Anonymous OpenAI researcher Rune on Codex, open source AI, and why LLMs are upending software engineering

probably have fun on there. Um but uh be careful because you might just uh wind up in infinite jest world. Anyway, we have our next guest, Rune, the anonymous poster, the world famous Rune in the TBP and Ultradome. There he is to the stream. How are you doing, Rune? Hello. Hello. I'm good. How are you? We're fantastic.

at long last. What I mean, it's a printing your posts. We've been printing your posts for months. Uh so, thank you for everything that you do for the timeline. Always providing insight and levity. You guys mentioned me four times this week and I was like, I got to come on just to round off the week, you know.

It's uh close it out. Yeah, John. Yeah. And thank you brothers. Thank you for walking around the and and you know reminding the researchers that you guys are you know holding up the global economy. Yes. It really is important work what they're doing and also you to just make them remember.

I mean I know you guys comment on this like all the time but like most of the growth in the equity markets is just AI you know for the past like three years. Yeah. There's like really nothing else. So, you know, uh, you got to not make mistakes. Got to keep pumpking out great models. Yeah. Yeah.

Please don't make mistakes. That's all that's all that's all we're But but yeah, I mean, we we talked about it earlier today. The number of companies that are now, you know, uh, barnacles on the whale of open AI is uh, the entire global economy is sort of Yeah. But you can even get you can even get more specific. Yes.

And um, you know, the the oracles, the core weaves, you know, you can make a really long list. Yeah. Uh how are you feeling generally? The the the timeline was extremely split uh this morning and yesterday. Uh a lot of incredibly great vibes around uh Chat GPT Pulse. Uh less great vibes around the Vibes app.

Uh are you happy with the direction that the productization of AI is going? Are you happy with the the roll out and the uh the actual implementation of these models so far? Uh how how have you been processing the way these models are making their way into every everyday lives? Yeah.

Um I don't know like okay the chatbot medium that's one thing it's uh I I think the pulse is like a fine and good idea that makes a lot of sense. Yeah. um vibes. I don't know. But like the the the thing I am like manic about is Codeex and you've probably seen me tweeting about it like three times a day.

But like more than that. More than that. Yeah. Uh like Codeex and I guess like Claude Code are like genuinely amazing. Uh like I don't want to like you know glaze the company's products too much but like it's like I don't understand how I did software engineering before this.

I have like 20 terminal agents open and like writing like five different things at once and I have an idea and I just like open up a new terminal and I just like get it working on it and it's like a completely different way of life than anything I was doing before which is like you like linearly slog through one script and then like there's like three or four bugs and like I don't know um like it's clear to me that this alone is like completely upends the software industry.

And you know like we're talking about massive data center capex across the industry and like what is it going to be spent on and like even the inference alone for this kind of thing is like first of all like super expensive and then like super in demand I think once it like really percolates through the economy.

So yeah I mean that's a thing that is uh fueling my mania these days like it it uh reignited my you know fuel agi fire. Well, yeah. So, you said completely upending the software industry. What what does that actually look like in your mind?

Because, you know, one thing is if developers are getting higher output, developers are going to be fine. A lot of people other people have been like, you know, 90% of software engineering will be automated, but maybe we do 10 times as much. There's the same amount. Like, what's the long-term view for you? Yeah.

I I like really don't know about the jobs or whatever. It's like extremely nonline nonlinear and hard to predict. Yeah.

I feel like there's so much um you know like I see the product managers at my company, they're like vibe coding useful ops all the time and like people who are not even engineers are becoming like programmers because of these things.

And I'm sure that's the case for like the most junior level engineers who are like bright are like now they're like way better than they were before.

I remember my first year of programming and I was like running to my mentor for like uh you know like how do I run this whatever basic thing and like they just don't have to do that anymore like they have to they can avoid all that embarrassment.

they have this like amazing agent or like the chat GBT itself or whatever it is. To me, it would seem like the productivity of an entry- level engineer is like just way higher than it would have been like 5 years ago. Um I can't, you know, speak to jobs impact in like the two years or three years or whatever.

What about in terms of like consumer impact? I feel like you still hear people complaining about like, uh, my United Airlines app is not as great as it could be.

And it feels like in a world where software engineering gets way easier, more reliable, higher leverage, like we should start seeing higher quality software diffuse through the everyday lives of people that interface with technology.

Does it feel like that's happening, going to start happening, or is there something sort of intractable about uh an airline booking app that just can't be solved? Yeah. I mean, like, first of all, I don't trust people's vibes on this.

Like, I don't actually think they're paying good attention how the United app looked like three years ago or whatever. Sure. Sure. I I'm sure it's actually a lot better. Yeah. Uh like I've seen the inflight displays on United and they're like like what is this? It's like space age. It's way better than it was before.

Um, and okay, I I don't know that United engineers are using any of these tools. Maybe they're not. Maybe their company doesn't allow them. Maybe they use like some third party sourcer. I have no idea. But, you know, that's what I was saying.

Of course, these tools I have no idea what time it takes for them to percolate through an economy or like find people in like the, you know, like the the non- tech Fortune 500s are using them or anything like that.

But I I from everyone I've spoken to that works at like say like GM or Ford or something, they're like, "What's Codex? " You know, it's not there yet. Of course not. Um, I hope it will be, but the fact that it even exists is just like exhilarating to me.

Do you think Codeex has the potential to be a hit consumer product and get, you know, end up eating into uh the market share of some of these uh new vibe vibe coding, you know, companies?

Um, of course, but also I think they're very excited like even just the cursor people because I don't know like there's like a competition between OpenAI and like Anthropic and uh that's probably good for them. I I don't know.

Um it's hard for me to say but I'm pretty sure that Curse is killing it and they're very excited. Yeah. How are you using the various products? It seems like you're using uh codecs all the time. Are you using codecs on mobile as well?

Are you in sort of like a like you know wake up in the morning and you have an idea or like a shower thought and I find myself doing this with uh with just GPT5 Pro queries or deep research queries. I have some idea, I fire it off and I just wait. Uh but I'm not writing a ton of software.

Are you having a similar experience that I'm having with just knowledge retrieval in code generation where you might be firing it off like when you're in the bathroom and getting dinner with someone and you just have a second you just fire something off?

No, I'm just like not a hardcore enough engineer to be doing stuff like that. But I I uh I definitely use like Five Pro or whatever all the time just to whatever random idea comes to my head.

I wake up, I'm like researching some I don't know some statistic or like setting off some query like what percent of the economy is like software engineers or how much how much do you think you personally cost OpenAI monthly like how much are you a lot a lot I I I can't say but a lot a lot uh what about other other use cases I mean we saw that a lot of people are using um these models for like uh like life coaches and just people chatting and I've never really gone down that path.

I don't know, maybe I'm just busy or I talk to Jordy all day long, so I I don't have a lot of room for that. I I hit knowledge retrieval constantly like the GDP of some country. Yeah, give me the full history of it. Build a I did a deep research report yesterday on the capital of California.

Thought thought thought for 59 seconds. Yeah, it it ran a hundred searches. I I hit that stuff all the time, but I but I I I haven't gotten in any sort of flow of of asking for uh interpersonal advice, life coaching, therapy, any of that stuff.

Have you had any luck or is that just like a is that a skill issue or just personality or certain type of person that uses things that way? Like do how do you think about that use case? Um, I've definitely used it in like a tough corner. Like what decision do I make? And I like have it discussed with me. Yeah.

Uh, I don't know. I'm probably not a typical user cuz I like read what it says and I'm like, "Oh, this this one seems kind of like and this is like a great point and I like weigh it and ignore it and whatever.

" But uh yeah, I think actually new sonnet was an amazing model um for this and that this was uh what they called claude 3. 5 sonnet new. Yep. Um, and I think that kind of invented this genre of like I as Claude for everything in my life.

And I think like the latest GPT models can also do this, but like um Yeah, there was like memes going around about the the Claude boys back then who just like Yeah. Like literally everything they do, they'd run it through Claude. Yeah. It's uh it's a bit weird for sure.

Like are you you know, are you being puppeted by an AI? like it's it's a bit loss of control vibes. It's uh I feel like my interactions with AI are getting less back and forth over the last two years. I used to chat back and forth and ask a lot of follow-up questions.

And I feel like now if I just talk to GPT5 Pro on voice mode for a minute and give it a bunch of context, it's going to come back with a very thorough answer that I might have one or two follow-ups on, but I'm not really I certainly don't have this concept of like there's one chat that's like fine-tuned at this point on this particular that I oh yeah, my economics chat I go over here when I'm asking an economics question.

everything is done at just like a new chat router level basically. Is that kind of mirror your experience? Yeah, I I I do think um the open models are built for that. It's like uh single turn think really hard get you the right answer type of thing.

What about what about like the flavor of the uh of the responses like with the GPT3 API? It it felt like if you if you prompted it with like the green text, you'd get a green text back. And it felt like there was a lot of uh a lot of variation in the style of writing. And now we're seeing it.

Maybe it's just the way people are prompting, but it feels like there's a little bit of like a a consolidation in the in the textual style, the stylistic flourishes, the it's not the it's not that, it's this, and the mdash and all of that.

like it feels very much like if I opened up your app and asked you and asked the same question, I would get the same stylistic flourishes.

And I'm wondering if that's something that you think will continue or maybe that's just a temporary moment of time and we'll wind up in the future where someone's talking to that the vibes of Claude 3.

5 maybe in a different completely in a completely different model but they're they have those vibes over there and then someone has 40 vibes over here but they're not talking to literal 40 or literal 3. 5 but they've kind of landed in a local minima of this is the tone that they like for all of their interactions.

Um, yeah, these are great questions. I mean, like the the GPD3 model to me was like my first feel the AGI kind of like hair raising moment. Um, I remember playing with it and just being like, oh, this thing's intelligent in ways that are like, yeah, just like non-trivial, creative, etc.

But people are also like too rosy about like the base models. they they just don't uh they're like extremely uh random.

They like they don't even have a concept of uh you know like we talk about hallucination but what the base model is doing is only hallucination like it's not uh you know maybe one out of 20 completions will have a coherent uh like front to back like sensical text and there are people who can do like magical things with it but there's a reason there was no like GPD3 moment and there was like a a chat GPT moment where like the like it took off among consumers and same with like claude and and whatever right but um there is like at all these companies there's this um what do you call it like a path dependence of post training so like the models sample and you collect comparisons between model samples and you're like um I shouldn't talk too much about post training, but like the point being like you are very reliant on the previous generation of models to train your next model.

And so there's like some style burn-in that goes on and like the late Claud resemble early club and like the late Grocks resemble early Grock and so on.

Um, and yeah, like they all have some similarities probably because of like well I mean they're all training on the same giant data set of the internet more or less and they also have like these common data sets. I guess I just wonder if like there's there's demand from people for more variation in vibes.

I was uh I was asking Mark Chen about this where um like I was like everyone was saying GPT 4.

5 had a big model smell better like flavor to it better writer uh but deep research was the one that could get you all the results and I was like so is the best practice for me now that I should like run a deep research report and then take that to GPT4.

5 and say hey rewrite this better and and I'm wondering if there's some world where there's just demand for like I like a lowercase person to chat with. I like really punchy. I mean, I guess you can kind of put that in the prompts, but I'm wondering if it's if that's like at some point going to be an emergent property.

Um I think you should try the file thinking model uh if you I have some use case like that. But yeah, it's the same thing basically. Does that work for you? Yeah. Yeah. I mean, it's not it's honestly like I'm I'm fine with the with the I like the uh the the flavor and the vibe that it comes back to me with.

Uh it's just it is impersonal, but but that's exactly what I'm But but that's exactly what I'm hitting it with. And I want that. I want it to be a an assistant. I want it to be a helpful assistant. I want it to be like a textbook. I want bullet points. I want m dashes. I want clear analogies. And so I'm very satisfied.

So, I'm wondering if m I mean maybe it's just a revealed preference versus stated preference, but it seems like there's a lot of stated preference out there that people are sick of M dashes and if it's not this, it's that. Uh, and it's certainly like a tell that someone used an LLM.

But um but I I was just I was just thinking more about um like how if there really are people out there with the root with the revealed preference with the true preference for something and and and will the models eventually be able to be many things to many people show a different face to every single different Yeah, absolutely.

I mean, I think the end stage is like extremely personalized models that speak to you guess even before you even tell them what it is that you want the model to sound like. Yeah. Uh and I see no real reason that's like not possible. Yeah. Just a matter of doing it. And the what else?

Do you think um do you think memory will start to be like feel like more of a moat as uh products like Pulse get more popular? It feels like this is the first time that uh you know I I I don't feel Hold on. Can I ask you guys a question please? Yeah.

When I saw the first episode it was called the Technology Brothers Podcast Network. What happened? It was just technology brothers and and then we had a big problem which was that which was that our community teapot and and the and the tech community broadly honored that we were calling it technology brothers.

But when journalists would write about technology brothers, they would introduce us as John and Jordy the tech bros and they would shorten it which was massively disrespectful.

Uh, and so we and once we realized we were doing our lives life's work, we did not want to spend the next 30 years everywhere we go being introduced as the tech bros. And so yeah, that's that's disrespectful.

If they had called you Jordy and John the Technology Brothers, that's we would we would have kept riding with it, but it was a big problem. That's the battle we lost. But uh Okay. Uh I believe you credited me with that one. You pull up a tweet on the the first episode.

Yeah, you were you were one of the first people to to truly, you know, expand Tech Bros into Technology Brothers. So, thank you for popularizing the term. What do you think? Your your change. Yeah. What do you think of uh what Thinking Machines has been putting out lately?

Um, I think it's really cool that, you know, like new labs, they're incentivized to speak very openly about their research cuz, you know, like it's a great recruiting mechanism.

You're like, there's some cool research you're doing and like, let me come join and then like I don't know, maybe I guess Facebook was pursuing this strategy a while back uh or meta or whatever they're called now. They they would uh they open source everything.

They write a lot of public papers and and it was a big draw for people. Um I haven't read the latest paper but I read the last one. It was something I I had been wondering about for a while and like was a common question in the industry about uh determinism and non-determinism and so on. Yeah, it's it's good stuff.

It's good stuff. It's good stuff. I don't know like it's great stuff. I don't know specifically what they're working on. I've heard rumors but I won't I can't I shouldn't talk about it. What about hardware form factors?

What what's uh uh you know speak generally, you don't have to talk about what you guys are working on internally, but what what's exciting to you uh in hardware broadly? Do we need new dedicated devices uh or is the phone uh good enough? I don't know. I don't have the answers to that.

I like genuinely don't think about the hardware side very much. Well, that's that says something in itself. Yeah. I mean, I'm pretty happy with my uh chat GBT on my phone, but I What was on your pulse today? I don't think I checked. Mine was a check right now. Check right now. Mine was Mine was shockingly interesting.

Pulse is a window into the mind. It was It was It was extremely interesting talking about bottlenecks and compute revolutions and retired coal plants powering AI campuses. And uh it even had one I was searching for different options for bonded cellular networks.

So if you want to bond a Verizon and AT&T uh like SIM card together to get like a really powerful Wi-Fi hotspot on the go for live streaming obviously and it had a whole deep dive on that and it's like yeah I'd searched for that but I would love to have that resurface to me in a different format. Uh, right.

And so I was I was very pleased with the results. And it also felt like deeply maybe it's just like the way I've been using the product, but it felt deeply anti-brain rot. It felt like, oh, these are these are these are articles.

I don't even know what you would call them, but summaries that um certainly push me to go deeper into something that's actually hard work learning as opposed to something that's more like candy and helping me just kind of tune out things. I I I enjoyed that. Look, here here's like mine is just exactly what you'd expect.

It's like tracking AI's impact on power markets. It's like Chinese infrastructure spend. It's it's all this kind of stuff. Yeah, it's I mean a very predictable predictable business. Now I want to know send me share me share that article with me.

I want to know about Chinese infrastructure investments and how they're tracking with the Well, maybe that's a feature. Run shared his pulse today. I think so. You can check it out. Uh I mean you can already share the link so yeah you'll need be able to share it in the app.

Um anyway what uh what else is uh is keeping you up these days? Uh you were you were talking you you kind of laid out we reviewed the post on the show um um just talking more broadly about the nature of process power and Dan Wang's book.

Um, have you uh where do you sit on the uh the general characterization of the lawyerly society in America? Does it feel like we are becoming more lawyerly or less lawyerly? Well, some prominent uh engineers, founders and AI have certainly gotten latigious recently. So maybe America could do both.

At the same time, many engineers are crossing over into the government. Yeah.

Um, I guess what is interesting to me is like the extreme depth of the capital markets in the US where like one company can raise hundreds of billions of dollars on infrastructure spend and you know together the industry is spending I have no idea like you guys probably know the numbers better than I do.

Um, and that's like that's very interesting.

So where are the like you know if it's like such a lawyerly society why are we able to do this like why are we able to spend so much um and yeah of course it's true like if you look at like the west side of San Francisco it's like it should not be this like rows and rows of single family houses it's like quite a waste it's quite a shame but um there's many things like that that people have talked about at Infinitum But in general, it's like, okay, like Elon Musk showed up and he revolutionized two industries, right?

Like the space travel um and the electric cars. And I'm sure he fought with regulators and like there was a bunch of lawsuits and whatever, but like you know his like his I guess his will to power or whatever made it work.

And to I I am always a bit skeptical when people say like you know the US is impossible to innovate in because of regulations or it seems like it's it's getting done people do it.

Um yeah, you know, maybe like the the highest level of government is more lawyerly in the US than China, but I don't um there's also benefits to that which Dan Wang talks about, you know, like there's no uh I I I don't want to get into it, but I think you guys know the the upsides of uh the great rule of law and whatnot.

I yeah I do wonder if uh yeah I mean that that's kind of the odd uh counter take is that uh maybe the depth of the financial markets is actually a outgrowth of the lawyerly society when you have a bunch of lawyers who can create robust contracts where somebody feels comfortable parting with 10 billion dollars hundred billion dollars.

Is that a function of engineering?

Uh it might be a function of financial engineering in some ways but it is also a function of uh strong legal contracts and the work of lawyers to be creative but also fair in a way that instills confidence in the capital market such that a company's willing to to uh that seems right to to to deploy a hundred billion dollars and know that there's a at least there's an understanding of the pathway to the return on investment and that it's it's less there's less uncertainty.

Well, on the other hand, you have Tik Tok allegedly selling for 14 billion. Sure. Which I'm sure is less than what was spent to grow Tik Tok in the US. I don't know. But that's that's more government worth a couple more zeros than that. But yeah, just a couple.

It does seem Do you think uh the Gartner hype cycle is being overapplied in the context of AI? Do you think we could just have a straight line acceleration and there's no there's no trough or are you trough pilled?

I just don't get this thing where people like overlay graphs and they like do this like highlevel pattern matching when it's like okay like let's look at the basics which is to me the basics are codecs which is like uh like literally as of two months ago this was not a thing and now it's a thing and it's like to me a mind-blowing product just uh I can't I can't stress this enough um and okay like that alone is going to create just enormous revenue streams for both OpenAI and Anthropic I assume and cursor.

Um, so like I don't know. I I just don't it doesn't seem worth the Yeah, of course. Like Yes. Will it create the whole world is a trillion? Will it create a So So will it create a trillion dollar revenue stream potentially?

Um I don't think it needs to create a trillion dollar revenue stream to to keep up with the uh the data center plans, but that's um Yeah. for me to say. You should you should get on uh the the CFO or something, you know, one of these companies and ask them.

Um what what about uh give us your read on the state of open source today?

uh OpenAI's open- source models were demanded, they were released, and then you don't hear uh you know, there haven't been a lot of noise surrounding them, but certainly a lot of uh a lot of noise coming around some of the Chinese open source models, but what's what's your kind of high level thinking open source models?

Yeah. Yeah. Um well, there's two takes on that. One is that people demanded it but they didn't actually want it. The other is that maybe the OSS model just wasn't meeting the market needs or like it was not good at whatever it is the market wanted.

Um I don't know maybe it's a stated I mean to me it's a stated versus revealed preference at least for a lot of the people that were most loudly demanding open source. I mean this has happened I will say some of these the Chinese The Chinese Kimi K2 model is like really quite excellent to me. It's uh it's a great rival.

Okay. Um yeah, it's like the model I compare against most internally. It's uh kind of cool. I mean, people have people have been demanding open source versions of technology for a long time. There were open-source social networks and open source there's an open source version of Shopify.

uh and yet Shopify acrewed all the value and uh no one's forking that particular uh Ruby on Rails project just to get their e-commerce store up and running. Uh because I think there's there's like two problems with the the kind of hype around open source.

One is like okay you've open source the model but it's also clear that that doesn't immediately make it accessible to like a billion people. Totally.

um like obviously the chat GPT product and you know like whatever else rock is hugely downloaded Gemini whatever but like the um it's like by far the easiest way for someone to access AI and for free they're like doing it for free um so like on that angle it's not really all that decentralized just because the model weights are like flying around.

Um, second of all, if everything was always open source, you guys know as well as I do that it's like whole industry would never support this kind of capex y um and this R&D spend and whatever.

And so like there's just no there's no way like that that's just the only way to do things and like that's you know just just normal capitalism you know. Yeah. Uh question from the chat. Did you read one piece? No, I have not. I have a bunch of friends who swear by it, but I've never never gone into that.

Are there any law? Are there any other uh cultural influences that you pull from? Favorite movies, books, comic books, anything. Oh, yeah. Of course. All the time. Top of the stack. My favorite show ever is probably this anime called Evangelian. Oh, yeah. Uh it's like a 90s TV show.

It's like kind of grim but like in like is this a staple of like the Mecca fighter genre? Yeah. Are there are there metaphors from is it Evangelion or Evangelian? Uh are is is there a metaphor? Real heads call it Evangelion. Yeah. Evangelion. Yeah. It's a because that's like the way the Japanese say it. Yeah.

Is there are are there themes in that story that uh people pull from and apply to like the modern world and they go, "Oh, this is just like Evangelion. " Evangelion. Um I certainly think there's like a dense religious metaphor in there for like I don't know.

Um it's like a classic coming of age tale where um Bill Dunerman. Exactly. That's what they call it. Yeah. And this this kid, he's kind of like a depressive anti- like he's got none of the good qualities of a protagonist.

He's like constantly crashing out and he's like uh there's a bunch of girls who are into him and he's like not capable of talking to any of them and he you know but he's like magically like the best guy in the world at piloting these robots to beat these aliens and like humanity keeps calling on him even though he's like being mentally destroyed to come combat the aliens and it's like a story about like yeah if you can crash out and not talk to women But if you're good at if you're good at driving the robots, history will remember remember your name for sure.

Totally. What's the best way to get started with Evangelion? Start with episode one. There's only just watch it. It's not like, oh, you got to read the books first, that type of thing. You know how people will say that about like don't watch the movie until you see the read the books or whatever.

Um, if you want to get really pedantic anal about it, there's like a different subtitles that came out in the 90s that are better than the Netflix subtitles, but like it's not really that important, you know? Like it's a pretty big barrier entry. Okay. There's 26 episodes in a movie. It's pretty straightforward.

It's not like some crazy watch order and whatnot. Okay, man. Check it out. Yeah. Yeah. That's a bunch of media I love. Uh, what's the current state of the talent wars from your view? Did we peak over the summer or is it still just as intense? What's your read on it?

I mean it seems like okay if a researcher is a manager of compute and the compute numbers are growing and then like the effectiveness of a researcher is more and more important then I don't know like that line is uh trending up and to the right as far as I can tell but do you guys agree does that logic make sense I don't know 100% I think the logic makes sense.

I just think it's interesting that we're seeing essentially the same thing play out that we've seen for I don't know for generations in entrepreneurship where a founder can marshall capital and have leverage and incredible economic upside.

Now we're seeing founder level wealth and financial outcomes applied to employees which is just a different thing. It used to be in order to get the big big numbers, you had to go build something externally and then the board would justify paying you a bunch to bring it in in house.

And now it feels like companies are more they've actually just cracked the code on, hey, come work here, build this thing that's going to deliver a bunch of value and we will pay you like you're a founder that was out in the wilderness and we acquired in.

Um, and it feels like it's in some ways it always should have been this way. It feels like an aqua hire to me. Like just some sort of distributed aquaire. Totally. Yeah, that was that was that was my framing for it. You just have to look at it as an unauthorized distributed aqua hire. Yeah. It's like Okay.

And like how much did Zach pay in total for all these researchers over the course of four years? Yeah. Is it like that enormous? Is it like I don't know like comparative market for him? Certainly not relative to market cap.

Uh especially if you just think about like even if you take the boring view of like AI is just another business line that they need to be in. You know, it's like oh they're they're spinning up a a new division. It's like well if the new division drives 10% of their revenue or 20% it becomes a significant business line.

Uh investing 1% of your market cap makes a ton of sense to actually just like be in there. There's way more complicated dynamics there. But even if you just saw it as like we need an infrastructure team, we need a database team, we need a front-end team. It's like you need an AI team.

Uh and you're going to have leverage in in this case. Uh what what do you think about this uh Rich uh Rich Sutton take that's going out just dropped on the Dores show. LLMs are maybe not bitter lesson pilled. Um this idea that we need, you know, entirely new format.

People have been saying this for a very very long time. Deep learning is going to hit a wall, then we come out with something new. Uh wh what are your thoughts on uh uh the father of reinforcement learning potentially not thinking that LLMs are bit bitter lesson pilled?

Um I think this is a common sentiment like deep mind that they were always like they kind of held language models with a bit of like this is icky. This is not pure because we're not starting from scratch.

um we're learning from the human internet and that's like that's not uh you know it's not pure in some sense unlike say Alpha Zero which is learned from nothing like you just selfplay yourself to superhuman performance.

Um, so yeah, like maybe it's not completely the most bitter lesson paradigm in the world, but like I don't know that doesn't mean to me that it won't like reach extremely useful levels of intelligence. And I don't you should ask Ilia about that. It already is really useful.

Like yeah, it's like the question of like uh will we get value out of this is like answered by DAUs and just a million anecdotes of people that use the products. in my in my opinion. Jordy, anything else? I got a lot more stuff, but we're over time. Yeah. Well, thanks so much for coming on the show.

We glad we glad we finally made this happen. Thank you for posting the posts that make this show possible. We truly enjoy covering everyone. John, you want to hit the gong for Rune? Yes, congratulations. Well, can't get out without a gong. 250. What an honor. 250,000 followers on X, I believe, is roughly the number.

Something like that. It's definitely gong worthy.