Joshua Tan

Mm-hmm.

Now, to be clear, like if I'm president of the EU, the EU has like five percent of global AI compute capacity, right?

It's like less all of the EU has less than Elon does in Texas. Right? So

Go.

Yes.

even if you brought all that together, you're still operating from a huge deficit.

Yeah.

So I think you need to be investing in.

A very like different strategy. Like you you can't not be using compute and training. Like you you need to have that going. That's like table stakes. you should be investing more in your human capital and your research. You should be, I mean, frankly, trying to attract more of the companies, all the European companies that have ditched the EU for San Francisco, try get more of them back.

Okay.

There's a talent migration thing that is a huge part of it. you should be, I mean, I'm biased here. My sense is like if the only way Europe can potentially catch up is to make this almost like a continent-wide

you

public works project,

Welcome to Tastebench. I'm Philip Zahn, my co host is Bruno Marnette, and this is a podcast about taste and the creative side of AI. Our guest today is Joshua Tan, a mathematician and computer scientist working at the intersection of artificial intelligence and collective intelligence. He is interim CTO at Current AI, leads product at Public AI, and works on innovation at Stanford Open Lab.

Thanks a lot, Josh, for joining. We are very glad to have you. You're a man of many talents, as they say. You have a lot of things that you do, many different things, but all circling around digital worlds. So let me start actually with the question What are you up to these days? What are you working on?

So I just recently took a new job as the interim CTO of this organization called Current AI. so Current AI is you know, global public-private partnership launched by President Macron and 10 other countries to build the public interest AI stack. And what I'm doing is leading technology there and kind of leading the product team that's building out a let's call it like a full stack.

open source alternative to Chat GPT and Claude

you

Perfect. Before we go into more details of what is actually happening in that company, you're already making this contrast to there are the companies, all we all know, and then there's a public initiative. you have been a lot involved in the past also on AI democratization politics. And somewhere you I read this and you correct me if it's wrong, you argued that the safety movement is failing as a political movement. Can you expand on what you mean by that?

.

Find the most provocative thing I've said on the internet. What did you mean? No, it's this is probably

That's not that that's not it. Trust me.

you

not the most provocative thing. this is really just pointing out something that I think that was like any AI safety person would say, which is that capabilities in AI were increasing at a seemingly accelerating rate, that all the attempts to pause AI development or cape AI capabilities research were not working.

Okay.

and that, you know, AI safety is a movement. I mean, there are other things that was trying to accomplish that I think have there's like a lot of AI research, AI safety research happening. But the main thing that was trying to achieve at the say at the time, that was not working. Capabilities were increasing, and it did not seem like that AI safety research or alignment research was catching

Okay.

up in a realistic way.

Your company that you're working with. I mean it's a public initiative, so it's already different on that level. Is there anything else structurally that you are doing differently compared to the Frontier Labs, let's say?

.

current AI strategy is really focused on, how do I say it, bringing people together, orchestrating things that are already out there in the existing open source ecosystem. in that sense, it's already quite different, right? If we think of I don't know.

ChatGBT as a kind of closed source stack on top of a closed source model. You can think of perplexity as a closed source stack on top of a you know notionally open, open weights model, let's call it. What we want to build is a fundamentally open source model on top of an open source stack on top of an open source model, like a fully open source model. So that's in some sense like clearly different, right?

We're trying to really invest

.

in the open source ecosystem in an attempt to build more public goods that everybody can share, that everybody in the world can benefit from.

Aren't you also concerned about the fact that it's open? That this is actually I mean, this is one of the arguments brought forward that if you're developing in the open then who god knows anyone can misuse it basically, and much easier than let's say if it's stored within the realm of a private company.

partly this is why we're first off, no, the open source model, the open way models are already out there, right?

Hm.

in some sense what we're trying to do is like help people adopt them over the existing closed source ones. Now, when it comes to safety, this is a real concern.

In the services, in our deployments, we're gonna be incorporating a lot of I think like trust and safety architecture and resources. For example, building on the work of this amazing nonprofit project called Roost, I think robust open safety tools.

to make sure that these models and the you know the sort of the the deployments on top of them really are safe, that they're not empowering bad actors in the world.

You worked on this is a bit of a step, but y in the past you also worked on governance broadly defined in the context of decentralized systems, blockchains, et cetera. What's the transfer, if any, that you can apply from what you learned there, maybe also failure wise, to what is going on in the AI space?

Well, actually one thing I was just writing about yesterday is you know open source AI, for better or worse, is at this point a very hype-driven attention-based market. it is mostly hype. you can look at examples like I think PewDiePie put out this like, you know, this thing called Odysseus, which is like a new chat front end. There was

Mila Jovovich, you know, the the actress from The Fifth Element, putting out Mem Palace, which is like memory stack. a lot of what determines captures attention and popularity in stars in open source AI these days, is actually driven by eventually hype, you know, attention, influencers.

In practice, that actually looks a l we we've had some experience with this before, right? You know.

Mmm.

I have

no idea what you're talking about.

Yeah.

Yeah.

so actually I think understanding the, you know, the nature and the fundamentals of the crypto market and the way it functioned back then, vis-a-vis tokens, is actually a good way of at least understanding a little bit of how open source AI, you know, acts like now and how to shape it and what to build it inside of it. To be clear, open source A AI is just one part of the ecosystem. There are giant companies, namely

anthropic ChatGPT that don't quite exist in the same way that they did in crypto. So that's that's like you know giant gorilla in the room. But I think if you looked at just open source AI, like this is something that you could actually learn a lot from looking at you know Pax PASS experience in Web3.

Maybe digging into the kind of decentralization aspect. if I understood correctly, one of your kind of common motivation in all your kind of philanthropic endeavors, there's always been, I guess you've always given a priority towards decentralization. In the crypto world, I think it's pretty well understood why people want decentralization in terms of like freedom from the banks and resistance from authoritarian regimes and all this kind of very...

understandable motivations. the case of AI, what do

Yes.

you think are the of the consequences of a future where AI is either centralized or decentralized? Like, what makes decentralization important in the case of AI?

Yeah, I mean, think a lot of people are very, very concerned about the concentration of power within AI today. how much power the large companies have, how much you know, extractive potential they have, the ability to just like, hey, you know, startup Y. You know, we have all your data because you send all your data and your queries directly to my API.

We can literally just take all the data, create an in-house copy, and you're like, sorry. Or just like cloud gets keeps getting better and better, and like you just don't need any like additional harness, right? You just call the API, solves your problem. this is like a perspective that you might take from a you know like a company to company perspective.

Hmm.

Obviously, there's just lots of questions about like, hey, if this is gonna be like taking away, you know.

A lot of jobs. That is a extremely sensitive economic proposition. that effectively a few companies, you know, are potentially like replacing entire labor markets. It's not just like competing in a given like product market.

That is a huge deal, I think, that scares a lot of people. motivating, I think, many governments to sort of like weigh it in and say, hey, you know, you can't just do this willy-nilly. There needs to be something in place to, you know, to to make sure that people have a have a voice or that their interest is represented in the deployment of some of these technologies.

I'm curious if you imagine if you were from entropy perspective, they obviously think that the good guys, because everyone thinks they are the good guys. They're trying to gather inputs on what they should do. obviously the economy consequences are important regardless, but it

imagining jobs is not an issue and imagining that the central kind of AI company is the benevolent type. Would there still be an issue with having just kind of one AI model that everybody uses? Or do you see it being important to have plurality of models around?

Yeah,

I think of this more as a research question rather than as a norm per se.

Mm-hmm.

So

I personally think it is a good idea to have many models. to have a world where there are many models.

It may be the fact, unfortunately, that this is an industry or like a product category where it just makes sense

Mm-hmm.

for there to be like one giant model where invest like one or two giant models that you just invest in, right? And you can like once you have the really, really big model, you can distill from that smaller

Hehe.

models that are more application specific.

But it might just be like actually, you know, technologically, from an engineering perspective, it's just more efficient to do it this other way. I think there are a lot of different technological variables that go into determining which one is gonna win, whether you have like, you know.

A mixture or an ensemble of different small models, which are all diverse, which more focus on particular things. And that probably is some version of that is already running running underneath the hood inside places like Anthropic. Or you just have one really, really large model and you spin up different copies of this where you distill like smaller versions of this, and that achieves the best performance. Like, I don't know yet.

Maybe I

Okay.

understand the argument in the time of a, there's indeed like a technical layer. Like there's a question, how do you achieve the same output with one or more models? But for me, there's also a question of the kind of product layer or like the behavioral layer, right? I think right now one concern people could have say in the coding space is that if everybody's using Cloud to code, it means that whenever a front-ender is by coding an app,

it's gonna be purple, right? They always default to like this average colors, kind of average taste is, you know, the same rounded corner and so on, right? There's a certain kind of a uniformization of outputs. Likewise, if half of the posts on LinkedIn are generated with Cloud, everything you read on LinkedIn now, you know, sound the same. Do

Yeah.

you think that's something that a plurality of model or like more decentralization could help solve in some ways?

I think it can help for sure. I also think like there are other ways of tackling that problem. Like you know, just specialize in your agent harness, even if you're using cloud

Mm-hmm.

underneath the hood. Maybe I mean this is also like a research question, right? It's not clear because at the end of the day, most agents are just like very thin layers of customization over Claude or or T or whatever it is, right?

And clearly like Gemini, Claude, and Chat 2D have their differences and they do have slightly different styles. And you can configure those those, right? but if you were just looking for like you know different slight differences in design, different choices in like how like front ends get spit out, I don't think you have to go all the way to like a completely different model, right?

Yeah, I am.

I think there are other reasons where you might want a different model. For example, like, you know, if you were representing like a s entire culture, like if you're Swedish, let's say, and you start using ChatGPT and you realize, hey, it speaks like decent Swedish, but everything that it says to you in Swedish sounds like it's an American talking to you in Swedish.

It's like they're just like they'll say say things that like you know Swedish people wouldn't really say or like do things that like there are lots of like these cultural aspects that are embedded in language that kind of get overwritten if you're just using, you know, a model that's like everything on the English internet and like all that, right?

for code it might be a little bit different, but you can also imagine like there being like code communities that want models that are very specialized to their way of doing things, right? And I could imagine like a model that's like fine-tuned specifically for that, right? so you have some continuous RL loop on it, and it's like continuously sort of like advancing the stylistic preferences and the way

Mm-hmm.

that people code with this group.

And that's actually, I think, to me quite cool. I mean, code also has its own cultures, right?

Hmm.

and insofar as you're using cloud, you are a little bit just buying into like a model culture.

Let me do the following question. This is hyp kind of putting you into a hypothetical situation. Say, you know, we make you president of the European Union. You can decide what you want. We have a lot of middle powers that are s decent, they are very good in some sense on certain research frontiers, but clearly are not on a economic level and also not on a hyperscalar level, any anything close to to the United States.

What would you do?

I would take all the hundreds of millions of dollars being thrown at national models and build one giant. I would I would just like bring all those teams together,

Hmm.

ideally co-locate them as many of them as possible, and just start working on I wouldn't call it like a massive model, but like a like like a deep seek level model, right?

Just so we're not like wasting resources.

Mm-hmm.

Now, to be clear, like if I'm president of the EU, the EU has like five percent of global AI compute capacity, right?

It's like less all of the EU has less than Elon does in Texas. Right? So

Go.

Yes.

even if you brought all that together, you're still operating from a huge deficit.

Yeah.

So I think you need to be investing in.

A very like different strategy. Like you you can't not be using compute and training. Like you you need to have that going. That's like table stakes. you should be investing more in your human capital and your research. You should be, I mean, frankly, trying to attract more of the companies, all the European companies that have ditched the EU for San Francisco, try get more of them back.

Okay.

There's a talent migration thing that is a huge part of it. you should be, I mean, I'm biased here. My sense is like if the only way Europe can potentially catch up is to make this almost like a continent-wide

you

public works project, right?

where we try to in a very like policy forward way.

Okay.

know orient our societies as a whole around some of these like data flywheels, right? And the you know what enables this is that you know how we build AI and who benefits is not just like private company. Nobody's gonna trust just like one company to do this. But if this is actually like a

a society or community-wide effort to like you know almost like a moon landing type thing, right?

Mm-hmm.

The entire country works together, where the entire continent works together to like achieve this vision of like a competitive, an aligned, a public AI.

Even then you're still at a, you know, there's like lots of problems. At the end of the day,

Yes.

Yeah.

Enthropic and ChatGPT can just pay for a bunch of data. Like where this would be useful is like if you got the entire like enough parties on your side and really rallied your country, your community, and your, you know, your your culture, you could have a lot of data. Perhaps orders of magnitude more data than ChatGPT and Athropic have.

You still don't have as much money as some of these

Yes. Yeah.

companies do. It's just like insane how much money they have. you also just frankly don't have compute, right? Even if you started now. I mean, my god, I'm not even gonna talk about the gigafactories thing, but the if you just started putting as much compute as you could into the ground, you would still be like running short all the way up to like 2020 eight, very likely.

because at the end of the day there's one other American company called Nvidia that like controls the pipeline for these things. and there's like a

Yeah.

you know just a huge backlog. So anyways, make the place you can, focus on coordinating, defragmenting.

Yeah.

I'd love to,

this is such a deep stack, obviously the AI stack is very complex, there's so many layers. Maybe just to go back again, more on the, guess, closer to the final user. One thing we've seen in Europe sometime is adoption sometimes works within one country. Like know Mistral has been able to leverage a bit of the famous French kind of patriotism to...

get some traction within France, but even the neighbors in Germany don't really care about trying mistrial, right? I imagine the, generally speaking, you this power that institutions would have, just even like being something government-backed doesn't necessarily mean that people would, you know, trust it more, right? That's another

Yeah.

issue of our modern times, right? What we see, for instance, in the UK, know, institution like the BBC used to be, you know, we used to think that because it's public, it's going to be

some sort of, know, different category of institution is going to be more trusted, but obviously it has its own challenges and so on. I imagine this is something you must be thinking about these days, the question of adoption, since you are building kind of a competitor to the private offerings. What do you think of the unlock? Like what do you think is the most promising way we might end up having, you know,

people moving away from their chat GPT addiction and so on to embrace the one you're working on, for instance, or other similar efforts of more like sovereign AI products.

Yeah.

well first off to be frank, I think Mistra has around like one percent market adoption in France at the moment. at least not.

Yeah, that didn't last very long.

so it's I yeah, even in the home market, I I think it's it's quite hard to compete. I mean, obviously they're they're making quite a lot of money on enterprise deals. but it's

Yeah, i in practice like there there's clearly like a dominant player in the global market today.

So what would it take to change that? Or what would take to get even to say 5 % of the world using something else, something more public?

So

I think it's a combination of like price slash unit economics and what I think of is like localization. okay, let me say one thing about sovereignty and sovereign AI.

Mm-hmm.

I think policymakers care a lot about sovereignty. if you're a national defense, obviously you care about that a lot.

If you're a normal user, it just it doesn't matter as much, right? vast majority of users would just use whatever is most convenient and works the best. and is cheapest. The so saying that you're the I don't know, French sovereign option on French with French data, I to be

Yeah.

clear, like it appeals to some people, though hilariously

It's very niche.

I realize that a lot of it is actually just on AWS servers in France.

Yeah.

The but it's you know it it doesn't matter to most people. They just like it's just that's just in stuff that's effectively invisible to them. What's not invisible to them is the price. So the price for ChatGPT for Cloud that is going to increase. It's already increasing.

Yes.

It will increase more. there's a huge amount of subsidy that's happening right now. companies

Yeah, 100%.

are just spending a huge amount of money to lock as many people in.

Get them used to Cloud or used to ChatGPT and then they're gonna just like jack it up. And you're gonna keep using it because like these things are economically valuable, right?

So I think like we need to be competing at the very least just on price while we can, you know, even now and then absolutely later, right? We need to figure out, figure out some sort of different unit economics that allows us to compete on price. in the long term, because in the long term, like the VC capital is gonna disappear at some point, right? and the question is like, what is there, you know,

Mm-hmm.

and that's when the prices get jacked up and

We can reason can we reasonably offer like a $50 a year option to anthropics, like a thousand dollar a year option, right? If we can, and if the quality of that product is still good enough, then I think we can like very sincerely compete. a I think an important question here is also like distribution. It's about like easy ease of access. Like that's something you really have to like take into account.

You can talk about like this thing all you want, but like nobody knows about it, nobody ever hears about it. you don't have the distribution. People can't get access to it. It's like the access in France looks different from the way it looks in Germany, looks way in Sweden. I go, yeah. That that that can't happen. It just needs to be as

Okay. Okay.

easy to access as other that's like a key part of the product quality.

So I said, price and what was the other thing I said?

Access, quality, locality.

Locality.

Yeah, locality, right?

I think by the nature part of the nature of these products is that they are like fundamentally like very global. Like you get the same Claude in Singapore as you get in the US. which is sometimes a feature because you know plot cloud is pretty good at many use cases. but sometimes it's not a feature because you know, like you you want something that's more like reflected in local local custom, local culture, or just local data.

Right. right now we have a lot of sovereign AIs that are worse on pretty much almost every category, every benchmark relative to the big frontier models.

Hmm.

I wish there was a way that there could be better on a set of like local benchmarks. Right. how do we execute that strategy? How do we build like actual real local AI? That's a big question. How we think about like, does it make sense for there to be multiple models? Right? Do people actually care about this? Or what maybe there are like it's possible also that just like you know, to be frank, sometimes

the people, you know, the three of us, but also people maybe like listening to this, maybe l live in a slight bubble. and that there are lots of people who are not definitely not definitely no bubble around me.

No, no, no, no. Of course not. We no Which bubble?

Mm-hmm.

Mm-hmm.

the there are a lot of people who are not adopting AI, who are, you know, more integrated in the quote unquote real economy in their various countries.

And I would say like maybe their use cases are just not actually well solved by AI today. And maybe those use cases are actually more quote unquote local. And if we had a more local strategy, and to be clear, l often people when people talk about local AI, they're often talking about like sort of like on your local computer. I mean something a little bit different.

But if we could realize that kind of like no more quote unquote local or community scale or industry scale strategy, that could be actually like a completely different approach to competing with the big AIs.

Maybe what you're talking about remind me a lot of electronic frontier foundation principles, like technology being more custom to what people want. And there's usually this tension between closed source proprietary products, which would be built with specific incentives, which are not necessarily aligned with the user's need. The most famous misalignment that everyone

tend to think about is addiction, Like SaaS company would try to make the product as addictive as possible. We see that maybe charge GPT would always try to keep you in a conversation longer. We'd always prompt to say, hey, I can do this for you now. And you have other kind of addictive behavior perhaps in the, maybe the personality of Claude, the friendly, slightly fun, charming personality is also potentially.

align with people spending more time on cloud rather than less. Do you imagine that the kind of approach you're following now might lead in a different product experience, which is perhaps less addictive or has other benefits?

Yeah, I think from a persona perspective,

I mean, okay, there are elements of product strategy which I can't quite share yet. but I think generally speaking, yes, we do not anticipate on building deeply addictive products that just like get people to engage over and over again to their

Mm-hmm.

detriment. That we would want to do something that respects, you know, human autonomy, that tries to support.

human labor, for example, and human agency, as opposed to take it away in order to monetize it. And partly that's because we don't have I mean obviously, okay, to be clear, we do have financial incentives, right? Even as a nonprofit, the money has to come from somewhere.

But we are not there to making money by building like addictive products that force people or try to get people to be on their twenty-four-seven. That is not the business model.

Can I maybe move a bit in a different direction from here? But there is a connection. I mean, you mentioned now autonomy. there's a lot of discussions about what's the next frontier, right? Because we've basically I mean we the the people who develop these models basically have solved a lot of these things and tasks that were hard before, like coding. to a large degree you can do now things without knowing the details basically. You need to instruct the direction and the goal. So the

Okay.

Constraint seems to be moving into what people refer to as taste or judgment.

How do you see it?

The autonomy. So if you think about how people normally form taste and judgment and they are basically constantly relying on these tools, do you think this is an issue? In a sense it's it's kind of a problem for the taste formation, if you want? Or do you feel like p nah people will figure it out like before? It's just a new technology on that level.

I guess I'm more on the

Actually.

Okay. Intuitively, I would be on the lighter side, as in I believe that humans are pretty smart, and they'll tend to figure it out and develop various forms of taste. And that's just a matter of like you live your life like most of your early life. You don't ever touch a computer. I hopefully you're not touching a computer all the time. you know, you desert like people develop taste and preference, like often quite early on, right? And then kind of iterates.

And some of your world that you're constructing is going through like AI constructed things. But my I find like people like pretty quickly get a sense of like, if you're just spitting out claud-generated outputs, like people like become very skeptical of you very fast. Like, you literally just generated this for me.

Mm-hmm.

and you like we're starting to develop, you know, the ability to say, like, it's like pretty shallow, right?

So I think like a reasonable person would be able to sort of say, like, this isn't you know, this is not a good direction and change and adapt and evolve keep evolving their sense of taste. The

Do I think this is gonna have like larger impacts on culture? Like clearly it will. Right. if nothing else, if you know, we we talk about like the certain kinds of like mode collapse where, you know, if the AIs keep generating shitty, you know, output and then it just gets consumed by the next generation of AIs and it just like the

All right.

AIs at least will get worse. do I think it will have other impacts on culture? I suspect it will.

I think this like this impact is actually very, very much more visible to people who are not like primary English speakers or who are, you know, working maybe not even like German or French speakers, but working like a lower resource language model. you know, with fewer speakers. there the kind of like that taste for that cultural component becomes like much more salient and much more obvious.

when your kid comes back after like talking to Chat GPT and starts like s using Americanisms, right?

Mm.

becomes very obvious.

Well, to be frank, they

have we have been doing this for seventy years. So it's just you know, it was T V in the eighties or and then Netflix these days, so no, I I'm I mean I'm half joking, of course, but yes. Are you concerned about I mean you said basically people do you tend to be on the side of people figuring it out? Are you concerned with let's say a certain generation that is now exposed to it in a in a form of unprecedented and we don't know how to to deal with it rightly? Or is it like we feel like nah, it's not gonna

Even for them, it's not really not not a real issue.

So I try to look at the instructor like take lessons from the from social media, right? I think early on we thought, you know, what is this gonna do? Like people will figure it out. And like I I think it's fair to say most kids figure it out like, you know, like, okay, let's get off Facebook or this is like kind of unhealthy.

Actually, maybe not most. A pretty decent chunk of kids figure it out and said, like, okay, maybe this isn't like the thing I should be spending like 90 percent of my time on. That said, like these things are quite sticky. A lot of your friends on there. People stick around. And it just like generally, you know, causes people stress. It's like tobacco, you know, it's like the our generation's like digital tobacco, where it's like it's like we kind of all know it's not really good for us, but we keep doing it, keep stressing ourselves out just because like that's what everybody else is doing.

So like there are cases where like you know a certain number, or maybe even the majority of people will not be like strongly affected by it, but a huge percentage

Okay.

of our population will be affected by it in negative ways, right? And that has lots of ongoing impacts to society as a whole. to even if you don't have to put a moral lens on it, it's just like

Productivity, right?

Hmm.

mental health resources, load on load on those resources. you know, just like straight up like people like killing themselves because like a chat bot told them to. these are still like very early warning signs, but you know, anytime these technologies get deployed at scale, these impacts will happen. The like I'm not saying that.

I try to be realistic. Like technologies are gonna come, right? We're going to like, you know, impacts are going to happen that we're not aware of, but we can and should be more aware of them. We should be trying to invest resources into understanding, hey, this is where we we need to pull back. There's this is where we need to invest more resources in pushing forward. And ideally, if we were

Wiser as societies, we could say, like, no, maybe this is not the I mean, just like like Australia, like if you're under 21, you should not be using this, right? Or under 18 or something like that. In practice, like those bands never really work because it's really hard to control a bunch of teenagers. The but like you know, those were the kind of things that the if we could implement

Okay. Okay.

Maybe we should consider doing that.

Yes. let me backtrack a bit to my my question on on on taste and the role it plays basically and judgment and how how it is formed.

The reason why people describe it as a frontier and it's also like the tech industry also frames it as that's a new skill set you basically have to learn in order to be productive. there's this understanding that these mach the machines a as of now are very powerful, but they are lacking taste. tying it back to the localization you mentioned, do you think we could let's say come up with a French model that has sp exquisite taste on on a specific domain?

Pairing wine with the food.

pairing wine with food,

Okay.

or you know, assessing art or going to the Pret à Porte and saying, Yeah, this is fantastic, this is bad. Do do you think there are

You I would

y is it feasible technically or is it like nah, it's a different different generation of tech that will be required for this?

You know what? I'm going to say no. That the taste that we're talking about here is still somehow

Well

I think the French model will make different choices. Or I mean it does, right? Mistral makes different choices. you can literally look at like it has a different political bent from ChatGPT, which has a different bent from Claude. It'll make different choices in like the code it tends to recommend. but I think taste in like a more forward-looking sense of like on the on the boundaries, on the frontier, creating new things.

I don't think that is something that a like a this generation style model is going to be able to produce, is ever going to be able to produce, even if it's, you say it's like the French one built on like French data. I think that actually does require much more human investment. And that the idea of a more s specialized French model is not to replace human taste per se, it is to perhaps.

elicit more human sort of like input. Like, you

Mm-hmm.

know, I think AIs can be very good thought partners. Like you it's like you talk to the internet, you talk to the most generic person on the internet, but like they're very, very well informed. And it's like almost like a mirror, right? A mirror is not a person, but it's like a very useful tool for understanding, this is what I, you know, what I'm doing. This is where my body is. And you can configure and reflect on that. I think like

We think about different mirrors that try to like reflect different facets of ourselves. And this is what these more local models are should be doing, rather than trying to replace the function of human taste and judgment.

I guess there's sometimes a bit of blurry line between what concepts taste versus expertise or in the domain, right? Like imagine an interior designer, someone who is very skilled at

going to someone's place and telling them exactly what furniture would go well in different corner, right? So there's an aspect of subjectivity. You need to like know the person, what kind of vibe they want in the room and so on, right? But there's really, right now the comparison is clearly a win for humans, right? Humans are still much better than AI at this sort of like supporting someone in making kind of subjective decision, this sort of thing. Do you think there is a risk that AI will soon

become much better in these sort of use cases by just feeding more textbooks on interior designs? Or do you think there's kind of a natural limits because of how model trains or something like that?

So these models are keep getting larger. And it feels like they're getting smarter, but I think my honest sense is like a lot of what we think of as the intelligence is they're just like getting more and more data on very specific use cases. So it's it's almost like memorization, right? You are memorizing like some like really good designers' choices, and you just like pattern match enough to like this is the exact setting, and then you just replicate, right? You just like deliver that.

And that it's still probably not as good as a hu as a human is on like sort of off-distribution type tasks. That said, do I know what the future holds? No.

Yes.

I think this is like practically the first thing they train you as a computer scientist. Like, don't make predictions about the future more than six months out, because it's invariably you get

Yes.

ski bottles.

They should tell this the economists too.

I guess I'm curious if you ever had a, maybe just changing gears a little bit is, know, people often talk about the feeling the AGI, like having an AGI moment where you're like impressed by, you know, the quality of what models are able to do. I know coders often talk about this when there's a new model and they're impressed by, you know, something, a new capability, basically. I'm just curious, have you ever felt something similar in more subjective domains like?

finding an image you found aesthetically impressive or listening to an AI generating song and thinking, how this is actually really good. Or where are you in your current perception of AI when it goes outside of the more verifiable domains?

I mean my like shit moment was when it you know I think I first used Replit, but you know, like like when these AIs just like became really, really good at producing code. that was

Mm-hmm.

pretty awesome to me. Like finally this is here. to be honest,

I I I mean I use AI a lot, as I think you guys do too, and I'm constantly frustrated by it, right?

Hmm.

it doesn't produce the thing that I want to do, it doesn't like doesn't follow the f the stylistic choices I would make. I try to prompt it, and this is I'm talking about Clot here, right?

Mm.

it doesn't do things I want to do.

And maybe I could try my own, but then that takes a lot more resources. It's complicated. I would need to like organize a bunch of data.

It's its own sort of bag of worm, like can of worms, right?

Hmm.

I don't think I'm as like San Francisco as I have not been asset pilled in the way. I have not

Mm-hmm.

had that like, shit, our overlords are coming. we must like I I I don't know w what what verb you want to use, resist or acknowledge or praise

Hmm.

like that. The I don't think I'm quite there yet.

I still think of AI as fundamentally as more normal than extraordinary. that is, it's yes, it's a very, very general purpose and is incredibly useful in a lot of different use cases. but it has a lot of the same sort of fundamentals as existing technology. you need to harness it in the right ways to build it. it's obviously getting better.

But also it's sort of like better in like certain kind of predictable ways. You know, scaling laws are just one of them. orchestration being another. and that in that sense it's like you know, it's it's a technology that you can invest in, build around, rather than something that I'll put it another way, you need to be throwing every single resource you possibly have because it's gonna replace the entire economy and

. .

You know, there is no other game in town. I don't think that's the case. Put it that way.

Yeah, that resonates well. Maybe we already took up a lot of your time. maybe if you allow us a couple of maybe one or two more questions following up on this.

if you had basically a magic wand and you could create a certain capability that is not there for yourself or for the role you're acting out as as in the organization you are in now, what would it be?

If I could magically solve coordination problems. Magic tools

Huh.

solve all the coordination problems, collective action problems in the world. That would make my life a lot easier. You know what? I I would take even just solving Europe's coordination problem. I work with a lot of European entities, countries, institutions.

Okay.

Amen. Amen. I would take that too.

And it I mean I think it's

Without in in some way, like current AI, you know, represents I wouldn't say it's like just European, because there's a lot of countries involved, but it represents this idea of like we need bring countries together, many different institutions together in order to establish something for the public interest. But there are so many frictions that are involved there. there's a lot of coordination that needs to happen. A lot of humans that need to get on board and

you

convince and persuaded.

And then all the technical systems that need to get orchestrated. I would love to be able to just like, bam, let's solve that problem. I don't think AI is anywhere close to that.

Hmm.

Yes. Josh, thanks for being part of that. Thank you for coming.

Thanks.

Thanks for the flying.

Thanks so much.

Thanks for watching. We have got more episodes in the oven, so please like and subscribe to get notified when they are out. And please check out taste-bench.com to read about the taste benchmarks and the creative engine we are building towards.