Will RSI Break the Economy? | Aidan McLaughlin

Before we dive in, I want to shout out our sponsor, Lovable. You know the app you've been meaning to build, the internal tool, side project, or product you'd otherwise lose a weekend to. With

Lovable, you get software you can ship today. [music]

today. [music] That includes an editable codebase you can inspect and change, plus two-way GitHub [music] sync. Lovable also

handles the backend and infrastructure, including managed Postgress, O storage, hosting, payments, and more, so you don't have to wire it all up yourself.

And through Lovable's MCP [music] server, you can create and deploy your project directly from the agents you already use. Turn ideas into software

already use. Turn ideas into software people love with Lovable. lovable.dev.

[music] Now, on to the episode.

The thing I'm really curious about is not like the, you know, 2% to like 4% GDP growth, right? or like you know the GP increase but uh like after like you know what's like the highest GDP you could imagine by like 2050 right like

let's say we're like deep in this like super exponential and we are like you know like mining like the vacuum and we're like you know harvesting black holes that are like outside of the ly cone like whatever like whatever crazy like you know probably banned by physics

stuff so I think like people will go up a lot people go up a lot all right we have a very special guest with us in the studio today it's Aiden

Mclofflin who's a research scientist on Open Eyes core models team and one of my very oldest friends from Twitter all the way back to like 2023.

Yeah, I think you're like one of the longest relationships I've had like from like uh kind of teapot, right? Like

Yeah, absolutely. So, it's great having you here in person. Aiden, welcome to MTS.

Were you guys like both early to each other? Like what were you guys both

other? Like what were you guys both doing when you first I think I was like an investor in Theo at like 60 followers and like at the time like I had like 150 followers or something. It was like we were both very

something. It was like we were both very early on.

So I feel like it doesn't count to say it was like an investor. It was more just like you know we were like yeah too early like a Yeah. Yeah.

Yeah. Yeah.

Preed follower.

Pre exactly. Yeah.

Yeah.

Start. So much to talk about. We were

just talking about like uh posting at labs. You know I

labs. You know I I so appreciate how OpenAI lets its employees post and you know Aiden got uh into OpenAI in large part through posting very quality things about models all the time.

Yeah. I think it's a good um you know, if I were like a manager or recruiter, I think that like I would still pay a ton of attention to Twitter. I think it's like a high signal thing. I think it's kind of hard to fake like really loving technology, really having like good takes. Um

takes. Um Yeah, absolutely.

I think it's like, you know, the the uncheatable eval in some sense. Yeah.

Yeah, totally. I also think like you guys probably see it as a brand asset to have good posters.

Yeah.

Like which I think is entirely the right way to think about it.

Yeah. How much intangible brand value has Rune created for the last Probably a lot.

A sort amount. Yeah. Yeah.

Yeah. Yeah.

Aura. Aura as a service.

But also like I think, you know, it's not just like OpenAI brand value, but I think it like, you know, when you love technology, right? Like when you like

technology, right? Like when you like love AI, when you love the stuff you work on, when you like earnestly want it to go well, not just for like a company's sake, but for like, you know, just to like pro, you know, like fulfill that promise to like an earlier version of you when you were like 12 and you had

like read sci-fi books, right? Like

that's like the the aura I get from like Rune. And I think that like, you know,

Rune. And I think that like, you know, some of that is projected on OpenAI, which I'm grateful for. But I also think that like if some of that's just projected onto like San Francisco or like you know us like as people who like love AI and like you just want these things to go well. Um

Totally.

Yeah. I think we all we all benefit.

Yeah.

Yeah.

So the most important thing in the world right now which we've been talking about basically every day is recursive self-improvement.

Yes.

Specifically like whether recursive self-improvement will happen uh and like what yeah what what form it will take, when it will happen, how quickly it will happen. So, broadly, what are your takes

happen. So, broadly, what are your takes on this?

Yeah, I think like, you know, one thing I would love to do is just like sit down and like write out like 10,000 words my own thoughts on RSI. Uh, I think generally like, you know, Yeah, maybe I'll tweet it or something, but um I feel like sometimes I don't understand

these things until I like am more careful and like put it into an essay or something right?

Yeah, totally.

Um, but I think like my like one sentence to take is that like I align a lot with people who like expect things to go tremendously quick over the next few years. I think that like of you know

few years. I think that like of you know my like friends who are into AI I feel like I have like 75th to like you know 80th percentile like timelines right like in terms of quickness I think things are going to go more quick than

average. Um but at the same time though

average. Um but at the same time though I think that like um I am less sympathetic to like you know discontinuities in both like AI progress

um but also like in economic progress um where I expect like the events of the next few years to just kind of be this like one long um continuation of like kind of things that we've seen uh you

know maybe the last like 5 years of like AI uh maybe like the last like 50 years of like the market maybe the last like 200,000 years of like evolution right yeah but there has to some kind of discontinuity especially in the economy

right like the economy can't keep growing at 2% per year if we get like actual RSI and super intelligence I you know I've heard people make this take before like Darren Asamolu is like you know all of the AI progress Tyler Cowan

has a similar take uh Mark Andre where it's like all of the AI progress will add like an extra half of a percentage point a year to GDP growth which just sounds totally wrong that does seem wrong to me as well I

expect GDP to increase if GDP increases more more than 2% per year then you have a discontinuity so I think Like you know I I was chatting it was a while ago with like Leopold about this and I like kind of asked him like you know what's your

argument against like you know the GDP staying constant and he had like the best like one sentence description I think I've ever heard and he's like you know people just aren't looking at GDP on like a long enough time horizon right like what you really need to do is like look at GDP since like you know 200,000

years ago since like the emergence of homo sapiens since like you know kind of we evolved out of like the savannah right um and yeah like maybe that's a cheat answer to be like oh this is like just kind of one continuous trend that has happened for all of human history

right um But I do kind of think that like you know I don't know if it's like Bostonramm or who kind of like regressed this best but like you can just like really zoom out and draw like a hyperola on these things. And I think Luke Muhouser was it? Yeah. Okay. I need to get my

was it? Yeah. Okay. I need to get my history better. But um

history better. But um yeah I think that like a lot of this is like best regressed by a super exponential and long enough time horizons.

Yeah. And also like it could just GDP growth could just lag behind model capabilities like like it doesn't necessarily you know maybe this year we won't get like crazy

GDP growth but like in like 5 years or whatever uh you could actually start to tangibly see it.

Yeah. I don't think it will last.

I think there's only like if you have like actual super intelligence in the world that is actually a country of geniuses in a data center that is actually capable of like doing any cognitive task better than a

human. Um like it won't be long before

human. Um like it won't be long before you see GDP go up a lot. It's like the argument that I've heard, the best argument I've heard is from Ryan Greenblat of Redwood Research on this where he's like, um, if you talk to an economist and you're

like, there's going to be 100 million new high-skilled immigrants to the US, they're all like Nobel Prize level intelligent and they all are capable of working much faster and harder and longer hours than any human and they're

infinitely adaptable and they will work in whatever field you tell them to. Um,

and so on and so on and so on. Of

course, the economist will say, "Yeah, GDP will go up a hell of a lot." But if you tell the economist that it's an AI, then it will be like, oh no, GDP will stay growing at 3%. There's bottlenecks

and you're stupid for suggesting this.

I I do think that I personally like expect a GDP increase. Um, you know, I think it will like start slowly and then, you know, go all at once, right?

Um, but you know, like the thing I'm really curious about is not like the, you know, 2% to like 4% GDP growth, right? or like

you know the GP increase but uh like after like you know what's like the highest GDP you could imagine by like 2050 right like let's say we're like deep in this like super exponential and we are like you know like mining like the vacuum and we're like you know

harvesting black holes that are like outside of the ly cone like whatever like whatever crazy like you know probably banned by physics stuff. Um

yeah it'll be like GDP by planet and not GDP by country.

Yeah. Yeah. So I think like GDP will go up a lot.

GP will go up a lot. And I think that like, you know, if a future economist were to look back, I hope that they like kind of, you know, my intuition is that they would expect this to look pretty continuous still. Um, you know, they'll

continuous still. Um, you know, they'll be like, "Oh, wow." Like, things start accelerating. Um, but, you know, things

accelerating. Um, but, you know, things have always been accelerating, right?

Yeah. Do you think that this GDP growth like when we do see it will just sort of manifest in like certain countries and some will just show absolutely zero still?

That's a really good question. Um

I you know one thing that like I have been more concerned about recently is kind of like the you know the centrality that America has on like super intelligence right?

Didn't we talk about this the other day?

We might have. Yeah. Um

but yeah like I think you know a lot of the like economic plans I hear about like distributing AI benefits. A lot of the like kind of you know think pieces from like people on like the AI scene and like that stuff here. um a lot of

like you know maybe talologically like the the government plans that I see like laid out for like making sure that this goes well um whether by like Bernie Sanders or like by kind of like the Republican party um you know obviously

focus on America right and I think that like the world is much larger and um you know like the kind of maybe like you know UN Roosevelt and me like does want to see this kind of be more

international if it can yeah AI 2040 talks about this you know they have the US government paying a citizens dividend to every American and then after paying a smaller div

and America gets like hyper wealthy.

I mean first of all we have like immigrants from every country in the world who will happily give like large remittances to their families. Second of

all we have like foreign aid and charitable institutions that will do the same thing. And then third, we just have

same thing. And then third, we just have like normal trade with these countries and like we will help provide them with like infrastructure in exchange for all kinds of things like you know all of the super intelligence in maybe not all the super intelligence in the world but

super intelligence up to a certain level is not going to figure out how to like grow coffee beans in the United States because like we have nowhere outside of like Hawaii they can grow coffee beans.

So we'll still be trading with Ethiopia for coffee beans. Maybe coffee will become more expensive and we'll take Ethiopia's coffee in exchange for our super intelligent robots and it'll help develop their country. So, I'm not like terribly worried about you worry about

like inequality, but I'm not worried about like the third world remaining in poverty after AGI.

Yeah, I think it's hard to imagine like if AI goes really well, right, and like we don't see like uh catastrophic mis land or something that like truly like any human is actually like impoverished on long time horizons. I agree with this. Yeah.

this. Yeah.

Yeah. Yeah. Um I would love to like read like a more like uh you know let's say like in the next like 5 years like economic like forecast about like what happens though because I do think that like you know it matters a lot kind of

the the political like temperature right of like the globe um as these things are like taking off right well I mean like the you know forgetting about misalignment like one real bare case I could see there is like um you

have like full automation of labor and then the wage for human labor falls below subsistence level um and then you know people are reliant on on handouts basically.

Agreed. Yeah.

Yeah. So I guess talking about recursive self-improvement a little bit more. Um

what are I guess your like aspiration how does it go but also what are some of your fears around it? [clears throat]

Yeah I like you know I tweeted this recently so sorry to like regurgitate my own stuff but like you either like die a capabilities researcher you live like long enough to see research.

That was a good tweet. Yeah. Um, so I think like for me and a lot of other people at OpenAI, um, our jobs which had like previously been like kind of a smattering of mostly like you know some good mix of like capabilities alignment

research are in my opinion like starting to like move toward uh alignment research, right? I think that we are

research, right? I think that we are like all kind of like getting all hands on deck here. Um, I think that like it's, you know, as we've seen the last few months like important for these

things to go well so that we can earn public trust. Um, so you know, what are

public trust. Um, so you know, what are my fears? I think like maybe I don't

my fears? I think like maybe I don't have like super creative fears. I think

I like fear like just the normal things some people fear. Um, I am like very optimistic though. I think that like you

optimistic though. I think that like you know it seems like there's been a real like coming together moment for the industry uh for San Francisco.

Um, certainly like my colleagues and I like you know have never felt like so locked in, right? Um, yeah.

This is uh one thing I think the safetyists get wrong like they're so habituated to this world where like people just don't care about AI alignment. you know, nobody's on the

alignment. you know, nobody's on the ball in EGI alignment and uh nobody like, you know, we're we're rapidly like speeding towards the apocalypse and nobody's doing anything. But it seems like now like labs are really starting

to like uh shift gears into alignment, especially OpenAI.

Totally.

I think uh at this point anthropic might be focused more on alignment than or rather OpenAI might be more focused on alignment than anthropic right now.

This seems to be based on vibes. Um, but

you know, like I I think I chatted with you about this recently, Theo, but like I forget like what race car driver said this, but like you know, the fastest way to like take a turn uh is to like use the brakes sometimes, right? Like

otherwise you literally just do like you know kind of accelerate in a straight line and like don't go anywhere. Um we

can check the physics later but [laughter] um I think that like you know if one thing that I think a lot of people get wrong about like uh arms source dynamics uh is that like it makes

sense that like if you are in you know a tight race between countries um over something like the you know atomic weapons um that like kind of your personal like safety goes out the

window, right? Um where like you are

window, right? Um where like you are like kind of motivated to do these like kind of crazy scenarios um and like kind of throw caution to the wind and this of course like ends badly. Um but for like public companies right like I think this

is like uh interestingly different right like I think in some sense like public companies before AI are like pausing all the time right like you know like there are car companies who are like we need to like fix this like engine issue everybody lock in for like the next few

months let's really like you know kind of get a recall out um you know this like has happened I'm sure with like every major tech company.

This is such a good point like pharma too.

Exactly. And I think that like you know it is very rarely the case like that I've heard a tech company like you know or heard like a company in general be like ah if we don't fix this like major like legal you know or like uh

problematic like thing here we like you know a competitor will just like bypass us like no like you know you just you have to do these things otherwise like you know the public doesn't trust you otherwise like you know revenue goes down right because people don't want to use a product they can't trust

um or like you know there are like legal consequences right so basically in like one sentence I think that like it is the nature of uh good companies to like constantly, you know, have their like

foot on the gas and then sometimes off the gas in order to like make things go smoothly right?

Yeah. Misalign models.

Yeah. Exact. Exactly. Alignment is not mutually exclusive with like profitability and scaling.

Yeah. I think they're actually quite aligned like that.

Yeah. I mean, I'm a capitalist, but one of the most damaging takes ever is that like profitability is inherently bad.

Yeah, agreed.

We'll continue monitoring right after this message from our sponsors. 11 Labs,

AI that communicates at human level across every channel and modality.

11labs.io/mts.

Scale your startup on Neon. Millions of

developers and startups have already chosen Neon for their backend. Start on

the free plan or get up to $100,000 in credits for your startup at neon.com/mts.

neon.com/mts.

Special thanks to our sponsor Kong, the AI connectivity platform. Connect APIs,

LLMs, agents, and systems with serious security and governance. Konghq.com.

So, what was actually the internal vibe at OpenAI to the extent that you can sort of share after the like hugging face incident? Cuz you mentioned, you

face incident? Cuz you mentioned, you know, like I I'm sure that this aligned everyone. This like had everybody, you

everyone. This like had everybody, you know, mutually like have put so much care into like what they were doing and just kind of changed things.

Yeah. I think like you know, I am really grateful where like the people I've worked with um even if they don't come with like a traditional like safety background or even if like they were not on like you know less wrong and like you know 2021

um they do just all like very homogeneously care about things going well right um so in some sense like it wasn't uh like I I think it's not like before people were like you know super chill

and now they're like people like you know jumping out the windows and stuff.

like um you know we have been like locked in and I think we like continue to be like very focused on uh trying to make great models and like you know a really important subset of this is

making like aligned models um so what is like the vibe I think you know it's intense uh but it's always been intense um yeah how much are openi researchers right now

do you think being meaningfully uplifted by model capabilities like we we did talk about this the other day but not on camera yeah um a lot like I I think you know one great thing about open AAI like you

know join open AAI but if you join open AAI you uh get access to like as much codec as you can eat right like it's kind of the all you can eat codex buffet

um and you know like truly now I feel like the manager of like a pretty competent team um I think that like you know you mean team of codeexes a team of codeexes yeah exactly and of

course like if you're using like 5.6 six on like ultra like you know my team of codeexes have themselves like team of codeexes right and like multi- aent um yeah that's part of the and I kind of just feel like this is actually like a great metaphor for like

um you know like kind of internal use in general right where like it might have been that like in after GB5's release in like 2025 like it was like a really bad intern where like you know I was kind of like looking at what it was up to like

every like 10 minutes like a sucks I can only have really one thread at a time um you know and like as capabilities get better the amount that you have to like check in the amount that you have to like course correct um decreases. Um but

as that decreases though, right, like your own ability to like spend more things in parallel increases, right? So

I don't know if like the fixed amount of time that I spend per day like checking in on my codeex like my codeex has really changed. Um but the amount that I

really changed. Um but the amount that I can get done in a day has like massively changed if that makes sense. Well, when

are we actually going to see this in like macro productivity statistics? You

know, presumably, you know, enterprises all over the country have access to models only what a month or two behind what open eye people have. And yet like labor productivity has not really increased due to AI yet. Even still,

even now, even now that we have agents, even now that we have multi- aent systems, even now that it's so cheap, it's so abundant. Um it seems like yes, they automate some tasks, but then the bottleneck shifts to other tasks and

productivity doesn't increase that much and you really need like uh systems that are capable of automating much more of jobs to like meaningfully increase macro productivity.

So what's up with that?

Yeah, I I don't know. I would love to like chat with uh people who like you know have a stronger economics background than I do. Um

my like really rough intuition on this and I I'm curious for your guys' thoughts is just that like uh sometimes we lack like the precision in the tools to like measure the effects of these things and like short time horizons, right? That like it is

horizons, right? That like it is actually possible that like there is meaningful uplift and if we like put together the right plot like we'd be like wow like look at that. Um but this is hard. I think it's like you know it

is hard. I think it's like you know it takes a while for these things to get published. I think that like you know so

published. I think that like you know so far like institutions that are really like looking at this for AI are often like anthropic open AI meter like kind of like you know already AIcentric organizations which like makes it harder

to um kind of gain credibility um right like you know we're selling products so obviously it's like uh hard for us to like publish kind of like really um uh like widely accepted like academic

economics work. Um,

economics work. Um, ironically opening I just did this. They

published like chatbt in the enterprise.

We talked to David Holtz who was one of the co-authors about it.

I like I am very grateful for economics team. I remember like Tom who worked at

team. I remember like Tom who worked at me who works at meter now too is like yeah exactly like uh such a genius. I

like feel like I learned so much talking to him. Um but I think like for people

to him. Um but I think like for people outside the labs right to like start talking about like kind of the productivity increase from AI I think it will just take a bit more time unfortunately.

Right. So one take I hear a lot these days is uh people who are interested in AI and concerned about making AI go well should not work at labs. Instead they

should work at like independent verification auditing orgs like Meter like uh Avery like epoch maybe. So you

know you still work at a lab presumably you think that more people should work in labs.

Yeah. Uh why do you think that people think that you should work at like meter or epoch? I'm curious.

or epoch? I'm curious.

uh they may think that like joining a lab will uplift capabilities and at this point we shouldn't uplift capabilities anymore. I don't agree with this take

anymore. I don't agree with this take but I'm curious about I feel like it's like a checks and balances sort of thing.

Yeah, that kind of thing.

Yeah.

Yeah. Honestly, I'll like recycle what I said a few minutes ago where I feel that like in order for you to like make a great product, in order for like people to love like the next version of CEX or something right?

Um it has to be aligned, right? And like

there's immense like economic pressure uh for like these things to to be aligned, for them to like you know not deceive us, for them to like you know kind of like intuit what you mean without you having to say it which I

hope has some like you know really um helpful kind of like generalization as like capabilities get better. Uh so yeah like you know I just think that like what I do is increasingly alignment research and I think that if you are somebody who does care about things

going well um you know it's it's an interesting uh it's an interesting trend right that like you know labs are doing increasingly more alignment research like maybe you should go to the place where like the rate of alignment research is increasing if that makes

sense.

Yeah. Do you see like certain uh talent like pipelines like for example like you mentioned capabilities researchers turning into alignment researchers like are you seeing basically people with a certain background or specification

pivot to a certain sub field?

Yes. [laughter]

All alignment. Um yeah, I think like again I think uh you know it's it's hard where like if you were somebody who worked on like a really niche product like capabilities like a few years ago as capabilities

have gotten better and as it's time to like bring your like you know kind of theoretical or mathematical insight like into the product like into the real world and you try to like scale it up.

Yeah. Um often that process like involves a lot of alignment research, right? Like then you have to like start

right? Like then you have to like start caring about your data. Like then you have to start caring about like um you know the experience for like a user, the experience for like you know humanity, right? Um so I think that like it is

right? Um so I think that like it is maybe the nature of capabilities research that it like often graduates to alignment research.

Um but I also just feel that like you know people uh you know like whereas like in 2023 2024 this is maybe more theoretical now it is like less theoretical. I think that like

another part of this is also just people like wanting to, you know, people like having a new kind of fire under their ass here and like wanting to make it go better. It's a good question. Yeah.

better. It's a good question. Yeah.

Yeah. Yeah. I think [clears throat] something interesting that people don't think about is like if there were to be if somebody were to literally cook up like a you know biological weapon in

their backyard using say like an open AI model or some other labs model and then news about that were if it were to like you know breach containment or news about that were to surface then there

would be so many pressures for that company to absolutely change everything about like the way that they do things or for people to start new companies. So

I think that I'm less concerned about like like you know alignment not happening in the like especially obviously there's you know potential for things to go wrong but there's so many

mechanisms for basically selfcorrection.

Yeah.

I'm curious if you see it that way as well.

Definitely see it that way. Yeah. I

strongly agree. No.

Yeah. [laughter] Yeah.

Yeah. What else uh you know what is it like to work at OpenAI and what is it like like what's your background as well? like how did you even get in in

well? like how did you even get in in the first place?

Yeah, I uh I don't know if you I'm not sure if we've like chat about this too, but it's kind of like a fun fun thing, but like back in like 2017 2016 um I was like really into chess. Um and you know

my like interest in chess like quickly moved to like my interest in like chess engines. Um at the time like you know

engines. Um at the time like you know the best chess engine like the best chess playing computer in the world is called like Stockfish. Um, and you know, many people I'm sure like know what Stockfish is, but it's just like this, you know, especially in 2016, like this

like, you know, grinded out like billions of positions of chess to like tell you what the best move is like. And

it's so so accurate that like, you know, even in like 2016, uh, this like Stockfish just like this brute force kind of like algorithm could like beat Magnus Carlson if it were like on your Apple Watch, right? Like it's just really impressive like piece of

software. Um, but of course like at this

software. Um, but of course like at this time, this is when Alpha Zero came out and like taught itself how to play chess by playing against itself. Um, and like you know I I hear some friends say that

like oh the thing that got me into AI was like the uh Tim Urban Post, right?

Like you know about like kind of uh super intelligence or like maybe it was like Boston or like whatever.

Um for me it was like this video by this like one British Grandmaster looking at the moves of like Alpha Zero. It was

just like you know if you look at the YouTube video I can send it to you guys.

It's just like a chessboard and he's like a this is so beautiful. And it like felt to me at the time as if like um you know I always wonder growing up what it would be like to meet aliens uh like super intelligent aliens and have them tell us that they play chess. And

there's this moment in like 2017 2018 where I'm like wow like I now I know know you know what I mean this is crazy.

Um so anyways I like my my interest went from like chess to like chess engines to like just kind of reinforcement learning generally. Um and I wanted to like help

generally. Um and I wanted to like help work on these engines. I did like a bit of work anonymously with um Leela Chess Zero which is like kind of the open source like version of Alpha Zero. Um

and yeah like I think I I was really interested in RL then uh I continued to be like an RL researcher now. Um I had some other stuff uh in the interim. I

went to college uh worked at a startup I was you know really excited about. Um

but you know RL is where my heart is. Uh

and I I think that like in the next few years we'll see uh increasingly like more Alpha Zero moments like that too. I

think be really fun for everybody else.

And you forgot one thing which is Aiden Bench.

Yeah. Yeah. Aiden bench. What what

happened?

What are we getting more?

I you know unfortunately or fortunately all of the Aiden Bench authors now work at Anthropic or Open AAI. So we I guess like good Aiden Bench legacy but we need to if anybody out there like is watching this or on Twitter um reach out to me.

I'd be happy to like fund this personally and like get this back. I

think it's like important that the results like aren't published by a lab because I think that like you know it's it's good for credibility for these things to be like independently kind of worked on. I think it's important

worked on. I think it's important obviously to stay open source. Um, but

for those who aren't familiar, Aiden bench is like a benchmark in the end of like 2024 um that measured like model creativity. Um, and the way that it like

creativity. Um, and the way that it like worked is um we would give models at the time this was like 01 01 preview um you know the good old days. Yeah. Exactly.

Like s 3.7 my kids about prime1 preview days.

Exactly. Yeah.

Um you know let's discuss is like how great of a jump one preview to1 was but conversation for another day. Uh but on Aiden bench though like what we do is we'd ask the models a question like how do I solve traffic in Los Angeles and then we would have them like give an

answer and one of the answers be like ah like you know put more lanes in your freeway or something you know I mean maybe not a great answer I'm not sure.

Um but then we would ask for another answer and be like okay like that's one answer like what's another example and then what's another example and we would do this like hundreds or thousands of times. So the models would have to like

times. So the models would have to like be forced to come up with like like a thousand unique solutions to like traffic in Los Angeles. And as you did this, right, as you like kind of like kept, you know, forcing them to like

spit out answers here, um, one of two things would happen. The first is that like they would just say something like purely nonsensical, right? Where they'd

be like, oh, like a great way to solve traffic in Los Angeles is have everybody like drive a flying car. Like, okay,

that doesn't really make sense, right?

Right.

Um, but another thing and a more common thing is that they would just repeat themselves. They'd be like, you know,

themselves. They'd be like, you know, their 500th answer would just be to like add more freeways again, right? Um and

it turns out that like you know the models that could spit out the most like unique answers without like collapsing and saying something crazy mode collapse. Exactly. And also not

mode collapse. Exactly. And also not like repeating themselves which is also I think like some part of mode collapse.

Um you know really aligned with I think a lot of people the times like intuitions for like front model capabilities right the models best on this like did feel like the most creative models. They did feel like the

creative models. They did feel like the most capable models.

Um so I think you know had like a good run. Uh I I think that it's important to

run. Uh I I think that it's important to like continue to measure like model creativity and diversity. I think that like as the models, you know, take like a larger part in our lives. I think like it's important to make sure that um the

model that you talk to isn't like just doing the exact same thing as like the model that like you know another user is talking to, right? Um I think this could lead to like a mode collapse world, not just mode collapse model.

Yes, totally.

I ranted. I apologize. Yeah. Yeah. Yeah.

Absolutely. Well, we're coming up on time. Oh my god. so much stuff that we

time. Oh my god. so much stuff that we didn't even get to like you know what you expect in the future after AGI. So we'll have to do a round two at some point.

Yeah, totally guys.

Great having you on Aiden. Thanks so

much for coming to the MTS studio.

Good to see you guys. I'm This is awesome. I'm grateful

awesome. I'm grateful and we'll be right back.

See you guys.

A huge thanks to MTS sponsors. Blitzy,

autonomous software development for enterprise code bases. Ship 5x faster.

blitzy.com. [music]

Adqu make your brand a billboard. Out of

home advertising as easy to scale as digital. adquick.com.

digital. adquick.com.

Arena, measuring AI performance in the real world. Arena.ai.

real world. Arena.ai.

Support for the show comes from VCX, the public ticker for private tech. The US

stock market started history's greatest wave of wealth creation. From factory

workers in Detroit to farmers in Omaha, anyone could [music] own a piece of the great American companies. But today, our most innovative companies are staying private longer, which means everyday

Americans are missing out [music] until now. Introducing VCX, a public ticker

now. Introducing VCX, a public ticker for private tech. [music] Visit

getvcx.com for more info. That's get vcx.com.

Carefully consider the investment material before investing, including objectives, risk, charges, and expenses.

This and other information can be found in the innovations fund perspectus at getvcx.com. This is a paid sponsorship.

getvcx.com. This is a paid sponsorship.