[High Agency] AI Engineer World's Fair Preview

24:09I'm always about trying to get people to meet each other in sort of many to many races rather than just through me or meeting to be rate limited by me. So the most obvious transition is the transition from a single track conference to a multi-track conference. And we definitely overdid that one. We went from one track to nine tracks. Didn't really expect that at the start, but just this sheer amount of interest and clear swim lanes of things that people were working on that I thought were worthy of their own conference. I basically picked those things up. Last year, one track. And then this year, we have the sort of standbys, the rag and the code generation stuff, which we also had last year.

24:48But then this year, we added multimodality as its own track. This year, we added evals and ops as their own track, which is super, super popular. Agents, we had a bit of last year, but this year, we had much more organized tracks. And then, for example, I wanted to have tracks to address specific criticisms of AI. For example, a lot of people would say that AI is adopted by startups. And that's absolutely true. But, you know, there's interesting stories to tell in the Fortune 500 and larger scale deployments. So I have straight up AI in the Fortune 500. And just putting that in the title attracted a different kind of speaker and different kind of audience, which is great.

25:21I actually thought last year was too startups focused. That gets you in this sort of naval gazey Silicon Valley type bubble where everyone actually like talks a big game and raises a lot of money, but doesn't actually make a ton of revenue. and uh you know i think the revenue is basically in a capitalist society the only thing that keeps you honest and so i'm really proud to initiate that track uh you know we not everyone in there is actually formerly fortune 500 but they're they're at least household names that everyone would care about coinbase and salesforce and then finally uh we for the first time we're also doing the sort of vps of ai track right like trying to and that's the one that you're speaking on which is addressing some team and leadership level conversations around AI strategy and growing a team, organizing a team and setting principles for what that should look like.

26:08For me, the ultimate win condition for this conference is actually like you don't even go for the talks. You just go because you know everyone else is going and you just show up and you're like, I know I should be going for the talks, but my conversations here are so great. I don't need anything else. That's actually a win for me. But these conferences are expensive affairs and people have to expense it. So I'm definitely creating something for people to expense and to be work relevant and to level up, to find jobs, to hire, to launch their products. I think we have three foundation models launching, by the way.

26:37That's sort of, yeah, I never said that publicly. You heard it here first. Are we going to get the big one? Not the big, big one. No comment. We have decent, so I don't think we've reached that tier where people want to give us their big launch. Every year, we're getting more legit. This is the second iteration. The first year, we had OpenAI and Anthropic, which is a good way to ask. How many people have you got coming this year? We're targeting 2 ,000. We're at 1 ,500 now. I don't know what the exact number is going to be. Typically, the online audience. Last year, our in-person was 500, and the online audience was 20 ,000.

27:11Then the asynchronous audience was 150K-ish for a single one of our top talks. Then this year, we're just basically trying to make everything four times larger. Right. I think relative to almost anyone in the space, you are incredibly well positioned to speak to a lot of different types of people through the Latent Space podcast, through your work as a developer relations engineer. I think you have a pretty unique perspective. And so one thing I wanted to ask you about is, you know, what advice do you have for product and AI and leaders? And I have specific questions there, but I want to give you the generic version of the question first to see where you take it.

27:44So if I'm an AI engineer or I'm a product engineer building today, maybe I'm starting a new project, what are the gotchas? What is the best advice that you would give? Oh, boy. I totally don't feel qualified to do it. I can repeat the wisdom of people much smarter and more experienced and more successful than me because that is my role as, I guess, a content person and community person. Okay, I'll give you a common take and I'll give you a harder take. Okay. Common take would be that you should move fast and much more so than a typical YC startup mentality. If you're taking like three months to ship a thing, what could you do to make it three weeks?

28:23You know, that kind of mentality of try to move fast. Because if your typical development timeline assumptions are based on traditional ways of making AI products, then maybe you should think about what is unlocked by foundation models that you can do differently, that is faster, that is mockable by a really shitty prompt. But whatever, it kind of works and get it out there. A core part of the thesis of the original essay was that you should adopt a fire ready aim approach instead of a ready aim fire approach. and where the old ML engineering process was much more deliberative, here you actually win by moving really quickly and getting information from the market, from shipping products, so that you can iterate much faster.

29:05Maybe, you know, in a way, like the gradient descent is performed in the marketplace of the user rather than in a model weight because you don't need to start with data. You need to gradient descent your product. And that's something that resonates at least from what I've seen as well, in that I've seen a lot of companies ship a V1 that's good enough and use the data they gather in production, the eval data, the feedback data to rapidly improve that. And without that feedback data, it's really hard to do so. So there is a flywheel that you can get going if you're quick to deploy something. How do you trade that off with the caution that some larger companies might have about hallucinations or risks?

29:44Those are very important, but also I think probably overstated for a lot of companies because they can just set the expectation that, hey, this is a beta thing, or this is generative AI products. Wink, wink. You all know the risk that comes with that, but we're just trying our best. And that suffices for most companies that are not named google.com with the responsibility that google.com has. But most companies, you are not Google. Go ahead and experiment. Get out of your own head. Go try things. So you said your non-spicy take was people should definitely focus on moving fast and faster than they might in another circumstance.

30:18What was the spicy take? Yeah, it's not spicy because everyone should be moving faster and regardless of AI or not, right? Like actually that's a universal truth. The spicier take is something I've been sitting on myself for a while and, you know, maybe concerns you a little bit. What kinds of startups are performing better and seeing more success in the market? It's vertical rather than horizontal startups. And a lot of people try to build picks and shovels. And obviously some picks and shovels creators will win. But actually, the people that have the most proprietary data stand out the most, pursue the markets with the highest margin because they pursue very price insensitive, non-technical audiences that just have a problem that's the burning pain points that you can solve with AI.

30:58Like they want an AI solution, no one's solving solutions for them. Those vertical startups actually have the unique insight into their customers that they're making the most money and growing the fastest compared to the horizontal one. We have a company that's launching with us that is working in AI and construction, right? And most software engineers don't really like, you know, the real world. We're much happier building developer tools where, you know, we just talk APIs and can log everything. But like, you know, if you're willing to do the nasty work in the real world, then you get appropriately rewarded.

31:30I think that's entirely appropriate. But at the same time, you also have the most domains of specificity and unique knowledge that you can train your models or whatever. You can pursue high margin markets and you're most likely to not be steamrolled by OpenAI when that day comes when OpenAI does a launch and they're like, whoops, my company's gone. So for many, many reasons, I do think that vertical companies have it right. And I wonder if the horizontal players, there's going to be a shakeout where a lot of them fail. And that's just the brutality of the market. I mean, that's neither here nor there.

32:00But I think that people might underappreciate how easy it is to get started on the vertical stuff compared to on the horizontal stuff where you're one of many. And which vertical AI startups have really caught your attention? Like, what do you have in mind when you say that? Is it something like Harvey AI in the legal space or you mentioned construction? Like, what are other examples? Because the counter argument that people will often give is, you know, some of these vertical AI startups, the worry is that they'll get eaten by improvements in the models over time. Why is that concern wrong? And what are some examples that you are really excited about?

32:33Okay, so maybe we start with the examples, which is always challenging to recite at the drop of a hat. So I think Harvey is definitely well-known as an example. I do think like MidJourney is vertical, like it's serving that sort of creative market. You know, and classically, for those who don't know, MidJourney is making somewhere between$200 to$300 million a year, completely bootstrapped with a 50-person team, something of that order. The numbers might be like plus minus 10 % off, but those are the rough magnitude. And that is one of the most successful startups of all time already. So you have to get creative about what that vertical is.

33:09But I think increasingly you will see other verticals that just come up and take on incumbents directly. I am thinking of perplexity there as a very successful anti-Google, even though they sort of wink-wink. They're like, no, we're friends of everyone. No, they're trying to take on Google. And I think they're doing a decent job. It's an open question as to how profitable they are. But they definitely made a dent in terms of the public persona perception. And then like other kinds of verticals, I could name Peter Levils' photo AI and the sort of room, I forget what it is, I think it's called interior AI, right?

33:39And that focuses on real estate agents doing virtual staging for houses, right? Like whenever they're trying to put houses on sale, like they need that. There's a bunch of people serving those verticals. I mean, yes, one of the verticals is developer tooling. I think the sort of cursors and co-pilots of the world, like focusing on that is like, it's a very clear use case where like you win when everyone basically says like you need to have this or you're behind in that industry. And I think the Harvey customers would say that. I think the cursor and co-pilot customers would say that. And I think like there is a tool for like building the AI, you know, companion or friend or like sort of default tool of every single vertical out there.

34:17Like there's got to be someone for the medical side. I can't really name them just because I'm not that close to that side. But like the same exact thing as Harvey is doing, you would apply on the medical side. But Brightwave, our most recent podcast guest is doing that for hedge funds. It's obvious that people want to do research quickly. A lot of research analysis is very commoditizable using language models. You can do that a lot quicker. And yes, you're going to hallucinate some, but so do your analysts, by the way. And it's no different. So yeah, those are all verticals. And for me, on my very, very small, you know, perch in the world, new summarization is a vertical that no one's actually seriously pursued.

34:53Everyone, you know, builds it as a feature. I'm building it as a product. Like, it's so, it's too easy to win here. Like, it's, that's one thing, something that really, like, puzzles me that, like, you have to work really, really hard in horizontal, like, dev tools type things to, like, win. You know, everything has to be open source. Everything has to be, like, distributed and scalable. and you need all these checkboxes and you need to sock two and whatever in order to get your customer. And then you just build one vertical product that solves the pain point and everyone is raving about it even though they don't really know anything else about it.

35:24And that's fantastic. That's what PMF should look like. You've worked now, I guess, at three different dev tool companies, unicorn companies as in DevRel. I think you've developed a lot of taste for developer tools. You're working on small AI yourself. We just discussed that the horizontal developer tool space for AI is increasingly crowded. There's a ton of different options. If you're a buyer trying to navigate this, there's a lot of noise. So what advice would you have for someone, not on the vendor side, but on the buyer side, about what is smoke and mirrors versus what do people really need?

35:59Are there tools you're excited about? What's the stack, I guess, that people should be thinking about building? I think that there is actually a lot that you should buy from the beginning and then selectively build later on because the community has probably run into more problems than you know of. And actually, it serves you well to buy first, understand your own problems with the protection of the communities to figure this out. And then once you understand where you differ from the community, then you can eject if you want. So obviously, you want to think about lock-in. And it's more expensive to build because you're going to slow yourself down if you try to build everything in-house.

36:38The sort of not-invented-here syndrome is very big. And I think some things like you need some kind of evals platform makes absolute sense. You need some observability and monitoring on your APIs, just like other regular APIs. But now with an AI flavor, some try to charge you a lot for that. And obviously, there's a right price for this. But it does absolutely make sense that you should buy all these things just to move faster. going back to the organizing principle of what makes sense in this world. And then you can build later on if you're like, none of these solutions make sense for our case, or we're seriously overpaying for this.

37:10Fine, you only paid for two months of this. Go back to build, backfill, whatever you need. Only once you've understood the problem and you've explored the features that other people have built to serve other customers, because you're not going to be the only one running into the fact that, oh, you need portable keys or time-nuted keys for your API, key rotation for your OpenAI keys, or you need to track your inputs and outputs. so that you can do evaluations or fine tuning or whatever. Everyone has the same issues. You're not special. Just buy it because you're sharing the development cost with everyone.

37:39And then there's the AI product tooling, which is for you building products for your customers. And then there's the internal productivity tooling, which is a separate thing. And actually, that gets adopted a lot quicker. And that typically comes in the form of developer tool, right? Like in terms of your people using Copilot or Cursor or SourceGraph Kodi or anything of that sort. That's the baseline. That's the most proven one. At least most people in the AI community are using that. But then also there's the other productivity stack of like your meeting summarizers and what have you. You know, I'm not really in the business of picking productivity tooling, but I do think that there is a lot of advantages to that.

38:12Ultimately, the point that we're going to get to here is we should seriously think about virtual employees or like AI employees that would reliably perform parts of the jobs that you would normally assign to a human. That role is going to start small now and then going to grow over time. The people who are most able to take advantage of AI to do that tasks are going to be the most capital leverage, right? Like that they don't have to manage humans to do that. These things can work online for you. And that's fantastic. So like, yeah, we haven't really seen that yet. I think the closest I've seen is Lindy AI from my friend, Booker Velo.

38:43But again, like, you know, it's not really being adopted at massive scale yet. Like every year I have one essay that sends the test of time. This essay is The Sour Lesson, as opposed to The Bitter Lesson by Rich Sutton, which typically talks about the scaling of models and how you should basically never bet against scale. The sour lesson is always actually more about trying to compare humans and AI on that sort of equal footing. When actually they're, due to more of X paradox, they're actually good at different things. And they're just always going to learn differently, develop differently. And you should not try to directly compare humans and AIs.

39:11So like, I think the way that we interact with these AI employees, let's call it, is going to be very different than humans. And actually the analogy isn't going to be that light. Like there are going to be some things that we find where these are killer apps. and that we would never give a human and vice versa for the other side. I think I somewhat agree with that. I mean, for that reason, I think actually the idea of like human level in AI is probably a mistake because they're already superhuman in some dimensions and they're like much weaker in others, right? Like if we kind of look at the graph of skills, there are things that humans find trivial that are really hard for AI.

39:42And so the moment that AIs are able to match humans on the task where they're currently weaker than us, they'll already be superhuman, right? There's not going to be a moment where they're human level because they're already much better than us at many things, retrieval, memory, et cetera. Yeah, exactly. I'm actually going to move on because there's one final thing that I want to chat to you about before we wrap, which is just trends. There is so much noise. There's so much hype. There's so many new papers. There's something released every day. And I think you do a really good job of curating this for the community.

40:13You have your automatic summaries, but you also have the podcast and tweets, et cetera. If someone was going to focus on just a small number of trends or like a couple of things that, you know, they're trying to find the signal from the noise, what are the trends that you think people should really be paying attention to that are going on right now that maybe are slightly overlooked? Or maybe they're not overlooked, but just in the amount of noise there is, it's just hard to follow them. Okay, there are a few categories of this, which I could point out. We do occasional essays on Lanespace. And what I would highlight is the Four Wars essay and then the Research Directions essay, which I'll pick out here.

40:50And then there's other trends that didn't fit those buckets, but I'll just name them and people can go read up on that. The trends are that there are several key battlegrounds that you as a company, if you're going to be an AI company of any consequence, you're going to have a beachfront on this war or you're not really in the game at all because you're fighting a war where there's no opponent. because there's no limited resource and everyone is aligned in the same way. So for example, open source is not really a war, at least in the technical domain. It is in the political domain, but not in the technical domain because most people are pro-open source in tech.

41:22That's not a war. What is a war is fighting over data, fighting over GPUs, fighting over the God model versus the sort of domain-specific model, and then fighting over rag and ops. And so those are the four battlegrounds that I picked out that I think are sort of key battlegrounds where there will be winners, but all the participants cannot all win, and that's what makes it somewhat zero-sum and interesting as a war. That also means that there are several domains where I don't think there are wars yet, which is because they're not interesting or contentious. So like co-generation, very, very important problem, but no war because everyone's trying to make their own headway and no one's really figured out what's the next thing after co-pilot, right?

41:59Maybe it's dev and maybe it's not. Okay, those are the four wars. You can see our reading on that. And then research directions, So moving from the sort of commercial space into the research space. And I think having a filter for what kind of research matters actually really helps to survive on Twitter. Because you're just constantly DDoS by all these influencers going like, this changes everything. And, you know, this week there was like this like paper about how like you don't need matrix multiplications anymore. Oh, my God, NVIDIA is going to zero. And I'm like, what kind of, what are you smoking?

42:32I want that. But so like having a filter for like what is worthwhile research direction is actually important. I put myself on record as like ranking a list of directions, right? So for us, it's long inference as number one. Synthetic data, number two. Alternative architecture is number three. Mixture of experts and merging models, number four. And online algorithms, number five. And yeah, those are the main trends. You know, I can always talk about the other trends that we're seeing. So for example, having like what is a Moore's law of AI and what are the long-term trends you can bet on, right?

43:02And I think we're trying to define this for one of our podcasts that we're doing. But basically, the cost of a 70 MMLU every single year goes down by something like 5 to 10x. And so this is the observed trend, and it's going to keep going down. So it makes sense to build a product that loses money today because you're going to be ahead. You just have to wait it out because of Moore's Law. It's going to bail you out of whatever sort of bad inference ideas you may have. For people who might not know, I assume most of our audience would, but just briefly, what's MMLU? Just let people have the context.

43:32I think it's multimodal. No, it's not multimodal. Massively something. Language understanding. Language understanding. It's not multimodal. It's something about domain specifics. Basically, it's a conglomeration of all the professional exams that you could possibly ever want as a human. It's massive multitask are the two M's. Massive multitask, right? Fun fact, it's created by Dan Hendricks, who is a very, very noted AI safetyist. And that's a whole different discussion because it's very ironic that the thing that he created is like the primary tracker for AGI. Anyway, it's the primary number that every LLM basically compares each other by.

44:10And this is not something that you should fixate on because it's very likely within two to three years we'll move on to something else. Yeah, benchmarks are very much flyby. Once they saturate, we create the next one. Is one way to interpret what you're saying is this kind of Moore's law of AI. MMLU is a specific one you've chosen for now, but it's really a statement more on like the intelligence or the capabilities of the model at professional tasks we care about every year is going down by 5 or 10x or something like that, the cost to achieve the same performance. Yeah, same IQ. You can also call that.

44:40Like our GPT-4 level model was not possible in 2022. And suddenly in 2022, you could pay something like$20 per million tokens for it. Now the cost is$2 per million tokens. It is on its way to 0.5 or 0.25, depending on if you look at DeepSeq MOE, being a legitimate GPT-4 level model, and it will trend towards zero. So, okay, that's great. But then our levels are sort of bar for what acceptable AI intelligence will increase when GPT-5 drops. And so then when GPT-5 starts on a higher curve and goes down again. And that's a very classic sort of cost curve model that's typically what semiconductors, you know, typically operate on, like the different semiconductor process nodes.

45:23And they operate on that curve. And actually you want to understand understand that and build that into your product planning, right? So it's not just cost. I have a few trends here that I haven't really published, but I'll just kind of go through it. It's commodification of intelligence, which is the MMORPG cost going down over time. It's also inference speed as well. So like going from like maybe, you know, 170 to 100 tokens per second to 500. I've had very credible sources tell me that Brock is aiming for 5 ,000 tokens per second. So another 10X from where they're here. And so what do you do differently, right?

45:56Every 10x unlocks a different kind of product. So that's roughly 10 pages of text or a full essay every second in terms of inference speed. Yeah, I used to emphasize every AI API must incorporate streaming because you want to stream out tokens when you're auto-aggressively generating things. But when you're generating 5 ,000 tokens per second, you don't need it. It's crazy. Okay, so I'll keep going. Context is treading to infinity, right? We used to have 4 ,000 token context models, and I thought that was enough. And now we have a million token context models, and I don't know what to do with it.

46:31But people will find the use cases for that. It's really interesting. Multimodal everything is another trend that I call out. I mean, this is obvious now because it's GPT-4-0, but all modalities in, all modalities out is a very nice shorthand for that. There's nuances around that. And then finally, variance is a very interesting trend, which is definitely one of those underrated things that maybe could become more prominent over time. What this is, is basically a lot of the use cases that we talked about for work is temperature zero use cases. You receive this God model from the heavens. And the first thing you want to do is lock it down and make it do your retrieval augmented generation, right?

47:04Like, that's the most boring possible use case. Like, what if you just let it loose and try to make it think of things that you never thought of? you know, that's this whole comparative benefit and you're chaining it down, trying to force it to do this other very unnatural thing for it. What if hallucination was a feature and not a bug? Right. So there's this whole emerging category of temperature tool use cases is what I call it, where they're diametrically opposed to the temperature zero people, where like, hallucination is a feature, like, come on in, like, this is actually really great, because I never thought of that.

47:32And creativity is expensive on my team. And if I can, you know, spend a few dollars to add more creativity, absolutely, I should do that. Yeah, that last one, I think is a particularly interesting one that I resonate with. I think it's not just for creativity, but actually, if we want to be able to create new knowledge from these systems, I don't think we can do it with LLMs alone, but the combination of a model that can act as a conjecture machine, can chuck out possible explanations for things, coupled with some way to actually test that and measure it, I think is a way that you start to build AI or computer systems that can generate new knowledge.

48:05And obviously, you want those operating, as you say, sort of high temperature, not deterministic mode. All right, Sean, I think that's all we're going to have time for today. I feel like I could talk to you for hours and hours. Just before we go, if people want to attend the conference, how do they get tickets? Where do they find it? And where else can they find you on the internet? Yeah, probably the smartest thing I did when writing the essay is I bought the domain. The domain is ai.engineer. You go to ai.engineer and you always see the currently active conference. And I've created the code IAgency for people, for listeners who've come in this far and want to get last minute tickets.

48:38We'll hopefully get this out as soon as possible. But yeah, it's definitely the industry conference to be at. You can find me on Twitter at Swix and our podcast, as well as Latent.Space. We just love the short domains, right? I hope you get high.agency. That's already taken. But we got Latent.Space donated by a listener, actually, which is very fun. You can find more conversations that we have on our podcast. Fantastic. Well, Sean, it's been an absolute pleasure. Thanks for coming on. All right. That's it for today's conversation on high agency. I'm Reza Habib, and I hope you enjoyed our conversation.

49:11If you did enjoy the episode, please take a moment to rate and review us on your favorite podcast platform like Spotify, Apple Podcasts, or wherever you listen, and subscribe. It really helps us reach more AI builders like you. For extras, show notes, and more episodes of high agency, check out humanloop.com slash podcast. If today's conversation sparked any new ideas or insights, I'd really love to hear from you. Your feedback means a lot and helps us create the content that matters most to you. Email me at razahimulub.com or find me at razrasgal on X. Until next time, Raza.