call the fog of AI. And the fog
57:38of AI is the fact that the foundations
57:40are moving very quickly. And so you have
57:43to you have to have a vision of where
57:45you're going but you cannot paint it
57:47because if you paint beyond 6 months
57:51whatever you're painting will probably
57:53be shattered by the foundation shifting
57:55towards a different uh I mean the
57:57spaceime of the ecosystem shaping itself
58:00in a different ways and whatever you
58:02were painting will probably not be not
58:03true. So that's we that means that you
58:05have that fog of AI at six months which
58:07is very interestingly very problematic
58:10for like building a a high efficiency
58:12organization I I find because alignment
58:15is one of the thing that makes
58:16organization extremely efficient and
58:18here you you don't have that kind of
58:20continuity between the current products
58:22and the vision. You cannot paint that
58:25continuously because you have that fog
58:27barrier at six months which means that
58:28you have to jump from the road map for
58:31the next six months to the vision and
58:33that makes alignment of the team a
58:36challenge that is interesting in the way
58:37people think about where the product
58:39will be how they prioritize stuff. You
58:41want all of that to be as autonomous as
58:43possible and that makes it a really
58:45strong difficulty compared to what I've
58:47seen. I've I've been lucky to be at
58:48Stripe and at Stripe we had a very clear
58:51uh alignment because it was a simple
58:53developercentric product to an API and
58:56so you had that very strong alignment
58:58internally that allowed for the
58:59organization to grow and be efficient uh
59:02without a lot of processes.
59:03>> This is the lesser discussed AI
59:05alignment problem.
59:06>> Yeah, I find that I find that I find
59:08that one of the most challenging part of
59:09building an AI
59:10>> I bet. Yeah. Uh and so how have you
59:13thought about building the defensibility
59:15piece for Dust? Like where
59:16>> Oh yeah. So sorry as uh I mean we've
59:19we've we've managed to build a product
59:22that was slightly in advance of phase
59:23compared to the market and know we see
59:25the market moving to that and as the
59:26market moves to that the uh every big
59:29players in the market is waking up to
59:31it. So you have Sal force with agent
59:32force. You have agent space of Google.
59:34You I mean everybody's working to to it.
59:37We've had two years of building a
59:39product uh as best as we could that is
59:41really creating us a defensibility today
59:43meaning that our product is probably in
59:45a better state than uh most of those big
59:49players are are shipping today. But
59:51they're also big players with many
59:52developers. So they eventually ship the
59:54thing. I mean we can trust that. And so
59:56I think it's always a question of trying
59:58to build a few
1:00:01some time in advance. Yeah.
1:00:03>> Which
1:00:04>> is completely the contradictory with
1:00:06what I just said before, but that's part
1:00:09of the custom work of building an AI is
1:00:11that you must be building two years in
1:00:12advance even if you don't really see it
1:00:15yet. And so that's a real challenge.
1:00:17Building an infra is more a common
1:00:19center. We don't know what it looks like
1:00:20being something like work plans and
1:00:22stuff like that that we discussed on on
1:00:24the podcast. I think we don't know
1:00:25exactly where it looks like, but you
1:00:27have to be investing a lot of resources
1:00:29there because you you have to build
1:00:31conviction that it's where it's going to
1:00:33be in the future and you want to be
1:00:34building it now. And as you do those
1:00:36phases, the the bigger you get and the
1:00:38more gravitas you get and the more the
1:00:41more resources you get and you have
1:00:42maybe a chance of surviving to the open
1:00:44AAI and the Microsoft and the and the
1:00:46Google of the world. That's for us. What
1:00:49about you, Harrison? Yeah, I mean I I
1:00:51think like uh agree with everything you
1:00:54said around just it it being a chaotic
1:00:56time to build. I think like you know
1:00:58execution is is a moat and execution
1:01:00speed and like that's yeah to the point
1:01:02of you know building fast and thinking
1:01:04in the future like yeah I think like we
1:01:06we you know I I think honestly the the
1:01:09team that we have at link chain is
1:01:10fantastic and that's that is a big moat
1:01:12we have and I think we do execute like
1:01:13really fast and and and and really
1:01:15efficiently. I think like a you know a
1:01:17little bit um maybe more kind of like in
1:01:20in the details on that or or or other
1:01:23than that this the fact that there's so
1:01:24much going on in AI can actually be a
1:01:26blessing in some sense as well because
1:01:27competitors will get distracted by other
1:01:28things as well. So they might see
1:01:30something that you do and be like, "Oh,
1:01:31that's cool." But then they see
1:01:32something else that someone else does
1:01:33and they're like, "Oh, that's cool as
1:01:34well." And so they'll So I think like
1:01:35really trying to like understand the
1:01:37problem and uh having conviction in that
1:01:40and like building towards that in like a
1:01:43uh I I actually think like just like
1:01:45understanding in general is actually
1:01:47very hard. Um, and so from like a
1:01:49product point of view, just like
1:01:50understanding what you're building
1:01:51towards and having consistent kind of
1:01:53like product strategy and that or
1:01:55product experience and that a and then
1:01:57you know the features that you add,
1:01:59someone else may be able to copy them,
1:02:01but if they don't have that kind of like
1:02:02holistic understanding, they're not
1:02:03going to do as good a job and it's going
1:02:05to show up kind of like at the margins.
1:02:07And then the other thing that I'd say is
1:02:08like I think a lot of the early things
1:02:10we did were around that kind of like
1:02:12understanding the user experience and
1:02:14building towards that. And now we're
1:02:16starting to build we're we're starting
1:02:18to try to figure out like okay what are
1:02:19the kind of like um deeper technical
1:02:22bets that we can make that like these
1:02:24other things like all kind of like boil
1:02:26down to in some way and and and again
1:02:28like and there's like two or three in
1:02:31particular that that like we're kind of
1:02:32like thinking about and that's not that
1:02:33many right but we we'll weigh many more
1:02:35things but they all kind of like come
1:02:37back to this and so if you're looking
1:02:38from the outside in as a competitor you
1:02:39might say like oh my god they do like a
1:02:41hundred things but there's really like
1:02:42two or three kind of like tech deep
1:02:44technical things that we're betting on.
1:02:45And that's just, you know, having
1:02:47conviction and and kind of like you I I
1:02:50think it's tough because you do need to
1:02:51be moving fast, but you also need to
1:02:53kind of have some sort of consistent
1:02:55kind of like conviction or consistent
1:02:57kind of like technical bets that you're
1:02:59making to build up that that that mo
1:03:02that can't just be replicated in like a
1:03:03week. So, it's a yeah, it's a crazy
1:03:06time. Um uh but but that's how that's
1:03:09how we think about it. you know, at the
1:03:11time that we're talking, uh, not long
1:03:13ago, we had sort of the the wind surf
1:03:16drama, uh, opera of, uh, you know,
1:03:20OpenAI buying them, then that falling
1:03:22through, then, you know, management
1:03:24getting picked up by Google, and then,
1:03:27you know, uh, cognition sort of taking
1:03:30the the rest of the company and sort of
1:03:31saving the employees from from being
1:03:33left without anything from what was
1:03:35looking like a really massive
1:03:37acquisition. Do you think that's like a
1:03:39version of M&A that we're going to see
1:03:41more often? Was there enough of an organ
1:03:43rejection from the startup community
1:03:45towards that practice that like
1:03:48hopefully we don't see that as much? Are
1:03:49you seeing, you know, talent sort of re
1:03:52respond to these kinds of behaviors? I'm
1:03:54Yeah, just curious for your take as
1:03:57founders on the ground and and how Yeah.
1:04:00those those sorts of things are are
1:04:02changing the field perhaps.
1:04:03>> Weird. It's probably going to be
1:04:04controversial, but who cares? Uh I much
1:04:07I think I much rather prefer the um the
1:04:10whole windsoft setup than the um scali
1:04:14setup.
1:04:15Why? Uh the windsoft setup is actually
1:04:18an acquirer and there's been aquire
1:04:20forever and aquire have always been
1:04:23selective on the people they take in.
1:04:26The the weird thing about that aquire is
1:04:28that the the amount is just doesn't make
1:04:29sense. It's just way too big. And so the
1:04:31amount makes it like makes it really not
1:04:34great because you already paid 2.5
1:04:37billion for some folks. Why don't you
1:04:39get all the folks right? Uh generally
1:04:41the acquire is when the company is uh
1:04:44dying in a sense and it's an event that
1:04:46is great because it lets you join a
1:04:48bigger team but it's not with massive
1:04:52amount of money. So the kind of stardom
1:04:54system of AI etc makes those acriers
1:04:58completely weird but I much rather have
1:05:01that then scale is it feels to me
1:05:04slightly more weirder because it's both
1:05:07the CEO acquire but at the same time a
1:05:10not fully completed acquisition just a
1:05:13majority stake uh by which I don't know
1:05:17where what were the dynamics for in
1:05:18terms of returning to the employees and
1:05:21it feels kind of a a bit almost even a
1:05:23little bit more complex and I as a sense
1:05:25a little bit more perverse. But um
1:05:27obviously I I all that being said I do
1:05:31uh think that it was really not
1:05:34acceptable for the employees that were
1:05:36seeing a bunch of folks leave for 2.5 I
1:05:38don't remember the numbers but for
1:05:39billions of dollars and be left with
1:05:41something like what what are we doing
1:05:43guys and having somebody tap you in the
1:05:45back and say you've got a running
1:05:46company that's great go get it. that
1:05:49made the whole setup weird. But uh but
1:05:51at the same time, if you look back and
1:05:52forget about the amounts, it's it was
1:05:54still mostly in acquire. It's just it
1:05:56was weirded out by the uh the massive
1:05:58amounts involved.
1:05:59>> This isn't the first time it's happened.
1:06:00I mean, like Character, Inflection,
1:06:02Adapt, all had versions of this as well.
1:06:04I also think like you're seeing in the
1:06:06markets like there's just this insane
1:06:08price for talent that's going on with,
1:06:10you know, Meta and you know, the rumored
1:06:13salaries and stuff that they're they're
1:06:14paying people to try to get from Open
1:06:15AI. So I think it's like a really just
1:06:18like crazy time in the talent market and
1:06:20I think that's manifested in a few ways
1:06:22including these offers but also
1:06:23including these kind of like aqua hire
1:06:25acquisitions whatever or you know faux
1:06:28aqua hires whatever you want to call
1:06:29them. I feel like the wind surf news is
1:06:31is is new enough where I I actually
1:06:33haven't had that many kind of like
1:06:35detailed conversations with folks about
1:06:36it on on the ground. I think it happened
1:06:38what LA last week or something like that
1:06:40um or or week and a half ago. Yeah.
1:06:42Yeah, I mean I don't really know how it
1:06:46will affect things going forward. I do
1:06:49think that it's uh uh you know I imagine
1:06:52that's not what um the founders had in
1:06:56mind when they started the company or
1:06:58even you know 6 months ago or something
1:07:00like that. And so I don't think it's a
1:07:02great situation at all for kind of like
1:07:04the employees that were left there. I
1:07:05also don't I don't know like I I I I
1:07:08spent a lot of time thinking about why I
1:07:09wanted to start a company before I
1:07:10started a company. I came to the
1:07:11conclusion that I wanted to build
1:07:13something great with people I enjoy
1:07:14working with. Um, and I feel like, you
1:07:17know, I I feel like not only do we have
1:07:18that here at Ling Chain, the chance to
1:07:20build something great, but also like I
1:07:21really enjoy everyone that that that I
1:07:23work with here. And so I think like
1:07:25that's personally kind of like what what
1:07:27what motivates me. And so I think I have
1:07:29a tough time kind of like seeing kind of
1:07:32like myself or Lingchain going that
1:07:34route. But, you know, I I also think if
1:07:35you asked the Windsor founders a year
1:07:37ago, they probably would have said the
1:07:38same thing. Um, so I don't want to I'm
1:07:40not I'm not here to judge anyone. So
1:07:42>> yeah, I I think your point about the
1:07:45just the intensity for talent right now
1:07:47is like, you know, really such an
1:07:50important one as startups like h how do
1:07:53you think about where your competitive
1:07:55advantage is in in such a hot talent
1:07:57marketplace? Like where have you found
1:07:59you're able to to compete most
1:08:01effectively um and and get the folks
1:08:04that are like perfect for your
1:08:06particular mission? And for us, we're
1:08:08building from Paris, so we we have an
1:08:10easy easy an easy day on this one. Uh I
1:08:15think the there's still some competition
1:08:16and we we uh but but we uh it's uh it's
1:08:20uh not nowhere close to to SF and I
1:08:22think we are capable of creating I mean
1:08:25we we're spending a lot of time creating
1:08:26a brand that is really attracting in
1:08:28Paris and has been a really a great
1:08:31thing for us to build the team. We we as
1:08:34well super excited to be working with
1:08:36them. I think that's been our most
1:08:37mostly differentiated approach uh here
1:08:40is around the locality of the injuring
1:08:43team for sure.
1:08:44>> Yeah, we're we're mostly based in San
1:08:45Francisco, so it's been a lot tougher.
1:08:47Um I'd say like you know like one um we
1:08:50we we hire more kind of like just
1:08:54software engineers as opposed to
1:08:55research engineers. And so like the the
1:08:57folks who are getting a lot of the crazy
1:08:58salaries, we're probably not competing
1:09:00for them. That being said, like you
1:09:02know, OpenAI and Anthropic and everyone
1:09:03is also hiring a bunch of software
1:09:05engineers as well. I think, you know,
1:09:07we're we're a lot smaller than them.
1:09:08We're more of a startup. A lot of people
1:09:09want to work at a startup for a variety
1:09:11of reasons. And so that's been the main
1:09:12reason that someone would join us as
1:09:14opposed to one of the model labs.
1:09:16Amazing. Well, I want to just ask one
1:09:17more AI question before we do a bit of
1:09:20uh our our final wrap-ups, which is more
1:09:22of a general one, but what are your
1:09:25rough sort of frameworks for when we can
1:09:28expect sort of true AGI if you think we
1:09:31haven't hit it yet already? uh and you
1:09:33know ASI you know are you on the AI 2027
1:09:37timeline are you more bullish less
1:09:40bullish more scared less scared
1:09:41>> yeah I don't really have a timeline what
1:09:43I'm what what we've seen so far if if
1:09:45you look back is that investment in that
1:09:47ecosystem has been a leading indicator
1:09:50to the progress that has been made which
1:09:52makes sense and is pretty obvious I
1:09:54think uh looking at advancement it's
1:09:55nowhere close to being stopped so we can
1:09:57expect more progress the pace of it etc
1:10:00is very hard to anticipate in any way. I
1:10:02mean at the end of 2024 we were like ah
1:10:06start things start plateauing and all of
1:10:08the sudden uh you've got a new paradigm
1:10:10that kind of emerge and and push back
1:10:13again or push back the accelerator in
1:10:15terms of progress. So it's all very hard
1:10:17to anticipate. What is true is that it
1:10:19seems like the investment is still going
1:10:21uh crazy which means that there is no uh
1:10:24no limits to uh to the kind of uh
1:10:28resources that will be invested in in
1:10:30the in making those models better. It is
1:10:32hard to believe at the same time that
1:10:34there is kind of a hidden
1:10:37like uh limits
1:10:40that supposedly would be here around
1:10:43before we reach even better
1:10:45capabilities.
1:10:47So I think it's mostly a question of of
1:10:49pace to be honest. That's not something
1:10:51I'm spending too much time on because I
1:10:54think we are we and I probably assume
1:10:57that uh it's simparis but we are pretty
1:10:59agnostic to the or pretty um edged in
1:11:03the in both scenarios. I think the
1:11:05scenarios leading to something that is
1:11:07really super human and the machine
1:11:09takeover all work it will require some
1:11:11amount of time to transition to that and
1:11:13there's going to be a lot of product to
1:11:15manage that transition eventually
1:11:18everything's off anyway so in a sense
1:11:20you we're pretty neutral and if the
1:11:22technology kind of plateaus I think the
1:11:25uh the uh uh deployment of that
1:11:27technology in the society uh will still
1:11:30take years and years and years and so I
1:11:32think there's still also a lot of value
1:11:34to be constru to to be to be to be
1:11:36created around building the product that
1:11:38helps that deployment. So I think we're
1:11:40excited to be building in both world. Uh
1:11:43one of the fun key funny ways that
1:11:46motivated me to go back to building a
1:11:48product was uh that which is the last
1:11:51last train before AGI. Uh so it's kind
1:11:55of the last opportunity to be building a
1:11:56company and so that was kind of it's not
1:11:58the true motivation but one of the uh
1:12:00fun reasons why I moved back from open
1:12:02area and started building a startup. So
1:12:04uh so we'll see. I think it's uh to me
1:12:07um I don't have a timeline and I I'm I'm
1:12:09I'm just amazed by how fast it's been
1:12:12progressing and so I don't expect it to
1:12:14stop and so I'm really wondering where
1:12:16we're going to be in two years. Don't
1:12:18know if it's going to be AI 2027, but
1:12:20it's surely going to be uh uh crazy
1:12:22compared to where we're at today.
1:12:24>> Harrison, I saw you nodding along
1:12:25through through most of that. Does that
1:12:27sort of uh is that how you see things
1:12:29more or less?
1:12:31>> Yeah, I I think that's pretty spot on.
1:12:33Like I I I don't spend a ton of time
1:12:35thinking about that either. I think
1:12:36there's I I think even, you know, even
1:12:39if the models get really really good, in
1:12:41order to make them impactful, you'll
1:12:42still want to integrate them in some
1:12:44way. And and that has to happen somehow.
1:12:47And I think that's the type of work that
1:12:49that that we focus a bunch on. I don't
1:12:50have any particular insights and uh so I
1:12:53I I I don't claim to and I don't spend a
1:12:55ton of time thinking about it.
1:12:57>> Well, let's uh move to two sort of
1:12:59wrap-up questions if that sounds good to
1:13:01you. Harrison, let's stay with you. If
1:13:02you had unlimited resources and no
1:13:04operational constraints, what is an
1:13:06experiment you would love to run?
1:13:08>> I don't know the exact experiment, but
1:13:09maybe two areas. One one just one
1:13:12general one and then one more kind of
1:13:13like focused on this. I think memory is
1:13:15really really interesting. I think
1:13:16memory for AI and agents is is and
1:13:19personalization and learning or whatever
1:13:21you want to call it. And so I I don't
1:13:22know exactly what the experiment I would
1:13:24run is, but that's an area that I would
1:13:25absolutely kind of like explore. Um, and
1:13:28then one that's like not related to AI
1:13:29at all, but like uh how do I get the
1:13:32best sleep? Like I just want to sleep
1:13:34really well. Like I I need like eight
1:13:36hours or or I'm terrible the next day.
1:13:38And so like what conditions set me best
1:13:40up for success there? Like that's one
1:13:42that I personally would love to have an
1:13:43answer to. Well, we're seeing some
1:13:45amazing things happening around like
1:13:47sequencing short sleepers and trying to
1:13:49figure out, you know, how to turn that
1:13:51into some kind of therapeutic where, you
1:13:53know, on four hours a night, maybe you
1:13:54can have the same level of productivity
1:13:56and energy or more. So, who knows? Maybe
1:13:58maybe we're just a few years away.
1:14:00>> That would be great. Uh Stan, what about
1:14:02you? What would your your experiment be?
1:14:04At the end of the day, a lot of and
1:14:06that's what we're doing is just be about
1:14:08doing it faster to some extent. I think
1:14:10uh at the end of the day, those models
1:14:13are still pretty smart. They do a lot of
1:14:14really impressive stuff. It's mostly
1:14:16that we don't we don't have the pipes to
1:14:18connect them to the right actions and
1:14:20the right data. And so there's a lot of
1:14:22pipes missing. And so if you could send
1:14:25everybody I mean a very large team to do
1:14:27partnership with every platform up there
1:14:29and create all the pipes etc. and see
1:14:32how much and and be then able to really
1:14:35uh figure out how much of work can be
1:14:38taken by those models because I think we
1:14:40there's this we don't see it quite
1:14:42clearly not only because it there's a
1:14:44there's the model capabilities that is a
1:14:46a limitation there's also the
1:14:48availability of the actions availability
1:14:50of access to the data which is never
1:14:53perfect and so getting perfect there is
1:14:55purely an operational play it's about
1:14:57building stuff or getting partnerships
1:14:58and stuff and so that would be one of
1:15:01Uh the other thing that is a kind of a a
1:15:03side track is I'm I'm super interested
1:15:06into answering the question of whether
1:15:09there is a a maximum team size that
1:15:12exists for a given product especially in
1:15:15the age of AI. uh I think many people do
1:15:18believe that there is a an optimal team
1:15:21size for a given
1:15:24product let's say and that often the
1:15:27scale that some companies uh go into uh
1:15:30to thousands of employees and stuff like
1:15:32that is mostly people getting uh uh busy
1:15:36uh by doing uh by just organizing
1:15:38themsel in a sense I think with uh uh
1:15:41agents being ambient around them and
1:15:43being used in many ways this has come
1:15:45become more and more True. And so if I
1:15:48could really quickly find out, you know,
1:15:50the uh scale the team with really great
1:15:53talent at many different levels and
1:15:55quickly operate it to see how it feels.
1:15:58That would be something that would be
1:15:59very interested to figure out how if
1:16:02there is such maximum of efficiency at
1:16:04some point.
1:16:04>> I love that. Yeah. Where where you hit
1:16:06the limit. Um okay, this is a a question
1:16:08I love to end uh with usually. If you
1:16:11had the power to assign a book to
1:16:13everyone on earth to read and
1:16:15understand, uh what book would you like
1:16:17to assign? Uh and uh why don't we stick
1:16:20with you Stan and we can end with
1:16:22Harrison.
1:16:23>> Yeah, one book that I uh loved is uh is
1:16:26and that is I'm not sure I understood it
1:16:28completely.
1:16:30[laughter]
1:16:31Uh it's called it's a Greg Eaggan
1:16:33permitation city because I think it just
1:16:35uh sends you into thinking about the
1:16:39nature of consciousness in a way that is
1:16:41extremely interesting and so uh that's
1:16:43one uh that's one that comes to mind.
1:16:45>> Is it a novel fiction?
1:16:46>> It's a novel. Okay. Yeah, it sounded
1:16:48like it might be. Uh wow, that sounds
1:16:50really interesting. I can add that to my
1:16:52list. And what about you, Harrison?
1:16:54>> One one of my favorite books uh is is uh
1:16:58Range by David Epstein. Um, and it
1:17:02actually, you know, it's great for the
1:17:03the the title of the podcast, the
1:17:04generalist, because it's all about how
1:17:06generalists succeed in kind of like a a
1:17:08specialist world. And I thought that was
1:17:09really interesting. And you know, when
1:17:11they succeed and when it's also good to
1:17:12be a specialist, and you know, just, you
1:17:15know, how a lot of the great things in
1:17:16in humanity, I've come from just uh
1:17:19generalists with range connecting dots
1:17:21and and and and putting things together.
1:17:23And so I I really enjoyed that.
1:17:25>> What a perfect uh kod for for us uh
1:17:28here. Um, thank you both so much for
1:17:31taking the time. Uh, I learned, uh, so
1:17:33much and, uh, yeah, really enjoy
1:17:35chatting with you both. That's it. Thank
1:17:38you for listening to this episode of The
1:17:40Generalist Podcast. Please subscribe on
1:17:43Apple Podcast, Spotify, or your
1:17:45preferred podcast app. Ratings and
1:17:47reviews help others discover these
1:17:49discussions. So, if you enjoyed the
1:17:51conversation, I'd be grateful if you
1:17:52could take a moment to leave one. For
1:17:54all past episodes and more, visit us at
1:17:57the generalist.sub. substack.com.
1:18:00See you next time as we [music] continue
1:18:02to explore the future.

About this transcript

This page contains the full transcript of AI Agents Are Here: What They Can Already Do—and What’s Next by The Generalist, generated from the public captions YouTube serves with the video. The transcript has 16,645 words across 2,273 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.