transcribe

AI’s Memory Wall + Swyx on AI Engineering

Cognitive Revolution "How AI Changes Everything" · 2h 40m · transcribed Oct 2026
More from Cognitive Revolution "How AI Changes Everything" Business
𝕏 Share ▶ YouTube 📥 PDF 🤖 .md

Section Insights

# 0:00

Insights from The Curve Conference

What were the key takeaways from The Curve conference?

The Curve conference aimed to bring together contrasting opinions in the AI space, highlighting the need for broader perspectives amidst rapid advancements. Attendees noted that while many companies are making progress, there is a lack of time to step back and assess the overall landscape.

  • The event emphasized the importance of diverse viewpoints in AI.
  • Participants are often too focused on their specific advancements to see the bigger picture.
  • Networking with leading figures in AI can provide valuable insights.
# 16:02

The Competitive Landscape in AI

How do major players like Elon Musk and Mark Zuckerberg fit into the current AI landscape?

While traditionally seen as leaders, Musk and Zuckerberg are now viewed as distillers of AI advancements rather than primary innovators. The focus has shifted to a few key companies pushing the frontier of AI technology.

  • The AI landscape is narrowing to a few dominant players.
  • Musk and Zuckerberg's roles are evolving as they leverage existing advancements.
  • The cottage industry of RL environments is becoming crucial for innovation.
# 32:04

Introduction to Thomas Summers and Posetron AI

Who is Thomas Summers and what is Posetron AI's mission?

Thomas Summers is the co-founder of Posetron AI, which focuses on developing hardware for AI inference, aiming to enhance memory systems for efficiency and cost-effectiveness. His background includes significant roles in AI hardware development.

  • Posetron AI aims to revolutionize AI hardware with a focus on memory efficiency.
  • Thomas Summers has extensive experience in AI and hardware architecture.
  • The company is rapidly deploying its technology in data centers.
# 48:06

Advancements in Memory Technology

What innovations are being made in memory technology for AI?

Posetron AI's Osmo chip offers significantly higher memory capacity compared to current benchmarks, allowing for more efficient scaling and reducing the complexity of multi-device setups.

  • Higher memory capacity per chip can drastically reduce hardware needs.
  • Simplifying deployment can enhance overall performance.
  • Advancements in memory technology are crucial for AI scalability.
# 64:09

Verification and Emulation in Chip Design

How are AI advancements impacting chip design verification?

Recent AI models have improved the speed and efficiency of chip design verification processes, allowing for closed-loop iterations and automated documentation handling, which were previously time-consuming tasks.

  • AI is significantly speeding up chip design verification.
  • Automation in documentation and testing is becoming more prevalent.
  • The integration of AI in engineering processes is transforming traditional workflows.
# 80:11

Upcoming AI Engineer Conference

What can we expect from the upcoming AI Engineer conference?

The AI Engineer conference will focus on the latest trends in AI applications across various sectors, including finance and technology, and aims to gather insights from industry leaders.

  • The conference will address key trends in AI engineering.
  • Networking opportunities with industry experts will be available.
  • Focus areas include banking, investing, and financial technology.
# 96:13

Shifts in Software Engineering Practices

How is software engineering evolving with the rise of AI?

There is a shift from meticulous line-by-line code review to a focus on module coherence, allowing for faster development despite potential oversights in individual components.

  • Software engineering is adapting to prioritize speed over exhaustive detail.
  • Managing complexity in modular systems is becoming essential.
  • AI agents are changing traditional coding practices.
# 112:16

The Future of Software Demand

What is the outlook for software demand in the context of AI?

Despite potential market corrections, the demand for custom software remains high as businesses seek to automate and enhance their operations, turning traditional processes into software solutions.

  • There is a latent demand for software that remains unmet.
  • Automation of traditional tasks can lead to increased software demand.
  • Market fluctuations may impact companies, but opportunities still exist.
# 128:18

The Impact of AI on Business Operations

How is AI affecting business operations and employment?

AI is leading to significant changes in how businesses operate, with potential layoffs in tech sectors as companies adapt to new efficiencies. However, the demand for skilled workers in AI and software development continues to grow.

  • AI is reshaping business operations and workforce needs.
  • There may be layoffs in traditional tech roles as companies streamline.
  • The demand for AI skills is increasing, creating new job opportunities.
# 144:20

Permanent Changes Due to AI Integration

What are the long-term effects of AI on work and productivity?

AI has permanently altered workflows, compressing tasks that once took significant time into much shorter periods. This shift means that removing AI would require substantial human resources to replace the efficiencies gained.

  • AI has created lasting efficiencies in daily tasks.
  • Removing AI would necessitate a large workforce to maintain productivity.
  • The integration of AI tools has fundamentally changed how work is approached.

Transcript

1:20 Here we go. Good morning. It is Tuesday, October 6, 9:00 a.m. Nathan, good morning to you. >> Good morning, Pash. How are you today? >> I am very good. I we met over the weekend at Light Haven, which is now become an infamous venue. you know that that that den of infamy and I think Ligh Haven had a conference called the curve over the last few days. you attended what were the and also a good number of other people from the Frontier Labs attended.

2:03 I I saw Jack Clark there. I saw a number of our former guests, Max Nadau, Zuy Marshawitz. I saw, you know, I was walking around and I was like, "That guy looks really familiar." And I was like, "Oh, yeah, that's Jacob Simlman, the fields medalist. So, a good number of leading lights in the field. what was your takeaway from the curve?" >> Yeah, it was quite a jam-packed weekend and it was great to see you in person for a change. We vlogged a lot of hours here on the in the studio, but not too many in the flesh. So, that was fun.

2:40 So the curb is it's a really interesting and pretty unique event in the AI space started by a couple people who had the idea that we need to bring more contrasting opinions together. you know, I guess my kind of meta strategy for like what I'm doing in AI is just observing that everybody's so deep down their rabbit hole advancing their frontier, usually having success, but has, you know, everybody has so little time because the pace is so breathless to come up for air and see what's going on. I'm kind of doing the zag and trying to cultivate the broadest view possible. And in the event world, I would say typically, and we're going to talk to Swix today about his upcoming AI engineer event, which is a great example of a leading event where lots of people go and they do a lot of learning and you know, I've been to one of those too, and it was it was a great experience as well, but it's very focused on like a particular point of view, right? It's like how are we going to make these AIs, you know, help us become better engineers and build better products and, you know, power enterprises and and whatever. It's not really a place where people come together to hash things out all that much. So the organizers of the curve were really trying to create a space where people with quite different perspectives could come together and you know at least attempt a meeting of the minds and that that's what it's been about for each of the three years that it's run. It's been people from frontier companies, short timelines, people, fast takeoff people on the one hand, but then also, you know, folks who have published articles like AI is a normal technology, people from DC who are initially were very skeptical, but this was like more than, you know, more Silicon Valley hype, but they've they're coming around and, you know, trying to create a space for for sense making and and kind of shared deliberation. It's been really successful in that regard. people have u I think really come to it in a in a positive spirit and certainly not everybody leaves agreeing on all the key issues but I I do think it's it's earned a reputation as a place where people do have a real exchange of ideas across like paradigms you know ideological divides backgrounds coast you name it so >> well you as you say there's a there's a some divides there so can you speak to some of the divides that you saw over over the weekend.

5:05 >> Yeah. I mean, I think it one of the big observations in general in AI is that events have come pretty fast. And you would think if you were looking back, you know, if you were looking ahead a couple years ago that with two years of additional information and revelations and, you know, actually seeing how strong the AIS get and and what they do and who's been right or wrong, you would think that like people would start to agree on the state of things. And that has happened a lot less in the broader world than I would have expected, but I did feel that there was some of that felt at the curve. So, you know, there there's all these different splits, you know, pro- open-source because it's going to save us from concentration of power issues, anti-open source because it's going to create, you know, biorisk.

5:54 you know, there's just tons of these kind of dualities where typically I I think except on the sort of like our AI is going to get really powerful or not, you know, I think generally both sides have pretty good points and there's like pretty serious trade-offs involved. But I do think the big thing that people have kind of come together on at least to a significant degree is it does look like the AIS are going to get really powerful and this is something we're really going to have to reckon with and can't just sort of hope goes away or you know feel like maybe if we just wait you know the bubble will burst or the fever will break or what have you. This time around people were definitely like okay it's getting pretty serious. you know there's there's there were very few perspectives that as as Helen Toner once you know famously put it like long timelines have got super short. I would say you know another correlary to that was like what is called normal in terms of expectations is getting increasingly pretty incredible. even I would say the people there who had the most modest expectations for what AIS will do, be able to do were like, "Yeah, it's it's going to do an awful lot." And and the question now on the capabilities front has gone to like, "Is RSI going to fume? Is it going to, you know, be fast for a while and then kind of level off?

7:17 is it, you know, is that more than it's meant to be? Like, yeah, we've got these super powerful things, but is it really going to kind of run away from us?" Those were still kind of the capabilities questions that were getting asked, but there has been, I think, you know, some healthy updating across the community just based on what we've seen. So the were there any from the frontier labs were there any kind of limits or constraints that they say that they said you know should be enforced on players in the space including themselves and others.

7:58 Yeah, that was a huge topic. I mean, and probably the biggest divide in terms of perspective is really just inside the frontier companies, outside the frontier companies, right? Everybody knows that they're living in the future and they have, you know, visibility that the rest of us don't have. And so actually had a chance to talk to and hear from in some cases in sessions, in some cases just in you know kind of passing conversations all cadam house rules who all kind of abstract way but we're talking people who are like founders, executives, top researchers at you know frontier companies and I'm you know small set of them. I asked a couple of them why come here you know what what what's in it for you you know and and before doing a couple sessions where I was like interviewing or moderating I asked the same question what what makes this next hour a good use of your time and they repeatedly expressed that it's just important to them to share as much as they can with the broader community.

9:02 and again they they didn't want to you know have this kind of turn into media product directly but they did feel like this is an event where people are coming together in good faith. There's a lot of people they've known for a long time some new people too but a pretty high trust and you know high goodwill environment where they could come and share their honest perspectives. Definitely not sharing their you know trade secrets of course still but really trying to give us a a candid take on what are they seeing internally? what do they believe internally? You know, what's their their outlook? What's the plan? How can we come together and, you know, potentially do some of these constraints? And there were a number of things that I actually thought were like kind of newsworthy.

9:45 one let's say senior executive at a frontier lab first of all was was saying that just pre-training continues to deliver that the models are just at the base level of capability getting stronger and stronger and you know there doesn't seem to be an obvious stopping point to that scaling laws are holding or maybe even bending because the data quality is getting better they've got you know increasingly different architecture tricks. lots of work on synthetic data. Synthetic data, you know, seems to be definitely really delivering you basically, you know, you you need more pre-training data. What do you do? You like give it a give the model a bunch of data and think like what what could I transform this data into that if I then pre-trained on it would make the model smarter. And you're sort of converting these reasoning tokens. you're you know test time compute you're converting back into pre-training data which then you know goes in of of course to the next generation of models. So this is like one way in which recursive self-improvement is happening. Not even necessarily I mean there are of course opportunities for architectural improvements as well but a big emphasis was just on data quality. We can now just spend tokens on the existing data to augment it to you know to create better cleaner different you know lenses on it. We feed all that back in and the models get better. So this this one executive was really emphasizing pre-training continues to deliver the models are getting super powerful.

11:20 you know, RL is important too, but his take actually went as far as say as like he believes that the latest frontier models do have the ineffable research taste that would be needed for them to start making paradigm level breakthroughs. Excuse me, lost my voice a little bit there talking so much. so again, that's like a key key statement. So he believes the pre-trained models do have the capability to do these paradigm level breakthroughs. He believes it's currently an elicitation problem where the RL is not bringing it out well enough. They don't know quite how to bring it out well enough. And so it's kind of rare and this is why you need 10,000 agents, you know, to to do a Millennium Prize problem. U and of course that's verifiable, too. So like RL is, you know, is working better there. maybe you need a million agents, you know, at this point to like stumble onto something eventually where you'd get the kind of paradigmatic change at a at a you know, ML research level, but he believes that is there and it and it will be they'll figure out how to elicit it. And then he said a couple things that I thought were just like legitimately newsworthy. one was he believes that there is a level of intelligence or there is might be a little strong but there there likely is let's say a level of intelligence that we just shouldn't go past >> what >> yeah first time I ever heard that from a frontier lab leader and it was not like minced words it wasn't a like there's definitely a level or I can articulate what that level is but like this was you know person who again like everybody would recognize you know their name and and position saying I think it is likely that there's a level that we shouldn't go past. Now was that not ever or just you know not right now? I think that was also a little bit ambiguous but it was a it was a pretty firm statement. Definitely the the strongest statement I've heard about imposing kind of a hard cap on capabilities even for a time. What what what confused me what confuses me about that that statement is that some of the increases intelligence that we were seeing number one they're coming as a result of compute increases in compute which has traditionally been what has happened in the last you know decade and a half. Number two, they are coming as a result of refining the input data. Like the input data is being refined by greater intelligences and so you have input data which is much better. When we say that we shouldn't go exceed a point of intelligence that automatically means that you have to look at the inputs going in and the inputs going in are going to be compute and this refinement of data. And so if you say that we shouldn't go beyond that p that point that also means that you shouldn't build compute beyond that point or you shouldn't refine data further beyond that point which which which means a lot really it means like the end of the end of like the compute buildout which is which is pretty significant right >> yeah well I don't I mean I think there's a lot of uses for compute so so I had a chance to follow up on this you okay this this combination of pre-training is really delivering and scaling laws if anything maybe are bending more favorably. Although it's, you know, it's tricky to know what you count, right? Do you count all those flops that you use to like create the synthetic data? Maybe you should. How to count these things is always super vexing. but between this, you know, pre-training thing and the there might be a cap on how far we should go. Then I just asked a follow-up like, okay, well, how do we operationalize that? Would you be prepared to sign on to something like a limit on the number of flops that go into the next pre-training run? You know, should we and I don't know what that number would be, but and and again, what exactly would count, but like should we say no more than 10 to the 27 flops in the next pre-train or something like that? And I kind of anticipated that being a bit of like a straw man, you know, for him to react to or like tell me what was bad about that and maybe ideally give me a better idea. But what actually happened was he basically said, "Yeah, I think that could be reasonable."

15:41 The end. not not really much of a fight there at all. Again, the details will be like incredibly important if we go down that path. because the more data you enrich, you know, the more flops you put into enriching the data, you know, you can kind of you can definitely play a sort of shell game of hide the compute. I think these days but the sentiment was like yeah we we might have to do something like that. And then of course your question I can anticipate I know you well enough to to predict the next Pash token at this point is what about Elon? What about Zuck?

16:18 >> And the answer there there were a couple really interesting things people said along the way. one they said you know in nobody this is not from one individual but sort of a takeaway from the event in general is that the let's say two to3 Gemini 4 being announced now we maybe have a three horse you know very frontier race again we'll see as that comes online but the the take was kind of the two to three companies are really pushing the frontier and even the other two that we would usually to think of in the top five, Elon and Zuck are actually distilling a lot whether they know it or not. Mhm.

17:01 >> And the way that they are distilling these days, it's not that they're, you know, calling the cloud API and getting traces and feeding them in, but rather the cottage industry of RL environment, makers and sellers is in in fact functioning as a distillation conduit because what they're all doing is having Claude create these RL environments, selling the RL environments to other frontier frontier model makers, and Then they do their internal RL on an environment that only exists because cloud was smart enough to make it and that's how they are kind of re you know constantly capturing the the capabilities advances that the frontier you know the true frontier companies had. So there wasn't actually a ton of talk about competition beyond the top three. one person went as far as to say and this was another like you know founder executive type at at one of these companies. yes they have kept up obviously there's been a gap you know but and that gap has been persistent but they've kind of maintained kind of consistent following distance and this is in contrast to Anthropic's old prediction from their fundraising deck that I always think about when they said you know maybe in 2026ish the companies that train the best models will be so far ahead that nobody will be able to catch up. Obviously that hasn't happened. But the the kind of synthesis of all this is if they hadn't been releasing the models and if people weren't able to do these like various sort of direct and indirect distillation techniques, they actually think that probably would have happened. that the other companies are not really keeping up but for the the sort of leak of intelligence in all these different ways including the RL industry. that is gradually finding ways to transfer from frontier models to their competitors.

19:02 >> I I did I did speak to some people at the at the party and one of the one of I think the prevailing wisdom conventional wisdom there was that Elon doesn't really have talent because no one wants to work with him anymore. and so Elon was not regarded as being serious in that sense because he's already chased away everyone who you know started working with him and I think I think on Elon's side Elon is very focused on the infra buildout. So Elon is making a making a a gamble that the infra matters more than the models and that he will be able to catch up basically using his infra buildout and so and he's and he's structuring things for 2029 2030 2031 like terapab things like that will come out much later the SpaceX like you know satellites that's like a 2030s So he he's he's structuring for a much longer period of time and I think anthropic and opening high are both structuring for more like the two to three year mark you know and they're and they're looking at a 2028 2029 AGI RSI which >> yeah honestly even sooner I would say was kind of the vibe and this is probably somewhat biased to be a short timelines crowd but again like the short timelines crowd includes the you know executives top researchers founders of these these companies they were talking more about just next year. it was really a kind of 2020 like critical period is beginning now.

20:45 RSI is kind of at hand. definitely not full agreement on how fast it goes or if or when it tops out or whatever, but a shared sense that like the decisions that get made over the coming months and and the way that compute is used over the next year could be really critical. there wasn't really even too much 2028 talk to be honest which is again like pretty crazy on the compute buildout. I mean I think you know we'll see where value occurs but I think that Elon's bet is not necessarily a bad one even if he can't really compete at the model layer and you know I don't count him out but you know just reporting the the news from the the front. but like in terms of you know will they still have need for compute or or does a cap on pre-training scale or other you know compute input cap that might constitute a pacing mechanism. Does that mean the end of the compute buildout? I would say no. another question I had the chance to ask was, of course, everybody's heard Jensen on Ezra Klein, I think, at this point. And the comment that he made was like, you know, these companies are going to have to mature. They're going to have to change their ratio of where they spend their resources. He said at Nvidia we spend like 20% of our effort on designing the chip and we spend 80% of the effort on verifying and validating and testing all the edge cases and making sure it's going to last a long time and be reliable and you know all these other things that are not just the original design. So his point was like they've been working really hard obviously until this point and they've been putting everything they can into making their models capable enough to be useful.

22:33 Congratulations, you've succeeded. Now you're going to enter the era where you're going to have to make them safe, reliable, trustworthy, etc. And so I had a chance to ask, you know, one of these people like, "What do you think about that take from Jensen?" And the response was basically like, yeah, that's that's kind of reasonable. You know, without again knowing exactly where the numbers will shake out, u the possibility that the majority of could compute in the future could be going into all sorts of different safety measures, you know, which could be monitoring, could be chain of thought monitoring.

23:09 People are pretty getting kind of bearish on chain of thought monitoring. >> Yeah. >> but then there's internal, you know, activation monitoring. They're already spending a significant percent of compute on monitoring and it sounds like that might go up a lot. Another thing that they're spending compute on in a big way right now is fixing the RL environments there. You know, of course, everybody understands at this point that like if you have sloppy RL environments that reward cheating, then you're going to get a lot of cheating and okay, now what to do about it, right? like it seems I wouldn't say it seems like we're quite to the point yet where they have a scaling law for cheating. but it sounds like they might be getting kind of close to a scaling law for cheating because they're looking at all these RL environments and they're like just going nuts, you know, with the they're just taking their best models, right? And this is kind of another way of kind of another form of recursive self-improvement, right? Like a model was trained on a bunch of hackable RL environments where cheating was rewarded. It's a bit of a cheater.

24:22 That's a big problem. but you know, you can still get like obviously lots of valuable work out of it as we do all the time. they are now applying these models to the RL environments themselves and saying hack this environment you know and now that is the direct task they're doing it they're finding you know all these places and and ways in which the environments can be hacked they're fixing those I'm sure they're throwing some environments out that are just fundamentally flawed whatever and gradually they're driving down the rate of flaws which means that they're also driving down the rate at which cheating would be rewarded which in turn means means you're going to get less cheating from the models as they actually come online.

25:05 And it I got the sense although it it was you know I think this is still like as they say an open research question. but I got the sense that they are seeing a pretty clear relationship between the more you clean up again excuse me the more you clean up the RL environments you do in fact get lower rates of cheating downstream and you know they're trying to push obviously they'd love to have zero you know hackable environments. That's going to be tough.

25:36 it doesn't seem like we're all the way to the point where we have like a scaling law but you know kind of we're seeing sort of a protocaling law it seems like in you know again the way you can spend a lot of compute driving this down and and getting better behavior on the other end. I heard repeatedly that they expect to be bottlenecked on safety and alignment, but it seems like for at least the next generation or two, there is still enough lowhanging fruit to pick in the safety and alignment domain in part just by fixing RL environments that we should expect things to continue to move pretty fast. We should expect like more powerful models that they will probably feel confident enough to release because they'll take like a you know pretty I would I would expect like maybe two orders of magnitude cheating out by taking I don't know you know this is where the exact shape of the scaling law gets gets pretty important right if you have one environment that rewards cheating you know is that enough to teach your model to cheat probably not I would guess not I don't know maybe it could be but you know, how low can they drive that? My guess is they'll take those those the rate of wi at which the environments are hackable super low and cheating will also go like much lower and we'll have to see what that ratio looks like. But it seems like they're probably going to be able to get comfortable releasing the next couple things even as they do consider themselves bottlenecked on alignment. But then past that they're like, "Yeah, all bets are off." we really hard to say what happens kind of past the next one to two generations.

27:26 >> I I I will say that you know being on an exponential it's not unusual that you can't see beyond two generations, right? That's it's it's it's part for the course. you would have to be literally a little bit like insane or a cultist to believe that you know in in five or 10 years we're going to have Dyson spheres. That's that's that's a tough one, right? So, I feel like that's not a you know, it's not unusual that you wouldn't be able to see past that past that period.

28:00 I wonder what comes next. because I think last year we had this end of year process where both companies dropped models around Thanksgiving and everyone kind of checked out for like a month and a half and people had a chance to use these things at home and when we came back in January we started to get like okay codic is kind of solved and by March you had like this you know January to March was just this enormous like you know all the way up right so I feel like right now Anthropic is rushing for the IPO. They're trying to IPO by it was supposed to be late October. Now it's pushed back to first week of November. last last call last call is just before Thanksgiving. So I think they're trying to rush out the IPO. Now the question for me is why?

28:54 and that I'm not too clear about. I don't know if they have a cash flow necessity because they are committed for like a hundred more than hundred billion dollars of commitments for next year. or do they want to get the shares into the hands of the public before the next ramp up in revenue and you know spread out spread out the wealth before before things happen? or is it something else? I I I don't know.

29:25 But there seems to be an intense pressure to get the companies to IPO to get the company to IPO this year which I think is rather unusual. So >> yeah, oddly that never really came up in my conversations at all. the assumptions seem to be at least these you know two companies will be able to raise the capital that they need one way or another whether it's public, private, whatever the case may be. I think I mean there was definitely some interesting discussion around how hard will anthropic press their advantage to the degree that they have an advantage right now. OpenAI folks seem to be like kind of not sure what to make of anthropic and there were a number of questions from the audience to anthropic people over time where it was like okay you know we're trying to get serious over here and going back way back to anthropics founding and there's a lot of people who you know at this event there are a lot of people who were there you know in the mix at and before Anthropic's founding and the sense that the AI safety community had when Anthropic was founded is that they were committed to not advancing the frontier. They wanted to be at the frontier so they could be relevant and do all the research, but they weren't going to push it because they didn't want to make the race condition worse. And people challenged some anthropic leadership on that. And their response was like, well, what we said was we would never publish methods that would advance the frontier.

30:53 And people were like, I don't think that's really what you were saying at the time. so there's, I think, a little bit of, if not, you know, I don't know exactly what was said obviously in every, you know, instance, but like people had the impression that they were not going to be pressing commercial advantage and now they've been asked like, "Okay, so right now you guys are you may you probably a little lead if anything." but it's like pretty competitive. You know, OpenAI's trying to get their house in order. You guys definitely have your, you know, work to do too, but you seem like you have maybe a little bit more in order than they do.

31:27 are you going to chill and like let them get their house in order and not try to like run away with the competition or not? And we didn't really get an answer on that. So I I don't know what to expect. and there actually wasn't much discussion really at all of like when will the next model launch or you know that that kind of thing. so I didn't come away with any major updates. I would hope that they would, you know, show a little bit of friendly restraint, but I'm not sure what to expect.

31:58 >> yeah, it's it's it's it's the story keeps changing, right? I mean, and it's not it's not unfair because as things on the ground change, you do have to change you know what you're doing. >> Yeah. >> So, >> as President Xi said, a man of wisdom adapts to circumstances. >> Indeed. and speaking of adapting to circumstances, let me you know, kick off for our >> great segue, >> a great segue to our guest for today.

32:38 Our guest for today is Thomas Summers. Thomas Summers is co-founder and chairman of Posetron AI, a company that builds hardware for inference, running a trained AI model to answer questions, write code, or perform other tasks. Posetron concentrates on the memory systems that feed data to a processor, aiming to make those tasks cheaper and more energyefficient. Thomas previously founded Rex Computing, a processor startup, and was a 2013 Theel fellow. He later became principal hardware architect at Lambda which provides access to graphics processors for AI developers and worked on its early cloud infrastructure. He then served as director of technology strategy at Grock another AI chip company. His experience includes both designing processors and putting other companies processors to work in data centers. Posetron was founded in April 2023 and shipped its first Atlas server in August 2024. Atlas uses programmable chips. The company's next generation Asimov is custom silicon intended for Titan servers. Posetron reports that more than 50 Atlas racks have been deployed at Oracle. Alongside that work, Thomas has questioned a familiar assumption about computing equipment that advancing software necessarily makes yesterday's machines less useful.

33:59 In February, he pointed to the first V 100 systems he deployed at Lambda and said they were still in service. Posatron has recently raised a series C I believe from Oracle and let's get Thomas up here. Let's see. Let's just wait a moment while he comes up. But it's it's interesting because I think we now have a series of very successful chip startups. at this point we have Cerebras.

34:39 Cerebras was for a moment last week was in trouble as semi analysis put out a note saying that Astra ultrafast is not using Cerebra's chips. semi analysis then put out a note saying that but that's because they're fully sold out but in that in that time period between those two notes Cerebras's stock had taken a beating. it seems that Cerebras is currently being used by our friends at Jane Street. so let me bring up Tom here.

35:15 >> Hi Thomas. >> Hey guys. Sorry for that technical difficulty. >> no worries. it's a pleasure to have you on the show. Thank you. Thanks for having me. >> the big news which is you guys have just concluded a series fundraising I believe from Oracle. tell me a bit about how that how that whole deal came together. >> Yeah. the funding was not from Oracle. So we we have Oracle as a customer. they've they're they're you know the first large scale deployment for us of our first generation product. So yeah I would say you know this new fundraising was led by Valor Equity Partners Trades Management NEA Andrew Capital and Jim Clark the co-founder of Netscape and SGI and so that is very much on the heels of the Oracle you know customer deployment and you know in anticipation of our next generation silicon taping off the end of the one of the things that you've done is you've speeded up the the process of designing and deploying the chips. Can you tell us a little bit about how you did that?

36:27 >> Yeah. so yeah, pos started about three and a half years ago. I would say like a core founding tenant was that we wanted to get chips out real hardware in customers hands as quickly as possible. It was about 15 months from cold started company to shipping first customer server. and then you know about three years total time to to win Oracle as a customer. And so pretty unheard of in the the new AI hardware space to to ship that quickly. And really that all came down to you know obviously great team but the fact that we're leveraging AI and really going pedal to the metal in terms of being able to iterate on that design and that that came down to you know us starting off with FPGAAS as the basis for our first generation products. This is field programmable data rays hardware that can actually be reconfigured and enable us to have very rapid iteration loop.

37:26 and and you know that was what what spearheaded the rapid development of our first gen product but for our second gen which is full custom silicon it's really really amazing the advancements in in AI over the past year specifically in our ability to close the iteration loop in in terms of everything from RTO development so all the front-end work related chip design doing the the verification which is the the biggest time limiter I would say for historically for chip designs all the way to doing the full background. So place and route and all of the the complex you know physics that's involved in that AI has been able to to drastically accelerate our our own development of chips to make AI cheaper and and more energy efficient.

38:14 >> Sounds like recursive self-improvement to me. >> indeed. >> Several things I I would love to dig in on there a little bit more. Maybe for starters, how do you taxonomize the different strategies that chip companies have at a high level? Right? Of course, like GPUs are like highly flexible and then, you know, we know all about the CUDA. We know some people are like going as far as burning in particular models into the silicon itself. how do you kind of see that constellation of like big picture strategies and and where do you guys fit into that?

38:51 >> Yeah. so pause from you know core focus that I from the beginning that I stated of of rapid iteration is number one. Number two is focusing on performance per dollar in TCO. so a lot of other you know chip companies have have you know tried to carve out niches be it ultra fast or you know that they're only going to focus on on you know one particular model and and get efficiency there. we want to be the TCO win and TCO wins require you to be fairly general and so you know we we start off saying that no we want to build a a fairly general linear algebra processor. If we think of what is the unifying factor of not just all you know transformer models today but what is the the unifying factor of if if I were to say could could an architecture that was designed 50 years ago you know still be be used today and and kind of with that thought cap on thinking you know can underlying hardware you designed today still be used for whatever you know computational applications exists 50 years from now and thankfully linear algebra the fundamentals of that math have not changed and and you know I I would make that bet that you know what we're building won't be fast in 50 years but it will still fundamentally function and so we we've kind of taken that lane and coupled kind of with that that general philosophy is that what we have to build also needs to be you know generally manufacturable and something that can scale and so there there's a whole another you know category large different categories, but kind of can be lumped in a bucket of people that are really trying like new you know, analog circuit designs or doing light based, you know, photonic computing and and all of that. And I think a lot of those things are very cool as as an engineer, but still I would say are much more in the science experiment stage state rather than something that could be mass market commercialized. and so you know be it with kind of our our philosophical underpinnings but also in just how we've executed and how we're we're developing and pushing our our nextg product it's all about how can we actually scale to hundreds of megawatts you know gigawatts of of compute as as quickly as possible. And so, you know, we're leveraging commodity memory technology as as the the you know, core part part of our our supply chain rather than the very expensive and very supply chain constrained HBM that everyone else in in the the high performance AI accelerator space is using. So, you know, we're we're taking a a pretty different approach with that that key tenant of can this thing both provide the best TCO for customers and can we actually scale this out to, you know, massive massive deployment sizes.

41:44 >> So, how does how do you deal with there? There have been increases in prices of all kinds of memory modules in the last year or so and pretty pretty significant increases. has that affected your projections on the TCO going forward? the the sad reality is that you know I I was looking at a quote that we had a year and a week ago and for memory and it's gone up four and a half times since then and our my expectation is you know there there I I do not I I really hope that this isn't the case but I wouldn't be surprised if this is going to end up being a another doubling over the next year just because I think that and and this gets to to TCO. So if if if all all other things being the same TCO comparative to other compute providers, everyone's in the same boat with having to increase their their cogs and and then thus prices. So that doesn't affect us that much from a competitive standpoint. I would say the the fear would be oh if if the cost of the products goes up and you're otherwise delivering the same value then then there's a problem but I would say a big reason why everyone has burdened these these cost increases is because you're driving more than that for 5x value compared to last year and so just from a you know sort of a primitive capitalist you know perspective of the the whole point of of pricing is to to enable you know things that are are of higher value to to receive those those goods that that are are limited. you know the the system is working in that sense because memory is getting to to the applications that have the actual greatest value and even if those that value is generating you know AI videos that we all share as memes. but but I think that the reality there is compared to a year ago, the capability of a model using x amount of gigabytes of memory is way more than 5x than than it was this time last year.

44:04 >> Yeah, we've got abundance of memes already, if not abundance of everything else we might want. Help us understand a little bit more the different I mean I know you know of course there's different kinds of memory. We know the the very basics, but like what kind of trade-offs does it involve or what kind of, you know, engineering challenges do you take on in order to get the advantage of being able to use commodity memory versus the, you know, the very high-end memory?

44:34 yeah, you know, how does that work for you and how does it flow, you know, all the way through to the customer experience? >> Yeah. I I wouldn't say I was I I definitely did not expect this level of price increase, but at the start of Positron, I I did think that memory was going to be the the limiting factor for for things going forward, even though you know people can debate on power and and other parts of the infrastructure stack. My fundamental belief was that there isn't going to be a stop anytime in the near future in terms of scaling laws and and getting increased value from from making models larger. So that eats into memory on one side. but then the other side, I think the main limiter today of AI being applied to to most applications is context length and being able to hold more context you know per user and then scaling that out to drastically more users and and really today I would say that the driver of quote unquote users or individual sessions is actually just having more agents. I've gone from, you know, three, four months ago where I had probably on average two to four agents like running constantly in the background to where I'm now running 15 to 20. So you know if you then multiply how many people are using AI and how many that that concurrent completely separate contexts adds up very very quickly. So that that memory driver had us say okay we have to use the commodity memory because that's the only thing that's going to be able to scale and be cost-ffective and when we set that as sort of the constraint in our architectural design we had to come up with with you know very very clever innovative solutions and and you know the the two main pieces of that and you know the thing that we showed with our first generation product is being able to achieve extremely high memory bandwidth utilization from a compute architecture perspect perspective. So while Nvidia GPUs, you know, I I would say on average are getting between 30 and 40% memory bandwidth utilization in the actual decode forward pass of of a transformer model. So basically that means even though they advertise you know 8 terabytes per second of theoretical memory bandwidth with you know a B300 you only actually are seeing you know something in the ballpark of like 2 terabytes per second of actual realized memory bandwidth. and that comes down to a whole bunch of you know GPU architecture details. comes down to the reuse patterns that really don't exist in transformers that they designed their hardware architecture around for for training and other workloads. and so with our firstg product, we were able to hit and sustain 93% of theoretical memory bandwidth. so that you know theoretical and realized basically you know become one. and so that you know isn't that massive, you know, 3x improvement right there. but to to really take advantage of you know the belief that the models and everything are just going to get larger we had to be able to scale to to massive more channels. And so we're we're you know partners with Credo semiconductor that and and have developed a memory chiplet solution that allows us to go from you know the maximum number of LPDDR same type of memory that is in phones, laptops etc.

47:57 The max number of channels that you can you know have find in any other products is on the order of you know 12 12 to 16 channels and we go up to 72 channels of of LPDDR5X with this decoupled memory chip solution. So walk me through a model that fits locally like which which parts of the networking or coordination start to disappear.

48:27 >> Yeah. Yeah. So a very big part of it is if you have all this memory. So with Osmo the the our chip we have up to 2.3 terabytes of memory capacity per chip. So when you compare you know the the B300 shipping today caps out at 288 GB. a lot of the the you know original deployments of them actually were were using you know 188 or excuse me 180 gigabytes as as the base skew for the for the B200s. but you know as you know reported in in by by analysts Nvidia for the next generation is actually cutting down the memory capacity because cost and everything involved. So they're only going to be at 192 GB with the Ribbon Ultra. And so but even if you use that 288 as like the current benchmark, we've got eight times more memory capacity per chip. And so that means on a per chip basis, you now can actually scale and what you would have needed from a memory perspective eight GPUs for we can do on a single device. And it's not just oh you've got that cost silicon savings. you know as as you pointed out when whenever you have to scale to having more than one device there's overheads involved with that. There's you know your all gathers all reduces that you have to do within for each layer and each really each multiply that you're doing of of sharding across those devices. And so by having that on a single device, you you remove a lot of the the both complexity of of that that that is both like software and and other burden from just a a deployment and and building standout standpoint, but also you know it comes down to real performance in the end as well.

50:18 So obviously the nature of the hardware that people can access determines in part what kind of models they can deploy. A huge topic recently has been on looping within the forward pass and what impact that might be having on our ability to monitor models effectively. Chain of thought is becoming you know less less reliable as a monitoring technique. My best understanding from what I've just heard from people is that it's not so much because of the looping and it's more just because of the pre-train getting more and more powerful and and being able to kind of juggle more advanced representations internally. But I'd love to get your kind of perspective and breakdown on how these hardware considerations relate to looping. And I I think I know the answers to some of these questions, but maybe start with like why do we loop?

51:14 you know, why is that even a tempting thing to do? and then you can, you know, branch it out a little bit across, you know, different hardware options and and how that might shape the future. >> Yeah, it's it's been very interesting from the loop transformers paper to that being the rumor of, you know, one of the big advancements with the GPT6 Astra that simply repeating the forward pass you know, gets you that improvement. But it's it's crazy that there was really work that I think was only being done in sort of the fringes of the open- source you know transformer community you know 2 3 years ago of taking like a llama 70DB and just duplicating layers within it. So it was taking like a 70D model and slightly strategically saying okay I'm going to just repeat these these sets of MAT balls in in sequence and turn it into like a 100 billion parameter model and you got better results from it even though it was repeating the exact same matball. So there there is an element >> imagine what you could do when you train it to work that way.

52:19 >> Exactly. So you know the I would say there and and there's been a couple interesting things years ago that preceded like the the multi-token predictor. So you know today from a speed up perspective you can add some layer you know a layer at the end that tries to actually guess what the next token will be after the one that the rest of the model has has predicted and you can do very cheap prefill on that to see if you know that's correct and you can you can basically get two two tokens for the price of 1 you know one. So so and there is a preceding thing to that idea called Medusa heads which I think is actually I I think that the reason I'm bringing this up is I think that this concept can be expanded upon in many different ways of being able to train actually you know either side cars or within layers themselves to be able to better predict what things are going to happen and and take some insight of where you are in the the layer. If you think that you are if if you already have very high confidence on what the next token is going to be and multiple layers have agreed on that, decide to early exit and go out. Or if you're really uncertain about something, actually, you know, find which layers in in network actually will have a higher likelihood of being able to increase that probability. same sort of like functions that the router does in terms of determining which expert should be selected for for on a on a per token basis. If you ex extend that concept to okay, if you're really unsure, if there's a a very large set of of equally weighted probabilities for what the next token should be and you're threequarters of the way through through the layer, do you actually just want to route back and and determine okay, we actually need a different set of of experts something like for this.

54:15 And so I think there's a ton of stuff that can happen kind of between with within the layer boundaries that we we have today that you know obviously doing this as part of pre and post training is is optimal but you know if if I had more free time and and thankfully I I can get some of it by just spinning up agents to work on these things. But you know it's it's nice to be able to have experiments of of seeing what can be done with the stuff in in terms of just open source models today. But I wouldn't be surprised if these things are already being done at scale in in the major model labs.

54:52 >> But c can you go just a little bit further into the hardware connection of that because I I'm I'm kind of coming in with assumptions along the lines of like you know just very fundamental terms like why loop at all? I think it has to do with this kind of memory bandwidth, right? Because now I can keep the same weights on the chip and I don't have to like shuffle things in and out as much. No, >> not not really. Because within when basically the assumption is if you're doing inference that you're going to have all of your weights locally in in DRAM, you're just sharting that over some number of devices. You do get some amterization, but just with the size of experts and the size of of caches on chips, there's not much of a reuse opportunity. You're, you know, by the time you're already done with the layer, you've already gone through the u amount of memory many times that of of what is in like the onchip SRAMs. I I would say the the biggest advantage is is if there there's kind of two ways that a loop could occur in ane you may reuse the same experts which would be an an advantage if if you don't have to go through like a fan out expert u u expert parallel routing decision again. but I would say that's fairly narrow and I I don't know what's being done in the big labs themselves but I would guess if you are looping you probably want to have the opportunity to select different experts than you did the first time to be able to get the improvement in quality of result. but that that's a guess from from like a hardware perspective like looping really is more memory capacity savings than memory bandwidth savings because like let's say that you trained two models with the same base set and you trained one to do with looping and so it only has 10 layers instead of 20 layers and and you know model model B is is 20 layers that 20 layer model it's doing very very simple is going to be double the size of the the 10 layer version. And if you find that the looped 10 layer version gets 95% of the same, you know, quality of result and it's half the size, you're probably going to deploy that one.

57:10 so that that's that's you're you're saving more capacity than bandwidth because doing the second loop, you still have to do the same memory fetches and do all of the same like same number of maples. there's the exact same number of bytes that need to be moved and the exact same number of flops that have to be done in that model A and model B scenario. it's just that model B the 20 layers actually has unique different weights for the the second group of 10 layers that have to be done. And so that that means you require twice the memory capacity >> when when you design a chip and there is a certain point in the design where you kind of freeze the design and you know hand it over to start the rest of the process. What are the decisions that you delay or defer until the very last moment before you decide to hand over?

58:11 >> yeah, the the the problem with that there there are some there there are a lot of features at different points of the chip development stages that we left on on the cutting room floor. and most of them are things that we still want to do in in future chips, but it's what can you actually accomplish, especially with our goal of iterating quickly. And it's better to get a product out quickly. even if it doesn't have all the features that you want, to get that out there, get all the the learnings, product feedback, revenue, everything that you get from actually having shipped something.

58:48 and so like huge amount of value in in getting that done. So I would say we we are on the side of cutting features in in to to save time primarily and it's it's not just the the design time but it's the verification of that component and you know figuring out how that integrates with the rest of the system. so, it's a, like everything that we I've cut, I'm I'm going through my mind and it's like I really hate that that that had to be cut, but I don't hate the fact that we did it because it it's what is enabling us to to move quickly.

59:32 I I would say just to to pick a a thing like a lot of the the things happening in the sparse sorry the the sparse tension space and and linear transformers etc. There are a bunch of optimizations that we could have made and done differently if if you know a year ago we thought that linear attention was was going to take off the way it has at least in the open source model space. I'm still not convinced that that's being done at the scale that we see in you know the open source and specifically Chinese open source models.

60:08 but in like the big model apps but it makes sense in the constraints that the Chinese ecosystem has in terms of hardware supply etc that that they're effectively being forced to go in in that direction and so they're innovating in in that way. and but as far as I'm aware, none of the major US model labs are doing anything like MLA or you know DSA or or any of these other techniques that the the Chinese labs have done. so I mean our focus primary customer focus is on like the the big big model of companies. So, I don't necessarily think we made the wrong choice there to to focus on the things that we did.

61:01 >> Can I do just one one more little follow up on the relationship between memory and this looping concept because I I just genuinely really want to understand it better than I do, I guess. how does how does KV cache type things relate to this? when you said like okay if you loop you can have you know half as many layers you have half as many weights does that also mean the KV cache is half as big and does that >> yeah so the per user yeah so the per user KV cache portion is you know on a per layer basis so so for the exact same reasons that I was saying all the weights you know are cut in half in that that simplistic example it would be the same for for KV caches or if you're doing like data gated delta nets you know the state matrix as well the the reason that's it's still a capacity win because even when you do that loop again you still need to fetch the KV cache it's the same KV cache you need to fetch so it's the it's reduced capacity but you still have to do the same total number of bytes moved if if you were doing 20 layers instead there there is quote unquote a second fetch of the same amount of data that you would have had if you were just doing it as as 10 layers repeated twice.

62:28 >> Gotcha. You mentioned how much AI is helping you on the verification side. in a notable interview Jensen Hong recently gave with Ezra Klene, he talked about how Nvidia spends something like 20% of its time and energy designing and then 80% verifying, validating, you know, ensuring long lifespan, reliability, etc., etc. what do your ratios look like on that? And you know what more color can you give us about like specifically where AI is really helping? I think in general there's this kind of question of like can AI now make you know like meaningful new discoveries sometimes they are obviously are but you know not in all domains.

63:16 What about in your domain? and does that matter if it can, you know, accelerate the other 80% so dramatically that it, you know, it still potentially creates like a phase change in your cycle times? >> Yeah. so for us, I would say it's a little bit different because we're starting from scratch. So there's a lot more like base layer base capability design that we have to to to start from scratch. So for you know this this our our first full custom silicon osmov it was about a it will end up being somewhere around maybe 60% you know design focus for us to 40% verification if I'm just thinking about from like the time from project start to tape out you know RTL 3ze was you know roughly halfway through it and then still need to, you know, fix fix some things in RTL that you that you find problems for in in the the back half. I would say like the probably most amazing thing that I I don't think would have been possible 3 or 6 months ago and and only really became possible with GPT6 Astra and and open now Opus 5.5 that that we're we're leveraging a lot in in this task is we we for for doing our verification work we use Cadence Palladium emulator systems. These are big giant racks that are full of custom AS6 that are specifically designed to just emulate do gate level emulation of of silicon.

65:02 And so this by itself is something like a hundfold improvement in you know the speed at which you can can emulate your chip design and do these verification tests. The crazy thing about Aster and five and Opus 5.5 is this fact that you're you're the the agents themselves have been able to completely do closed loop iteration and testing on on on this. So you know I I think people thought it was crazy but you know handing you know these very powerful agents you know full keys to our our internal infrastructure and saying here's Palladium piece and I highly highly doubt that these models had you know Palladium documentation in in their returning thing especially because a lot of documents are new as as software updates but the fact that it went and looked up the documentation when we're looking at these agent traces the reasoning traces and tool calls etc. It's reading the full documentation effectively compressing that itself by by you know reading the PDFs generating its own you know markdown files of of cheat sheets of what to do and then building out all of the testing infrastructure and its own harnesses for being able to access this. It's it's mind-blowing. And so while if you're just doing regular software RTL simulation that's running on the order of like 10 10 hertz for us in in in those cases if if you remove a lot of the debug pieces that slow things down etc. We can run the full chip emulation on the order of 500 kilhertz. And so you've got actually several orders of magnitude, you know, speed up and improvement from from that gate level hardware piece. And this just gets us closer to this this RSI loop. And right now that's really just focused on implementing test programs, finding cases where they fail, and then writing out reports which you know get reviewed by other agents and and by humans in the loop. But I I would say when this really started working, you know, six weeks ago, we roughly was beginning of September when when Astra 6 came out, it it was a huge like step function improvement from, you know, GPC 5.6 where it it could not do that full closed loop itself. It still required humans at different stages. So it's it's going to be amazing to see how that same like paradigm shift can apply for every FA part I like I don't see that not applying to every part of of the chip development process. So one of the commentaries on AI in general is that it's a lot of like process following work especially like for example taking a a manual for a particular piece of software or hardware and implementing it or following the rules behind it and the criticism has been that it is not really true new innovation. Have you have you seen any signs of really some unexpected innovation rather than just rule following in that sense?

68:29 >> I have not yet. I would say that's the the most disappointing elements of all like I'm as you can probably tell I think all this is insanely amazing and I think it we're going to continue this exponential trend. I I'm I'm somewhat proud of the fact that it has not figured out some of our cleverness of what we've done. Like it's it has like questioned and like this seems like a bad design decision until it gets explained to it why something is done or it actually runs test and sees and understands oh that's why you're not doing the traditional way a systolic array is done. like I would say and of course models it's a night and day difference now versus a year ago but I remember using Opus one of one of the early Opus 4 releases I I forget which one but was having a very long dialogue and I remember how amazing I thought that early Opus release was that I thought you know well we're we're days away from AGI or something you know and and now I you know wouldn't trust it with you know my dog. but I I remember the it it looking at some of our documentation like our internal you know design specification and asking questions and and having it compare and contrast how it thought about you know versus you know a TPU or you know traditional systolic array designs and like it pointed out differences and basically you know isolated and had a just regular opus thing you know design something with these specifications and and everything and it would take the traditional approach that's the training data it had. And so it it does make me feel a little proud even how just how unbelievably amazing the models are today that it won't come up with the I would say the ingenuity that we had as as human designers. Do I think that will hold for another six months? not not sure. But just because I in part why I actually think it's not going to be like the agent will come up with a better idea immediately. I think the amazing turn of events that we've just had in the past few weeks in terms of us having access to these models is is the fact that just because it can iterate so fast and do experiments and things by itself, I wouldn't be surprised if if it would come to the same conclusions we did or maybe you know make a better things than what we would do by ourselves just because it's going to be able to iterate so fast and and go through so many different design possibilities. I I think like the just the story of of OpenAI solving Navier Stokes like the same approach of yeah if you take 10,000 agents and and send them off for for you know millions of of manmurs worth of effort it found a solution I I it's brute force it's expensive but it I I think that is a valid strategy that has only become cap possible in very short past.

71:38 >> Not many people left standing in the AIS can't do what I do category. So, congratulations on your place for as long as it >> I'll hold it for a couple more weeks at least. So, but I I'm very eager to u you know, I for one welcome our new robot overlords. how how does this translate today to like token spend versus salary spend and then you know another kind of not sort of a money question but sort of a you know just kind of environment question obviously just raised you know nearly a billion dollars what is it like to go try to get capacity from manufacturers I imagine that's got to be very competitive and probably that's where a lot of this money is going to go even with, you know, nearly a billion dollars, you're competing against, you know, just massive whales all over the place that are throwing, you know, unbelievable sums around. So, yeah, how are you how are you spending internally on AIS versus humans and and how how are you managing to win, you know, a space in and I don't know who's manufacturing for you, but, you know, TSMC's heart, so to speak.

72:52 >> Yeah. so on on the token spend side it's become you know our single largest non manufacturing line item very very quickly. >> bigger than bigger than human salaries. >> Yeah. just recently. And it's interesting it it eclipsed human salaries. It came back down I'll explain in a moment. you know, our our token spend was, you know, equivalent to like our our monthly spend got was like the equivalent of having like the the token spend across all all employees in the company. granted, the company was half the size six months ago. but six months ago was like equivalent of a single employee. So that was very easy to budget and say, "Oh yeah, we've got claude plus plus GPT and a little bit of cursor and other stuff all for the cost of a single employee."

73:50 really started in in in May. you know, GPT 5.5 Opus 4.6 were just massive massive steps up and that Oh, you dropped off for a moment. >> We lost you. Come back. >> Oh, you're back. Oh, you were back for a second.

74:21 >> He's still He's still in the room. So, I think it's just >> Yeah, if he's like me, he's got a faulty cable that keeps coming loose at in opportun difficulties >> getting up. Yeah. All right. Am I actually back now? Okay. >> Yeah. >> sorry, it dated from my phone because my computer was having issues earlier. and that's what happens when you get a phone call. >> yeah. So, our token spend just massively increased where where, okay, six months ago is equivalent of one employee. in in June it started to now be multiple employees and now it's more than like I I guess but I think yeah so so our annualized expenditure so we're spending well over like we we were peing you know in the in the days after GPT6 and we're really trying to to push its boundaries etc we were we were spending over $100,000 a day on tokens that tempered down a little bit in the weeks after as we optimized a bit and and just weren't trying to run as many parallel experiments. but I would say the big thing that was like, you know, a saving grace was, you know, Opus 5.5 came out and in a lot of our tests, not everything, it was doing better than than Astra and was, you know, a quarter the price. and so you know that this is the the great moment to be alive that that that we're in right now of because of this competition in the market. You know we're we're not worried like we didn't really we didn't tell anyone to cut their spending or do anything even though we were seeing this exponential growth in in our our token spend. it it was really just the fact that we we both felt that we were getting good ROI on it. So we weren't going to to taper it. but also the fact that like we have very high confidence belief that costs are going to come down and yeah we're we're using you know the the major model labs you know their their models for everything today. but you know we also have faith that we're going to be able to move things to you know local models running on our own hardware. I would say like we we do a little tiny bit of this with GLM 5.3 today.

76:53 but personally like my philosophy is I really don't want anyone to be running like a real software or hardware development task on anything that is less than the best model. like I don't care if it's, you know, a tenth of the cost, you know, per token. it's just not worth the the the the expense. So >> So that's fascinating.

77:23 we're at time. I'll let you get back to work, but I would love to hear just very briefly what it's like competing for manufacturing allocation. >> Yeah. Yeah. I mean I would say a great position that we're in and this is largely you know our our COO Bernie Sardinia Pron is his lucky number 13th startup. he has the the you know honor I guess of having been the the very first customer of TSMC in North America back in 1987. And he's done now 13 startups all working with TSMC over these decades. and so we were able to have a very close connection with TSMC from from the startup company and we're you know a direct TSMC customer which is very rare these days. If you talk to basically any other AI chip startup they're all going through a value chain aggregator like the Broadcom or Marll or you know any of these others and we're we're we're direct with TSMC. and on top of that announced with our our series C is that we had Venture Tech Alliance which is a VC fund that their sole LP is TSMC. so through that way you know we have TSMC as an investor as well.

78:43 So, we've not had really any worries. And even with our like aggressive production ramp, you know, TSMC's TSMC is the best founder in the world. They're the ones scaling capacity like crazy. we're not aggressively chasing the latest process nodes. We're we're sticking at N3P. And so there's there's a lot of capacity out there for us. And we we really solved their core supply chain bottleneck which is is memory by kind of going in the opposite direction of of most people in the space.

79:16 >> Amazing. >> Yeah. Incredible stuff, man. Congratulations on the round and the early success. I'm sure there will be a lot more to come. I don't envy our editing agent who's going to have to try to find highlights out of this section. It's going to be a question of what will it cut. this is was action-packed. So, really appreciate it. And you've got a great great guest coming up next. So, >> yeah, indeed. Indeed. >> All right.

79:39 >> Thanks for being here. Great to meet you, John. >> Byebye. >> Great meeting you. Bye. >> That was that was incredible. And let me pull up our next guest. and >> graphics package. Our next guest is Swix or Sean Wang. he is the co-founder and CEO of AI Engineer, which runs conferences and workshops for people building software with AI. He's also co-founder and editor of latent space, a publication and podcast about AI and the people developing it. And he's also an adviser to Cognition, the company behind the coding agent Devon. he also founded Small AI, whose AI news project combines automated reading and summarization with human selection. The next AI engineer New York conference is scheduled for October 12th through 14th with a focus on applications across banking, investing, insurance, and financial technology.

80:46 Let's bring up Sean. >> Hey. >> Hey guys. >> How are you? >> Can you hear me? >> Yep. Yes. You're good. >> okay. All right. yeah, thanks for having me back on. I was just saying I hung out with Tom for two hours at Devday, and he's such a fountain of advice. He's very low-key, very humble, but he's building a a monster chip company in record time in at speeds that no other company of his like size and cohort has has done. So, yeah, definitely want to keep those two.

81:33 Yeah, it helps to be a galaxy brain genius. I guess one of I was just joking, one of the few people that can still credibly claim to be able to do something that models truly might not be able to do, at least for now. But even he it sounds like doesn't expect to hold on to that for all that much longer. so you've got another AI engineer coming up. What What are the latest kind of trends that you're watching? I mean, I know you are the benevolent dictator of the community and always scouting the horizon for like what are the new topics that you want to get ahead of? What have you decided are the big things that we need to be getting ahead of that you'll be featuring at the upcoming event?

82:16 >> yeah, so first of all, both of you are invited. I keep trying to invite Nathan specifically and then he's always like on some family trip. but you know, some at some point AIA will be more important than family. And we we do four we do it four times a year. Basically the idea is that every other conference is once once a year. that's horrible because AI proceeds at superhuman speed. and so the the next one is New York. This is our basically our New York flagship and we we we're coming back this our third New York conference and we're coming back with a focus on finance this time. which I'm pretty excited by, but it's also somewhat of a risk for us because you know we're very on on a home turf talking about code and coding agents.

83:01 but I definitely think like as far as like the other verticals that are breaking out like code. I would say finance is the next you can clearly see it from the labs who have all prioritized like finance plugins and finance verticals for their FDE operations all that good stuff. But I have a particular affinity to this one because I used to work in finance. So I I say it in my sort of opening address that I I'm drafting right now. my like it has basically taken the last 20 years of my life to build up to this because I spent the first seven to eight years. First of all, I went to college at Wharton.

83:40 Oh, cut out. Come back Swix. Come back Swix. So, >> everything's working great today except our guests hardware, it seems like. >> Oh, >> you're back. >> I think I got kicked out by Codeex who didn't like that I was chatting. so that's a C codeex alignment fail. I I just have stuff. By the way, we I'm like a super token billionaire now.

84:15 oh. And then then they kicked out Pash. Is he Is he here? Can you still hear me? >> Yeah. >> Yeah. >> I'm a super token billionaire now because like I just have always have stuff running in the background and operating my windows and but that was the first time to actually shut me down during a call. I mean usually I use Zoom but like you know you guys have your have your weird firecoded platform. U anyway sorry. Where was I?

84:43 Where was I? what was I started? yeah so finance I spent the last 20 years building onto this mostly because I spent the first half of my career in finance. I was a sellside options trader trading interest rate and currency der derivatives and then I switched to the buy side in in hedge funds doing basically quantitative portfolio management. and then I switched to tech. so like for people who only know me for my tech career there was a whole finance phase and so in some ways like I'm very much one of the few people in the world qualified to sort of bring AI finance together and it just so happens that I think this is also genuinely the next sort of breakout vertical that people should be watching.

85:25 so one when we talk about AI finance I mean with the AI code obviously the the work product is the code itself the code and the the functionality behind the code. When you talk about AI finance what is the work product that people are looking at? Is it code? Is it spreadsheets? Is it presentations? Yeah, spreadsheet actually I think is I considered having just a a entire track that is only about Excel and automating Excel and building models.

85:59 spreadsheet is very very much part of it. but you know it's just in the same way that I think code has a development life cycle and so there's like you know the CI/CD side of code, there's the app side of code, there's like the testing side of code. in finance there's, you know, back office, middle office, front office, there's insurance, there's payments, there's commercial, consumer banks, investment banks, and then there's hedge funds. and then there's all sorts of other information services providers that that provide sort of horizontal services for all these guys. so I think it's a little bit hard. Basically, this is knowledge work, but the most verifiable form of knowledge work, right? so it is like very logical that like you should that AI started at code which is verifi very verifiable then then you sort of proceed on to the next most verifiable domains in the so you guys are doing so much on latent space right now which I'm enjoying you're moving into science which I would emphasize as the that's always been a passion of mine AI for science generally. I'm I'm glad that you've joined into the AI for science coverage when it comes to just engineering, right? I mean, a couple years ago it was like all the engineers were kind of figuring out how to use these coding tools and you know, of course, they've gotten better and better. just the with the last couple models, I have felt like another phase change where I'm just like it's so effortless to create so many things that I really am bottlenecked on my ideas. you know, it's I'm getting things that basically just work the first time or if there's like a little missing piece, you know, it's just like one quick correction to fix it. What is the vibe that you're getting from people who have been software engineers building software applications? are they still feeling like super empowered?

88:00 Are they starting to feel threatened? Like you know the for a while people were like oh well there's going to be you know so much demand for more software that like it'll be fine you know there will be even more software engineers. What's the vibe you're feeling among those that are >> you know most in the know right now about how those things are actually playing out? Yeah, I think college kids are a bit worried. but other than that, like if you're relatively plugged in and very capable with AI engineering tooling, you're in more demand than you've ever been because your expected value is higher than it's ever been.

88:39 so this is definitely one of those Jevans paradox type things where the the sort of cost of software creating software has gone down. Therefore, the demand has increased a lot and but but spec for a very specific kind of demand for people who can manage coding agents productively instead of producing a whole bunch of slop. and I have been in that situation. So I am both an engineer but also an employer of engineers and people who are not engineers who are vibe coding. and you know, the the the phrase I've sort of recently taken to is I don't want to pay for someone else to go through LM psychosis, right? Like I can pay for my own psychosis, that's fine.

89:21 But like when you work for me and I'm paying your tokens, you better be like actually producing thoughtful stuff. And I have two or three employees right now who are basically under performance review because they are just giving me cloud slop. And that's really bad for them. And they don't understand it. They don't see it. They're like, "What do you mean? I think this this is perfectly fine." And I'm like, "Well, you're not like producing any value. I can't just prompt Claude. I don't need you."

89:46 and so I think there's there's that perspective. Most software engineers, >> you describe that gap a little more. Like I'm probably also mostly producing claud slop. >> What's the delta? trying to go for depth of insight rather than breath of coverage, you know, of like, you know, my my my YouTube guy will be like, "Oh, yeah, your your videos went up like 14.7% week on week." And I say, "Why?" And he doesn't know, right? Well, he's actually never watched the videos. He just used Claude to like analyze the videos. and Claude doesn't have video analysis or or like any contextual understanding or anything like that. And like I'm like dude like you are a human using these tools. You cannot let these tools think for you.

90:37 You got to actually think to to use these tools. And because the the tools simulate thinking and are so and and like provide so much false precision. Who gives a about like 14.7%. Like what did I do right? What did I what can I do better? like and like clearly you know cloud or GPT is not tuned to most specialist tasks like hey help me optimize my YouTube video. Like it's not something a thing that anthropic really cares about. and why why should they right like but like yeah I mean there's lots of areas for human domain expertise but then they send it they like substitute it for the like the lowest energy lowest cost means of production which is chuck something in the cloud and I hope that I don't notice. I I had a I had an agency actually I I was I was like you know how these like launch videos come up and then they're like the agency goes like oh yeah we made that.

91:32 And so like I I hired I I like reached out to one of those agencies that I that like to launch videos. I was like okay well I want to work with you. And they were like blah blah blah give me your requirements and they came back to me with like very clear cloud slot and clearly like we were not like maybe they have some talented humans in there that serve their high priority clients but then the rest of the customers that they're serving just cloud slop and I happen to be like have high enough standards that I was like first of all this like really doesn't get what we're going for and I think like to some extent cloud is fine right as long as it's like useful and good. but like this was this is like clearly not the case for for for this particular project that I was trying to hire for. And they they were just like, "Oh yeah, you caught us. we're going to move on."

92:17 And so like they're just hunting for for people with low standards or like just don't care or who who have low enough money they're just like, " I don't know anything about this business. I'm just going to take a shot." Whereas like I think the people with taste just care about the product and the the end product and will just have like an im immediate reaction to like yeah clearly you have no relationship whatsoever with this core content and you're just doing cloud stop and and I think that's the that's the best I mean you guys you you have a sort of clipper right like you have a agent clipper and all these things and I'm sure it takes a lot of tuning to like get somewhere but like probably if you put your full human attention on clipping like you do better like it's hard to that's hard to articulate what that gap is but you know it >> is that so this whole taste question to what degree is it coupled or correlated with a background in software engineering though I like I do feel you're absolutely right at least for now and along may it continue that there's a role for the human to kind of exercise taste and judgment and and push for depth of insight. But if you were to be in the market to replace a couple of people because they're just giving you cloud slop, would you feel that you would need and maybe it's like a ratio thing, maybe you still need a little bit, not as much as before, but how much do you need this like experience building software to come in and do that job versus just like product sense or like high standards in general, you know?

93:56 because I'm not reading any code. Most people I'm talking to aren't reading any code. There's like I think probably still some you know some stuff around like hyperscaling you know platforms that like I don't know that the agents can do super well. but do you care about the part of the resume that's like oh I've been coding for x years at this point or is it not so relevant? I I don't I don't care so much except for let's call it security roles and for anything sort of backend scalability like you want I I I really need you to know the differences between GCP and AWS and like what I can do on each each one of those things. you know like I just recorded an interview with the superb base founders who who you know are scaling Postgress to a scale that we've never seen before. and you like yeah good luck trying to hire somebody to scale Postgress like you know with with a complete self-managed sharding solution that works at YouTube scale.

95:01 there's only one person in the world that can do that and they hired the guy and like yeah like so like you better you better go and and you know make sure you don't lose data that you like don't have don't cause downtime that you scale you know economically because basically agents consume databases something like at a 500 to one ratio of of what what human developers do. then yeah so so like other than that like you can mostly buy code and then it's then it is about taste. and it's not so much about length of experience actually sometimes experience works against you because you just have a set way of doing things and you don't understand how to do more than one agent at a time. yeah at this point like you should be relatively comfortable juggling like five to like 10 ongoing multiple things. And it's it's definitely true that you're much more of a manager than a a individual contributor. I do think that people who look at data rather than code are more valuable these days. so being able to say like here's the logs, here's the traces, here's the schema.

96:08 here's the input output capturing it turning into an eval. all these are basically the sort of the merging of AI engineering and ML engineering that is that's happening that people need to upskill on and then like managing scaled up runs of that like right I think that the the person that has let's say 10 20 years of experience in software engineering really reviews every line and tries to make every line make sense whereas now we just need to make the the modules make sense and I I I can allow slop in there because it helps me to go faster as long as I contain the slop in things where I completely understand the whole system.

96:47 where you go wrong is you have too many modules, too many black boxes where you don't even know and actually the code also gets confused. So just as an example, I had you know I'm I I'm making my own sort of Slack competitor and one of the and like I saw a bug where like the the messages weren't load loading. I refresh and the messages were loading and I refreshed again and the messages were not loading and it's like well it's the same exact code what the hell's going on? Turns out there's two code bath code paths and there's a race condition. Why? Because two different coding agents worked on it at different times and because they they just like made their own thing. So like a human would never do that. an agent would potentially sometimes do it because like sometimes just things fall out of the context window. but like you need to have that oversight of the module and then whatever inside is inside inside of the module can be a black box.

97:42 >> So can you kind of describe so we're obviously seeing this transition from humans being the primary users of the internet to code to agents really being the primary users of the internet and as you pointed out 500 to1 ratio on things like databases. What are the other parts of the stack that need to be scaled at this point? Because it seems like we're going to end up overusing you know many many of the primitives are going to end up being overused. So which parts of the of the stack are really poorly designed for agents right now? compute like we covered this in like January, February where like the GPU shortage was spilling over into compute shortage.

98:34 if everyone knows the GPU shortage, everyone knows the memory shortage. but CPU shortage is the most consistently reported thing among all the people that I interview right now. so what does that mean? It does mean that we need sort of more fluid compute which is the versel term for it. more sort of different forms of serverlessness where like it's very ephemeral you can sort of execute and resume because you know agents need some time to do LLM calls or make network calls whatever but basically longunning states that are resumable. I used to work on this at temporal by the way. So like there there is existing technology that's doing super well and this is why temporal is doing super well which is there they've just been the longest in this game.

99:20 but there there are other competitors like restate that also work on a sort of orchestration problem but I think beyond that you really start to get into two things which I'm really thinking about like sort of bandwidth for networking where the latency starts to really matter and that's a opt that's an optimization of like well what point of presence are you communicating with or what region are you in versus like actually how much compute should you be doing in cloud versus locally or sort of edge devices. that is definitely happening a lot for the robotics companies that we're talking to. and basically don't treat it as just a robotics problem. The robotics are just the first harbinger of what that you guys will eventually be doing if you if you consume or do enough inference at their scale. so so there's there's all that there's all that and then finally to me on a more very very personal basis bandwidth. so like how many tokens per second can I get bas both on a single call basis and I just aggregate across all my calls and then aggregate across my company because I have so many I have enough people managing all these calls right so right now maybe my tokens per second personally as an individual is something to the order of 2,000 across all the agents that I'm managing and cognition is roughly two trillion right? Like that's like the rough, you know, this is an estimate. I this like vibed number don't don't treat this as like a real indication of the the spend.

100:59 It's in a rough ballpark. and like I think a scaled AI company that is like probably your overall total bandwidth of like productive AI usage that sums across agents and humans. >> So let's talk a little bit about that. you you've already said that you've had issues with I would say poorly spent tokens, right? Tokens which were you know parts of your organization that perhaps didn't have the critical faculty of evaluating the tokens that were being you know emitted with sufficient you know thinking. So how does that work as you kind of start to scale organizations and you get less ability for a human manager to kind of evaluate whether their subordinates are spending tokens properly?

101:58 >> Wait, so yeah, I mean yes, you need the ability, but like are you asking if I have a solution? >> Yeah, I mean how how how do you think that would work? Like would you would you would you see basically like maybe organizations are capped in terms of man of you know the amount of manpower they can you know successfully use because you don't you need people to monitor the humans who are using the tokens.

102:25 >> So so I think this is this is coming down to just management philosophy in general and I don't have any sort of core competitive advantage about that. I do talk with a lot of CEOs about who going through this problem. so on one hand yes you know I would say that the leading companies and the names that you all heard of are hiring have a hiring practical hiring limit because they just don't have enough managers to manage those people that they're hiring.

102:53 It's still pretty good. it's roughly let's call it something to the order of like a 100 to like 400 people a year. you know but like if you ask them to onboard 4,000 they would they would struggle right >> I think the other perspective so like that is normal like let's call it seed to series D series E hiring at the large large like more than 10,000 person company level nobody cares about the individual humans inside of a team anyway and actually you just manage by numbers like hey what are your three top goals per team and are you delivering those three top goals to me so that I can plug you into the rest of my org inside of that black box? I don't care.

103:42 Do whatever, right? I'll hold you to some cost efficiency metric, but hire the humans you want, spend the agents that you want, I don't care. and that actually is a much more scalable like that's like the bitter pill for bitter lesson for organizations which is you actually don't need to know every single person in your company like actually you just need to know your your team leads and your team leads need to know them but like I'm just going to fire the entire team if the team team lead is not performing. you know that that's like the really like you're managing Walmart, you're managing like Johnson and Johnson like Proctor and G like there's there's just no way you're you're doing this otherwise, right? Like which is fine like it's not very humane, but like I I do think that you know if you want to manage a very large human scale organization that like people have done spent centuries building up the disciplines for this. I don't think anyone's like particularly inspired by like this is how you manage you know a 200-year-old legacy Fortune 500 but this is this is what people do and like I think that's perfectly fine like everyone there like business managers they have P&Ls they have go other objectives that they that they optimize for actually Amazon are very like this right like you just have business units and every every business has a GM which is basically a CEO and they run they're costed they're they're run they're plugged into an overall strategy And I think it works well. I think probably more companies should try that. That sort of bitter lesson scaling of humans rather than let's say like making sure that everyone is like well attended to. you know at at the very limit you know take the take the one of the richest men in the world Jensen doesn't have any one-on- ones with any of his direct reports. he has 60 direct reports, right? Like this is like very contra to everything I've ever done in my tech career where like maximum you have eight. so like it it is just a different type of scaling and like I I do think that more companies should probably embrace that but it is a very big cultural change.

105:43 >> Indeed. let me segue a little bit. you did a kill my SAS project and >> not done yet but yes it was very it's actually very successful. >> It's actually very success. So so why don't you tell us a little bit about it? It came about because you had a product that you were using for AI engineer and you found that they were kind of over billing you or you you you were not happy with how much you were paying for the amount of you know work that they were doing and you decided okay let's do it over a weekend. I'm going to have this kill kill my SAS project. You're going to put up this SAS there and you know you're going to encourage people to come up with solutions. Is that is that is that the is that the kind of summary that you know you started off with?

106:26 >> Yeah, basically you can call this like you know a bounty on mid-tier SAS that should not exist and this definitely was one of them. basically like you know half a salary or one third of a salary for a for a piece of software that I don't own that nobody enjoys using. Let's just throw that out, right? Like I I let's like you know if I'm going to spend 40,000 in tokens, $40,000 on this SAS subscription, I can spend that in in amount of tokens and that buys me a heck of a lot of tokens.

106:57 and so that would be the ideal outcome. And the the only problem is that well for for people or first of all for people who are involved in kill my sess who haven't receive received a results yet sorry we're working on it. second of all this is the problem especially with like launching it without planning. like I was I just did it on a vibe to my entire audience. we have so many submissions that we then have to eval them. and so eval because so like we're no longer checking like here's like two or three requirements that you fit. it is we're checking the the entire UX of the entire flow from three different perspectives of like you know organizer attendee and sponsor or a speaker and like then you need different login then you need like all the different workflows and like to submit different applications across the the different formats that we do it is actually like the the the way that you eval software like it gives me an appreciation for like building frontier evals which is like well if you're if you're going to basically If I code or manufacture entire programs from scratch, then you need the ability to eval this. And I actually shipped evals for to help my participants on on the second day. But for really frontier things, the human just has to play play test through all of it >> on top of their regular day jobs. So yeah, that is the tough finding. And just be aware that like if you offer like a large bounty like we offer $10,000 as the prize you'll get a lot of sub submissions because people want to vibe code for fun and a lot of them will be low quality, >> right? They will just be like Claude, make no mistakes, go do this and like and then they'll make mistakes and they'll just submit. And so like the the sort of verification load is very imbalanced because they spent zero thought on this thing. they just threw it in there and then and then you have no idea if it's like low quality or not.

108:58 >> so yeah that that that'll be my sort of general lesson but overall very successful and the team initially so I have one of the most adversarial non-AI teams in AI. Okay, I hire like what is like what is my business? I hire event professionals who are like super old school. They do everything in spreadsheets. They work with unions like they have to worry about like placements of like the meter boards physically in like locations right like these are not glamorous like high-tech jobs then they're very super suspicious of anything that is new and vioded and techy right so I hire these guys and then I I like make them work for AI right that that's my job right and so initially they were like we'll never use this vibe coded thing like this is like like I want to use the tried and tested stuff that like has has been working for for Microsoft and all all these these other guys. and then they saw the quality of the submissions, then they s they looked at their existing platform, they were like, "Yeah, okay, we're going to switch, right?" Like it's like that initial hurdle needs you basically need to provide the sort of proof of like this is feasible and and then then the ongoing thing which is this is one of the the the benefits of me partnering with Cognition on is I just give them give all of them access to Devon to modify the code. So anything they don't like, they can just request a change and they have it pretty much in one or two hours. which they've never had before. Just to give you an idea of the the the the way we work with these SAS companies. right now when we request a change, they'll say, "Okay, that sounds pretty cool. It's on our Q3 road map."

110:30 >> Yeah. Yeah. >> Right. >> Someday. >> And like we don't have confidence that it'll actually be done in Q3. Like and by the way, this was done. We requested for a thing in Q1. It just landed. like like so like SAS is quite cooked if you're if you're mostly a CRUD app. I would say that there's a lot of UX issues left. there's you know if you look at Sweetbench and you're like oh wow we're like 90 on Sweetbench. You're looking at the wrong thing my guy like you know if you haven't actually tried to completely vibe code a SAS that you use you don't understand how models are still very bad at all this.

111:10 Yeah, indeed. it sounds as though you had to set up a SAS in order to evaluate your kill my SAS submissions. >> Yeah. Yeah. So, we had we had the the the the thing that we wanted to kill, which I which I tried to sort of keep them anonymous for for their protection. but yeah, then we just like completely evaluate against like a sideby-side playthrough of each each of those things. you know, that's what that's one of the beauty of like computer use being good now. is that you can basically just do a point andclick and screen capture and then compare compare compare completely document like we have a 200page document that that documents every flow and then just make sure that each each of that is represented in your in your clone and that's mostly it. the the the extra thing that like we put on ourselves and to to because this is also a hiring challenge is we asked them to also put in some human taste into like well what what would you change if like you could do things differently? So you are allowed to go off script and like why why would you not?

112:06 >> So how can you synthesize this experience with the first statement you made around if you're decent with vibe coding you'll never you you know you've never been in higher demand. Like it seems to me that we're headed for a world where we're going to there certainly there is a lot of latent demand for software that hasn't been met yet. But I have a hard time seeing how we don't kind of pick a lot of this lowhanging fruit and then end up in a sort of a lot of companies go out of business. And also I'm kind of thinking like layoffs from like meta, you know, that's another kind of canary in the coal mine. Like are we going to see mass layoffs from big tech?

112:47 >> How long can the party go on? >> Yeah. So this is the the thing I had to deal with when I was considering joining Cognition. because I was like, you look, you guys are at a $2 billion market cap evaluation. like you already, you know, you did it in two years. Congrats guys. Like this is faster than I've ever seen in my career. And now there, you know, 48 billion. the I mean the simple answer is like if you look at it as a fixed pile of software engineering, then that is the conclusion that you will arrive at.

113:21 but if you look at it as the point of like my co my competition or my TAM is spreadsheets. All spreadsheets made by anyone for with any sort of productivity tooling can be turned into custom software for them to put on rails and with a beautiful UX and with like you know automations or or then then then the then the the demand for custom software becomes a lot larger. and more beyond that. if it's all if it's phone calls, if it's emails that go back and forth, and and you know eventually physical meetings between people all of that can be incrementally turned into more and more and more and more custom software and hardware and and models by the way.

114:10 and so that basically like we're just growing into the long tail, right? Like why is Salesforce so damn big? is because people can customize Salesforce, but at some point they need to stop paying the 300K a year baseline subscription and spend 30K building their own personal CRM and they're good. In fact, they're more than good. They're a happier year because they can now modify it to whatever they want. And like so you should have an idea of like the sheer diversity of human needs and desires be over and above just ego. Like there's just a baseline ego of like oh I I made my own thing like this is my thing like and this you know you have your thing I have my thing there there's a lot of that right like there's a lot of like well you know I just want my own thing and so like there's some demand from that but beyond that like you immediately start customizing like I like the buttons this way I like the flow this way you know I I like my you know whatever like the the sheer amount of customization that people can do we don't even have the same amount of customization in our software as we do our handbags.

115:18 What the hell? You know what I mean? Like imagine like lux like clothes has been around longer than than computers, but we have so much customization in clothes, so much like demand in like different brands and different var varieties. Don't we don't really care about benchmarks for clothes. It's just like this is my the style that I like. I think what software engineering is basically only done when we basically have like Kate Spade versus Louis Vuitton versus Coach for software.

115:43 like we we're like not even at the point of like looking at features or like price. We're just like what's the brand feel and like I I I identify more with this brand like that. That's what really sort of true commodity software is which is you start you stop differentiating on substance and you start more differentiating on vibe. which there is a lot of vibe right now but like you know you know what I mean like I I basically trying to get the sense of like the the the pie of software is expanding because it's so easy to customize and therefore it will we just need to get the infra there and you need to scale inference the the real limit the the reason I said you know this whole thing about like scaling my personal token bandwidth versus scaling my organization token bandwidth the real limit on all this is the the the sum total of humanity's total token bandwidth, right? Which which is like literally the amount of silicon we can produce. and that's that like that's going to dictate pricing and availability and and rate limits and all these things. and so like the the real players in the room are only focused on that and then the rest of us are just fighting for scraps within the fixed pie that we already have.

116:52 >> Let's let's talk a little bit about fixed pie and non-fixed pie. So we had this release called Jev which was a calibrated decision-making model and it was it went viral. I think a lot of people were playing around with it. OpenAI during their dev day introduced a decisions API which seemed to be a Jevlike you know undertaking. What do you what do you what did you think about the Jev release and have you used Jev? Has that been useful in reducing token spend internally within the organization?

117:30 >> yeah, so not only have I used it, I was friends with Yogo for four years before the launch. and so I I I insisted that we were the first podcast that he did about Jeff and that I mean that is now our number one podcast ever within space. and and then secondly, we we're also in the trial for the decisions API with open eye. So we have a podcast on on the decisions API as well. which is not the same thing there. that that one is basically a Luna Luna with a a sort of different inference layer on top of it.

117:59 generally it's a very good idea. I wish it's it's so good that I wish I thought of it first but you know do earns the credit there. and I think it is probably like I think is what's also true among the sort of people that I talk to are that a lot of the use cases you see on Twitter are probably overdone or dumb in the in the sense of like you could use any other small model to do exactly this this thing but because Jeff is hot right now you're just creating a bunch of content about Jeff. or the other perspective is this is a another classifier. You can train classifiers all all this whole time. but you can but but you know now now it's sexy. So like now now everyone just people are throwing classifiers or using a classifier model where they would previously use an LLM because they didn't know anything else other than LM. So that's great like this is there's there's additional options in in the in the toolkit.

118:57 I think basically what the whole like people are missing the whole point why they don't call it why Jeff or DGO doesn't call it a decision model. they call it system one model because it's meant to be integrated inside of your software to go into the background of your software. as unremarkable as an if statement is the is the rough idea inside of your software. And so people aren't really doing that. they're like, "Let me just let me like replace this the kind, you know, every sort of point of structured output with with Jev." And they're probably going to run into issues because they don't care about intelligence. If you notice, most of these use cases aren't really evaluating like, well, does it actually play it very smart? Does it do any planning? None of that. it's just can you make a decision quickly? And like, yeah, you could always make a decision quickly. So, I think there's some nuance there. that's that's maybe being missed. Doesn't matter because I think the category creation was so successful that is an overall good thing for the industry.

119:58 >> Well, maybe offline I'll see if you can help me get off the Jev wait list so I can get into it firsthand myself. >> Send them Yeah, you got to send them good memes. but yeah. Yeah, they they are they are constrained by compute. I will say that >> I and I don't doubt it. I I didn't I don't know the numbers on that podcast, but I I can just say as a listener, I think it's not just one of my favorite of your podcast that I've heard, but it's one of my favorite podcast truly ever that I've heard.

120:28 because his passion and difference of vision for where AI can go, how we can work with it, what the what the experience can be like was just so invigorating that I felt like I wanted to, you know, go run through a wall or whatever when I was done listening to it. And that's before even using it. but I mean it's really, you know, somebody who has has clearly put a a ton of time and energy into developing his own positive vision and it was it was no hearing him, it was no surprise how much it it resonated with everybody. I know you got to go. We can let you get back to work. You got a lot coming up before the event. Why don't you tell us one more time when, where, who should come? and then we'll let you get on with your day.

121:15 Yeah, it's next week in New York October 12th through 14th and we're about to sell out tickets I think probably by the end of the week. so definitely get your tickets if you want. There's a there's a code on the latent space for lep space subscribers if you can dig in through your emails and it is on ai.engineernyc. >> And if you're not a latent space subscriber, what are you doing folks? Come on. It should already be in your inbox. That one's on you. Swix, thanks for joining us. Always a pleasure. Great to get your take on things. few people are more plugged in to what's going on in the Silicon Valley development world and it's always a pleasure to get your perspective. Good luck with the event.

122:01 >> Thank you. >> Byebye. >> Thanks, man. >> All right. yeah. I'm I'm I pity our our claude that's going to have to try to cut this down into a highlights episode. Man, I think this is banger content from start to finish today. >> We we we learned we learned quite a bit actually. so really a difference between the 100 to 400 person you know maximum ad additional you know people that you can take on up to a series you know seed to series D and then kind of like this mid-tier kind of gap between the kind of 2,000 person company and the 10,000 person company kind of where you're kind of struggling to manage and grow fast enough and really trying to figure out your management agement techniques because you're in this kind of gray zone where you can't just assign the team the project and expect to be completed or fire them which you can do at a larger firm and you can't at the same time you don't have the bandwidth to individually kind of you know interact with every team and see what they're doing and see the internals. So kind of kind of really a management gap there and my guess is that management gap gets filled by AI agents. So you you jump from two 2K to 10K people rather than you know scaling slowly from 2 to 10k. So I don't know have we seen any companies that have been super successful lately grow headcount in the way that you would expect them you would have expected them to if they were blown up years ago.

123:50 I I would >> I mean I might need to put a research agent on this, but it seems like everybody is sort of keeping the squad pretty small. like how many people did Cursor have, for example, at acquisition? I don't know, but it wasn't like >> I think it was less than a thousand people, I think. So >> yeah, that would be my impression, too. I mean, let's we'll do a quick >> between several hundred and a thousand, I think. Maybe 300 to 500 something.

124:18 yeah, I think I think the I think maybe the if you just look at revenue per employee and I think the re the requirements for revenue per employee are like very high now. They're like in the 5 to10 million per employee range and I think that's the scale at which openi is very big right now. they have they have they definitely have thousands of people and maybe close to 10K and maybe more but it's you know the revenue per employee is like very high right and the valuation per employee is also very very high anthropic is probably about half the size or 60% of the size of open AI and also revenue per employee is very very high also to note both firms tend to hire startup founders so they tend to hire startup founders and CTOs and and then have them have them be there to you know kind of self-motivate.

125:18 This is also a strategy used by Ripling. So Ripling is a kind of HR and back office software company. They had a strategy of hiring former founders and then assigning entire basically entire business lines because they they were absorbing they had absorbed like you know business line by business line of you know other HR companies and so they would assign these former founders to kind of lead teams to kind of build out these business lines that they had and I think that's pretty similar to what open anthropic have done I think they they've they've hired you people who have founded companies before in the space and then they tell them to they they hired people I think the CTO several CTOs of SAS companies were hired to you know run run run run run products which would eventually be competitors to their SAS their SAS firms. So I think that is that is definitely a strategy that people are using.

126:16 So Gemini meanwhile estimates indeed pretty much exactly what you said three to 700 employees it estimated for Curser U and it estimated about $10 million in revenue per employee at Curser which for comparison Meta and Google you know have always been like extreme standouts in that regard and they operate around like one to one and a half million in revenue per employee. So like almost a full order of magnitude higher revenue per employee at a cursor. Gemini also puts OpenAI and anthropic in certainly elite territory but a little lower than cursor more like 3 to 5 million in revenue per employee.

127:08 you know obviously a big part of the reason it has to be so high is because they're spending all not all the money but a lot of the money on tokens and in some cases like literally companies are still even losing money on the tokens that they're serving to customers. I know cursor probably at times has done that. I'm sure at this point they're you know, positive unit economics on on most users, but when you're selling tokens as part of your subscription, you know, obviously a huge huge part of your topline goes to that.

127:41 One thing, if we had had another minute with Swix, I would have been very interested to get his take on how people are reacting to the login with Open AI. So you can bring your tokens along with you to your SAS app. I didn't that that it was not the kind of event that I was at this weekend that you know where talk of that kind of thing was happening but I suspect that is going to be you know a huge win. I would be be surprised if it's not because that is just such a weird environment to be in to be like I have to make millions of dollars per employee because I'm turning around and se and you know spending in some cases a majority or even you know going into the red just to power the product with the tokens you know so much nicer if people can bring their tokens along with them at at least at first. so that may change what is required I I imagine a bit. but yeah these are you know heady heady times to say the least.

128:43 >> indeed let's maybe segue a little bit to other other news. so I think over over the weekend we've had a a series of math discoveries. I I I have still not caught up yet. It is an unending series of math discoveries at this point. major conjectures falling left and right. I think one which perhaps is worth maybe discussing is threesome and obviously people are making a bunch of jokes about this but anyway and I think the what's happening is that people and and this is this is an interesting story. So let let let's do let's do what what this is. If you take three numbers and you sum them up, can you get to a zero or not? So that's that's all that this is, right? And this is basically a search operation at the end of the day. And so 2 - 10 and plus 8 equals 0 is a threesome solution. So and it's basically ends up being a search operation over over the space of numbers. And this whole thing was that can you get this below n squ? So can you get the can you get the number of operations that you do below the n squed number? And the for 20 30 years now it's been no. The answer has been no.

130:17 and so you've had all of these people have been trying to prove it. and they haven't they haven't been able to get something below n². And these guys that published yesterday or today they managed to get it to n= 1.99992 by finding a combination of numbers in the number space that you know you could get the particular like search time to under n squ.

130:48 and this is pretty significant because it breaks the barrier and once the barrier has been broken other people will optimize find other optimizations to to further reduce the number and for the moment no real impact because these are extraordinary large numbers that these guys found. So it's not not a big deal. but now that you have techniques that can get you below n^ squ you'll find other techniques and in the end this will you know help matrix multiplication and you know other other u other basic very basic mathematical operations because as you can imagine summing three numbers is a very very basic mathematical operation that you could you could do. and the way that it was found was an anthropic employee basically decided to was gave a cryptography problem to you know Claude and asked Claude to solve it. And that cryptography problem relied on this n being true that the that the limit of how quickly you could do something was n^2. And Claude while doing this cryptogari problem which relied on this n squed being true broke broke it and came back and said well you asked me to you know solve the cryptoari problem and I kind of just solved it by undermining the entire basis for that set of problems.

132:19 and so on the one hand it is yet another kind of find a category find a single value that disproves the entire kind of theorem. So but on the other hand it's also you know absolutely amazing that Claude could actually do this at all and actually did it as an aside really. So a number of things will happen over the next I think few weeks and few months. this is somewhat related to P= NP. so threes sum was basically an answer was a P problem. It was answers which are quick to find. and NP is answers which are quick to check. and so far you know right now we are assuming that these two regions are not connected.

133:15 but this is starting to go into that territory of okay which of these problems are really you know unsolvable or you know faster than you know what you could solve at and which of the problems are you know essentially solvable. The other the other interesting thing about this problem is that after Claude solved it, Anthropic approached two mathematicians who were leaders in this field and asked them to publish paid them to publish, gave them the solution and then they kind of reformulated in a readable manner and then they published under their name. So they are listed as the first and second author. Claude is listed and the fact that Enthropic gave them the solution is listed in page 63.

134:08 And this is being claimed as the way that things are going to be in the future that we we we would have models solve these things and we'd hand the hand the solutions off to mathematicians to kind of verify and the mathematicians who verify or explain it would end up taking the credit. It's still still there's still a lot of unease in the in the in in the math community right now on what is actually the right thing to do. This this was the recommended process from the from the group from the math committee that was put together you know a few weeks ago. but it's notable that you know we come after the weekend and on Monday we have I I think there's like 20 or 30 30 you know math things that happened you know on Monday just dropped. So people basically spending the weekend using up their tokens and on Monday you know they they dropped these proofs which of things which have been standing for 20 or 30 years. So there you go.

135:17 Yeah, never dull moment. This is one of those things where it's like, I'd never heard of this problem. I don't know anything about this problem really. I'm not still entirely clear on how meaningful it is. The guidance I got from AIS was like it's not likely to be immediately super consequential, but that barrier busting effect can at times be super important. You know, the classic four-minute mile, right? Once it's demonstrated that somebody can do it, all of a sudden lots of people start doing it. So, it will be really interesting to see if all of a sudden there's like a wave of kind of follow-on optimizations that bring the the number down substantially further. This also seems to be going on right now in the nano GPT speedrun department where there was one of the and my g my understanding of the discourse around this also is it's like a little bit unclear like what should count because people are starting to take approaches that were sort of I guess allowed by the rules but like not necessarily what people expected you to be doing when you were doing this sort of thing. So, I need to study it all more. But just headlinewise, it seems like we've gone from something like 70 seconds on this this classic benchmark. This is something people have been working on for quite a while. And basically the the idea is to train a small model to get to a certain loss as fast as you can in terms of wall clock time. And this result here took almost half the time off.

136:58 >> Yeah. >> which was more than I think the last 45 improvements combined. And this one has been accepted. And then since then a number of other people have come on with additional claims. One was just like a few seconds off and there was one that you know was maybe like half of the time off again. And then there's one that's even under 10 seconds now. but that is still I guess under review and not not fully accepted. And I guess it it again uses some techniques that are kind of like maybe going to get rules lawyered or you know maybe maybe the maybe the techniques themselves sort of constitute like hacking the rules to a certain degree. One thing I did think was quite interesting about this particular one that has been accepted and kind of seems to have kicked off this wave of you know significant drops in in the time was that the author here said AI didn't play a big part in it.

137:56 like the person you know senior executive leader at the frontier company that I mentioned earlier their perspective was they believe that the model probably can come up with these quality of ideas but it's an elicitation problem and they didn't really rely on AI super heavily in this case so you know obviously you know getting lots of help. but in terms of core insight it was mostly attributed to the human not so much to the AI on this particular drop. This company is doing also like largecale training. So they did also make a point in their publication about this that the optimizer that accounted for a significant but I think not majority of the time savings that they did share is actually not as good as the one that they are using internally. so they have an even better optimizer improvement in house but they're keeping that to themselves because >> and this company is hyper hyperstition as a hyperstition and so this is I think I think the nano GPT's record was a form of you know performance marketing basically to intended to show the capabilities of their optimizer and their and their team in general. I I I'm surprised that they're still independent. And I'm surprised that they haven't been bought yet because given given the severe lack of talent, you know, a shortage of talent. I'm surprised that they haven't received offers or, you know, their staff haven't been poached as of yet. So, >> well, there could be a licensing deal in their future. Who knows?

139:47 yeah, man. Wild times. This has been busy weekend. intense information fire hose that I'm still kind of processing. I took thousands of words of notes at the various sessions and even just you know after talking to people sometimes informally I was like I'm going to write down what I heard here to make sure my feeble human brain keeps track of it. Here's a question for you. This this question actually I heard from Aja Katra in May. This goes back to recursive and I was I had occasion to reflect on how my answer might have changed this weekend. So in May she asked me and others in an audience how many copies of yourself would you need to have to do as much stuff as you are doing today with AI assistance and maybe at some point we'll have to like get into the sort of coordination problems even of having like a lot of clones of oneself. I wasn't really thinking of it in terms of you know stepping on one another's toes or needing to coordinate but just you know another way to think about it would maybe be how much faster would you have to run yourself you know if you if you could speed up your own inference how much would you have to speed yourself up to get as much output as you are getting with AI enhancement at that time I said 2x >> two clones or two or twice the speed.

141:28 >> So meaning meaning that let's say you don't have AI assistance, how many copies of yourself >> would you need in order to perform the same amount of work, >> right, that you are doing with AI assistance. Yeah. So I said 2x then and this weekend thinking about it again I was like I think it might be more like four or five now. >> Mhm. But then last night I came home and I was you know just asking for all kinds of enhancements to my my personal you know as Swick says like you know total made to measure agent console that I use and in that domain in software in particular it's like way higher u so you know what I did last night with like 12 prompts would be easily weeks worth of work. So that ratio is like honestly hard to calculate and kind of throws the throws a bit of a wrench into the whole thought experiment. of course the other answer would be like I wouldn't do that stuff but you know that's sort of fighting the hypothetical. but even subtracting out the software development and just imagining sort of you know the content the sense making writing you know whatever giving talks I still think I'm probably right now in the kind of four to five range. And you know, that's a finger to the wind.

142:56 It's I haven't measured it. I don't intend to I don't intend to deny myself AI for weeks to to really try to measure this sort of thing. But what's your intuition for you? >> I don't know, like maybe 10. So, so, so the the way I the way I look at it is that a lot of what we build for ourselves tends to be a couple a couple of different things that we tend to build. We tend to build like sensory organs. So, we want some fire hose of data and we want there's a lot of it and we want it to be compressed into things that we want to that that are important for us to see. and you can find this kind of like pattern in many places. For example, Swix created AI news. in AI news. He set up pipelines of you know unstructured data from discord channels on AI from archive from all of these like different places and they kind of got filtered and you know ranked etc etc until he came to be able to send out these emails which were you know very compressed kind of you know sensing or and those became kind of sensory organs for the entire community to kind of know what's going on in the community and I see this pattern repeat. I've seen so many repeats of this pattern you know for people building this kind of thing which is which I call you know sensory organs basically and I also see translation layers people build a lot of translation layers like translating between like a math thing to language or etc etc. So they they build that and once those things are built when you use them they basically displace a portion of what you would have done. and that and that is a permanent displacement.

144:47 For example, like if you if you had been, you know, if you were the kind of person to wake up in the morning and look through like Y cominator, hackernews and like four or five or six different pages to kind of keep up on AI and that was that used to take like you know 45 minutes of your time in the morning and all of a sudden that's compressed by AI news to like you know 5 minutes that's a permanent compression.

145:12 So when you say like let's take away the AI agents, you also have to say like let's take away all of these affordances that we've built for ourselves over the last year or so. And once you start taking away those affordances then you need to replace that with like an actual person kind of and I think that that is where you start like okay I have permanently displaced a lot of work that I used to do and in order to live without AI completely you would need you know quite a lot of people because we've we've replaced you know lots lots and lots of what we do right and permanent replacement too. it's not not not temporary permanent and getting getting more and more thorough and more and more more and more stuff is getting consumed by by the agents right and I think that's that's the main thing the main thing is that we have built these tools for ourselves and that's the that's the that's AI that's AI and software they they allow you to build tools for yourself and so we've built tools for ourselves you know Swix is building tools for himself continuously right continuously building things that you know will will change his organization and will alter how the entire organization does things going forward.

146:35 So I think the numbers are actually much higher than we think. It's not just the number of tokens that you spend on a daily basis but the fact that you've spent those tokens and that becomes kind of a capex and you have this capital asset that you've built for yourself and I think that's pretty significant. buckle up. You know, the takeaway, one other big takeaway maybe we can end on from the weekend at the curve in terms of expectations and this was definitely shared again across your frontier companies and again shows Swix's foresight and you know to a degree our own as well. AI for science is very much where they think we are headed in the notistant future.

147:28 like one one sentiment was you know are AI going to be superhuman at just some things like the verifiable tasks or are they going to be superhuman at everything? and the sort of middle position which I think is very credible is we are seeing superhuman performance at anything we care about and are you know really committed to investing in. so that doesn't mean it, you know, it's full generalization to every domain, but even in domains that are thought of as like not super inherently verifiable, they feel like when they put their focus on it and they, you know, go license what data they may need to license and they, you know, of course, apply lots of processing to that data to augment and, you know, create synthetic versions and what have you. and you know put some RL on it too. the whole the whole package basically is I think broadly understood to work on essentially any problem that they really choose to focus on. And so AI for science is coming next and the expectation is we will see one person went as far as to say in no more than a year it will not be possible to do the very best science without AI's playing a big role in it and you know we'll we'll see maybe somebody will prove that they can still do it you know Tom from earlier is still designing chips better than the agents can do it but he certainly couldn't do it without you know certainly not fully at anything like the speed that he's doing and the expectation is that comes to much of science over the next year.

149:25 >> I I I wonder how you know math is very nice. Ma math is like one theorem depends on another. So if you disprove an underlying theorem, you disprove everything on top. It's it's very nice that way, right? I wonder to what extent science like physics or biology or chemistry is dependent on these kind of like node like nodes on the graph and if you disprove one of the nodes like an entire theory falls or entire you know something else falls and then you have new new knowledge that can be created and I wonder to what extent you can do that kind of thing and kind of focus on the more important nodes which will allow you to open up like entire new spaces rather than just doing because because if not I think you end up in this kind of experimental like this cuz you know let's let's be honest maybe 30% to 50% of all science papers are probably you know non-replicable and are probably trash and that's what we're training the AIS on. you know, >> although they I bet you they are, if not yet, I bet you they will soon. I bet they already are doing the same thing with papers that they're doing with their own RL environments. I would bet you they are going through and poking, you know, and they're socially, you know, these folks are not always the most socially sensitive, but they're socially sensitive enough to know that they don't necessarily find it in their interest to come out and call out, you know, however what percentage of the scientific community for having published junk stuff. But I would bet you they have a growing sense inside the companies of what scientific literature is actually reliable and what is not. And that will of course continue to get refined as the models get better. but I I would be very surprised if that's not well underway at this point as well. Can you can you imagine the can you imagine the outrage the absolute outrage when the models come out and say like oh yeah you know this Nobel Prize winning theorist was wrong like you know and and the guy is still alive and well he's like what can can you know can you imagine the outrage well it's another one of these Jubilee questions you know that we're we're going to have coming our way sooner or later there's no doubt that you're right that there's a lot of junk science out there And if these AIs are going to get that good at science, which I think seems very likely, there's no g no doubt that they're going to be able to be able to separate the good from the bad with presumably imperfect reliability. But like, you know, it's you know, we know panggram isn't perfect either. We've demonstrated that a little bit with one of our past Gonzo experiments. Looking back at my own work, I was able to find like a couple things where it called me AI writing.

152:14 And I I you know feel confident that that's not a fair description. even though I'm not afraid of that you know label I'm not afraid of being panggrammed you know just in point of fact like there were a couple false positives. that I'm sure will happen in the science realm too but I would bet it will be accurate enough that people will just kind of generally believe it at you know take it at face value when the AI says that something is like is junk. and some people will be you know defending their position and you know probably it's going to be tough for them to defend that position and I think that you know at a minimum we are going to need to be graceful about that process because who knows you know why all these different junk things happen mostly it's not fraud I think you know it's it's all kinds of ways people can go wrong in the in the scientific process that's not fraud so I don't think we should take the mindset that it is by default but there will be a lot of stuff that it's going to you know watch out for mega retraction would be my expectation over the next can't be that much more than a year. I don't you know I don't I have a hard time fitting into my worldview where it wouldn't be possible in the next year.

153:31 so the president of Stanford was a Alzheimer's researcher and some kids at the u college newspaper took his old research research papers went through them. he had some diagrams, some pictures there and they kind of spotted that these were u fake basically. And then that drove this entire process of him kind of like trying to explain why he did it and it's like someone else's fault and not his fault etc etc and then he had to quit so he resigned. Mark Tessier Lavine MTL I think actually the early so so that's just image recognition that's that's not that's not even like LMS or reasoning or whatever that's just the image recognition you know u you know post Alex alexnet postresnet that's just a simple image recognition being applied you know to these to these diagrams people just have just not done done it at you know at scale yet right or the resolution at which they've been scanned in as PDFs is not high enough to have confidence. But if you had the high resolution images you know the originals which is why also there's this whole like push back some some of the some of the journals some of the sub submissions they're asking for raw data and there's this push back on submitting raw data because they don't want the raw data to be reanalyzed by someone else and then for the errors to be pointed out and then for the these outcomes to be basically you know disproven.

155:09 it's it's this whole mess and like just just as we saw in ma mathematics this whole thing about like you know the status and like who should be allowed to do it and etc etc like I can just imagine some like young undergrad who runs all of biologyy's papers at scale you know several you know spends like 100 grand runs all biology papers at scale and then let's say says like okay 30% all these guys are just faking right so I think there there is going to be an enormous kind of social push back on in in the research community on this this kind of thing because it's going to affect a lot of careers and livelihoods I think. So >> yeah, it's going to be fraught. No doubt about that. One other thing that I heard at the curve this weekend was on the topic of medicine in particular you know and my experience using AIs for medical situation is well documented going back a year. Models are of course much better today. but this is sort of a not not confidential thing but just somebody who works on the AI for medical advice domain said doctors have been surprisingly warm and receptive to AI. much less threatened by it than you might have thought. Much more inclined to embrace it. You know, of course, from my position, it sometimes felt slow in the hospital, but even on the level of like the AMA, which I had long assumed would come in and try to shut things down that, you know, we're starting to encroach on its turf. the point of view that I got from somebody who, you know, has been doing some of that work with with those kinds of organizations was like not really. You know, I I I came away from that conversation feeling like good on the doctors and good on the AMA if they really are as patient oriented as it sounds like they might in fact be. Of course, you know, medical field contains multitudes like everything else. But at a high level, I had expected walls to be put up kind of unreasonably in a protectionist way. And it sounds like, you know, from somebody in the no, much less of that has happened than we might have feared and that you know, the doctors are actually kind of living up, broadly speaking, to their oath of, you know, putting the the patients well-being first. I thought that was really exciting and I would hope to see the same thing from scientists. U, but, you know, we'll we'll see. It's it goes a little bit more it's a little more personal for the scientists, right?

157:51 It's one thing when the AIs can come in and just be super useful in general in medicine. And now it's like, okay, this is a whole new reality, but they didn't attack you specifically as as a doctor, you know, who is who did wrong. The mega retraction watch is going to be something that's going to touch, you know, a lot of individual names and that is u inherently going to be just a lot more contentious. But I did come away optimistic about not just of course the AI for medicine which I've you know long been optimistic about but actually also optimistic that like the powers that be will not fight it and might even embrace it and we might actually all get to use it you know in with their blessing and with their kind of support in a way that I had kind of always dreamed of but you know was was a little bit stealing myself against the possibility that it it might prove harder. sounds like it actually might be smoother sailing in that regard than I had dared to dream that it might be.

158:47 >> I I wondered to what extent it's because the expectations that people have for doctors to be current on all of this new medical technology is so high that they've been feeling like just overwhelmed, you know, for the last few years. you know post Google basically and just not able to keep up with the questions that are being asked by by patients who go in Google and and now they have something that that works for them, right? That that that works to keep them up to speed very quickly and accurately too. the the one of the problems with Google was the accuracy was the accuracy was very debatable and things were hard to search because you had to go through all of these links and then you it didn't connect to what you already knew and I think open evidence kind of solved a lot of that. So I think I think that's been pretty significant for for doctors.

159:47 >> Yeah, no doubt. Well, I think I need to go prompt some AIS to make sure I'm still hitting this five X acc acceleration. this was a great session today. I really enjoyed the two conversations and a lot of interesting conversation just between the two of us as well. So, we'll be back at it tomorrow. We're only doing two days this week, but we'll have another good one tomorrow. And we'll just keep rolling. Thanks for being with us on AI in the AM.

160:19 >> Thank you.

Summary

The discussion revolves around recent developments in AI, particularly in the context of a conference called "The Curve," where experts gathered to share contrasting views on AI's future. Key takeaways include the growing consensus that AI will become increasingly powerful and influential across various domains, including science and medicine, and the potential for AI to reshape traditional fields through enhanced capabilities and efficiency.

- The Curve conference emphasized the need for diverse perspectives in AI discussions, highlighting ideological divides within the community.
- Attendees acknowledged that AI's capabilities are rapidly advancing, with many expressing a sense of urgency to address the implications of powerful AI systems.
- There is a growing belief that AI will play a crucial role in scientific advancements, with expectations that it will soon be integral to high-level research.
- The conversation touched on the potential for AI to disrupt established scientific norms, including the verification of research and the reliability of published papers.
- A notable example discussed was the recent breakthrough in solving the "threesome" problem in mathematics, showcasing AI's potential to contribute to significant discoveries.
- The medical community appears more receptive to AI integration than anticipated, with doctors embracing AI tools to enhance patient care.
- The importance of building tools for efficiency and customization in software development was highlighted, suggesting a shift in how work is approached in tech.
- The conversation concluded with reflections on the evolving role of AI in various sectors and the implications for future research and development.

Questions Answered

What were the key takeaways from The Curve conference?

The Curve conference aimed to bring together contrasting opinions in the AI space, highlighting the need for broader perspectives amidst rapid advancements. Attendees noted that while many companies are making progress, there is a lack of time to step back and assess the overall landscape.

How do major players like Elon Musk and Mark Zuckerberg fit into the current AI landscape?

While traditionally seen as leaders, Musk and Zuckerberg are now viewed as distillers of AI advancements rather than primary innovators. The focus has shifted to a few key companies pushing the frontier of AI technology.

Who is Thomas Summers and what is Posetron AI's mission?

Thomas Summers is the co-founder of Posetron AI, which focuses on developing hardware for AI inference, aiming to enhance memory systems for efficiency and cost-effectiveness. His background includes significant roles in AI hardware development.

What innovations are being made in memory technology for AI?

Posetron AI's Osmo chip offers significantly higher memory capacity compared to current benchmarks, allowing for more efficient scaling and reducing the complexity of multi-device setups.

How are AI advancements impacting chip design verification?

Recent AI models have improved the speed and efficiency of chip design verification processes, allowing for closed-loop iterations and automated documentation handling, which were previously time-consuming tasks.

What can we expect from the upcoming AI Engineer conference?

The AI Engineer conference will focus on the latest trends in AI applications across various sectors, including finance and technology, and aims to gather insights from industry leaders.

How is software engineering evolving with the rise of AI?

There is a shift from meticulous line-by-line code review to a focus on module coherence, allowing for faster development despite potential oversights in individual components.

What is the outlook for software demand in the context of AI?

Despite potential market corrections, the demand for custom software remains high as businesses seek to automate and enhance their operations, turning traditional processes into software solutions.

How is AI affecting business operations and employment?

AI is leading to significant changes in how businesses operate, with potential layoffs in tech sectors as companies adapt to new efficiencies. However, the demand for skilled workers in AI and software development continues to grow.

What are the long-term effects of AI on work and productivity?

AI has permanently altered workflows, compressing tasks that once took significant time into much shorter periods. This shift means that removing AI would require substantial human resources to replace the efficiencies gained.

© transcribe · For agents Built with care and craft by Gokul Rajaram