You're watching TBPN. Today is Friday, September 11th, 2026. Wall-to-wall coverage in the Wall Street Journal. Remembering 25 years ago, the cover of the Wall Street Journal. Every single article except for one was about, of course, the World Trade Center. The headline that day was 'Terrorists destroy World Trade Center, hit Pentagon in raid with hijack jets.' Bin Laden is on here. The only story that, you know, attacks raise fear of a recession. Very interesting time capsule. Highly recommend picking up a copy of the journal today and taking a trip down, you know, in memory of the tragedy. The one piece of news that broke through this day that was not related to the terrorist attack was Xerox reached an equipment financing agreement with GE Capital that will let Xerox erase about five billion of debt. Very, very odd. Every other story, I mean, the market was closed. Actually, the internet got a very interesting shout out here. Says telecom systems were strained as terrorist attacks in New York and Washington knocked out telephone wireless services across the Northeast. The internet proved the most reliable way to communicate following the attacks. As the phone system sagged from severed lines and an extraordinary volume of calls, corporate executives used email to find employees across town or across the country. So interesting to see, but there's so much to go into and lots of interesting retrospectives across all the different media organizations. I don't know that I have a particular anything to dive into there, but it's interesting for you to go and dig into. Anyway, moving on. Let me tell you about ramp.com. Time is money. Save both. Easy to use corporate cards, bill pay and accounting, and a whole lot more all in one place now with musicals. Anyway, um, what are the AI doomers actually proposing? That's the question I was trying to answer this morning for a while.
That's one that. So yeah, there will be this weird translation layer between the thought leaders and the people that are writing policy papers effectively, and then what actually gets implemented, what gets, you know, votes basically, can be wildly different because to get the populace to actually support something, you need to wrap it in a different structure potentially. AI 2027 predicted that this month, you know, late in 2026, Congress would wake up, and that is in the journal: Congress is suddenly waking up to the AI doomsday threat. And so, this is happening all over the place. Was it Matt Damon who was caught on TMZ being interrogated about his thoughts on AI risk? There's now protests. I believe AI 2027 predicts a 10,000-person anti-AI protest by the end of the year. I was trying to figure out how big this protest in Chapel Hill, North Carolina earlier this month was. I think it came in sub-10k, but certainly tracking to it. It was about 150 people at this protest in Chapel Hill, but just two orders of magnitude away from the prediction from AI 2027 and AI 2027 in the sequel AI 2040. Frustratingly vague about impacts outside of the AI industry. So there's a lot of really, really amazing predictions about agentic capabilities and the amount of compute that will be marshaled and even lab revenue, but there's not as much predictive stats and calls around like what will it do to GDP? What will it do to employment? What will, you know, how often will people actually be using this? What will it actually be good at? The diffusion question is still sort of left unanswered. But they're clearly taking it very seriously. There's a huge press cycle around this and the journal says that Congress is suddenly waking up. So, I wanted to dig into like what the actual proposal is because the joke for a while has been just like everyone when they're pressed on this, they just say we got to talk about it. We got to talk about this. AI 2040. Daniel, uh...
One thing that stands out, it's interesting to me that there's been so much more seemingly grassroots mobilization around the anti-flock movement, Dlock, when you compare that and you don't see totally you don't see videos of people with, you know, 100,000 likes saying like, 'Here's how to cut power to your local data center.' Not for X-risk reasons, but you will see a post that's just like, 'I don't like AI image generation because I'm an artist,' and that will get a lot. So there's a lot of anti, but what I'm saying is like the sentiment, the anti-flock sentiment converted into actual physical actions by a bunch of just otherwise normal people.
Yeah. I'm saying we haven't seen that yet. With data centers, maybe we don't. Yeah. But it's notable. Yeah. Much harder. Data centers are in remote locations, highly fortified. There's fences. Totally. Just walk in. Block cameras are in your, you know, on your street maybe. Exactly. Yeah. So, yeah, there's degrees there. But AI in the what's weird is that, yeah, I mean, you're calling out grassroots taking down the data centers. The AI 2040 proposal is effectively the opposite. It's like harden the data centers even more. And the actual proposal is super concrete in AI 2040. And it's very interesting just to hear about how they want to slow things down. So the thing that I think a lot of people online who are like, 'Yay, an anti-AI sentiment that's going viral' are going to be depressed about is that like this is not stop AI at all. AI 2040 is like keep the inference flowing. The current models are great. Also let's keep doing capabilities research, but we're just we just want to reach superintelligence by 2040 instead of 2028 where we're not necessarily prepared. So they want it to be highly controlled by governments, nations, states, highly secured and there's a whole bunch of very, very tactical recommendations that they make around that. So the first mechanism is an AI pause. They want to pause training. They don't want to do any more new frontier training runs or R&D experiments. And to enforce this, they're calling for to apply inference-only verification to essentially all major AI data centers. So anyone who has more than 10,000 H100 equivalents, roughly a hundred million dollars of equipment, and that is like pretty easy to figure out. You're just like, big building over there. Let's send the inspector inside. Oh, says Nvidia on all these chips. Count them up. There's over 10,000 of them. You got to apply for this permit. You got to tell us what you're doing, right? Very easy to enforce, at least in the United States. And with an international body, you could kind of do the same thing internationally. So, the whole goal, you can only inference the current models and you got to verify your workloads with an independent auditor. Probably the government, maybe there's some sort of non-governmental organization that's doing this. You know, there's a whole bunch of different solutions that you can pull from across nuclear and non-proliferation work that's happened in the past. Major countries, they want them to declare AI compute inventories. Tell everyone, not just your local population, but also the international community, how many warheads you got? How many H100 equivalents do you have? Where are they? Everyone shares this. That's going to be a really tough sell because international agreements are really, really tough sells. It's much easier to have a groundswell of support for something that happens in America. America changes. We have a system that we don't really have the international rules to quickly implement that in a way that doesn't allow for a lot of defection, but they want to know who has compute. So major data center owners and semiconductor supply chain companies would be required to turn over sales records. Who'd you sell those chips to? Where'd they go after that? Foreign inspectors would do routine chip counts physically on site at facilities. Large transfers of chips would only be allowed to go to registered auditable counterparties. There's some interesting networking-specific components to this proposal too. They want to physically remove high-bandwidth east-west networking inside of data centers. So you can't do large distributed training runs, but you can still do inference. So again, anyone, all the anti-AI people who are like, yeah, I don't want LLMs around anymore, like these are not your guys. They're not fighting for that. They are fighting for stopping the next training run, which is probably a line. Those are overlapping circles. But it is not moving backwards in time. It is merely slowing down at this current moment. They also want to install passive optical network taps on anything leaving the data center to independently verify traffic. For any new AI R&D data centers, they want entirely new facilities built from scratch with nation-state-level physical security and verification. So they're saying like, okay, the data centers that are built, they can inference the current models, your Astras, your Fables, like your Groks, like those can run because we can deal with those, we can harness those, we have control over them, we're going to continue to align them and that's a solvable problem. But for the next run and the one after that and the one after that as it gets crazier, we want it in a new building built inside a Faraday cage so you can't communicate it from the side. We want a bunch of physical controls. It's like going to a nuclear facility. Highly verified who gets in the building and when. Air-gapped communications. This is one interesting proposal that they have that really shows how deep they thought this through. They want the R&D data center to be connected externally. If you want to communicate with it and you want to tell it what to do, okay, train the next run or do whatever, they will have a bandwidth-capped connection at one meg per second. So you can send little instructions, but if you say, 'Send me the weights because I'm taking them somewhere else,' it would take you like five years to exfiltrate it. So interesting like hardware solution to this. I, you know, how do you actually go and implement that? There's going to be a whole bunch of other things, but interesting that they're thinking about like the width of the pipe. So it'd be very obvious if you're stealing the model weights because it's like, wait, this one-meg pipe has been at full tilt for months. What's going on here? Someone's taking the stuff out of the data center. They want if when frontier model weights move from an R&D facility to an inference facility, they want it to be placed on physical storage devices encrypted independently by both the US and China. So both countries have to sign off and physically escorted by representatives of both countries to the destination.
It's a tall order. That one's a tall order for sure.
Yeah. And they actually want frontier models to be made deliberately larger than compute-optimal. So, they want the weights to be 100 terabytes instead of honing them down to something that's just one terabyte that could actually be moved around a little bit easier. There's a bunch of public disclosure proposals in there, restrictions on various tradeoffs. So labs would have to share the model specification, the fraction of compute devoted to internal AI use. So you don't get a lab that's just internally using way better models than what's available externally. They want qualitative descriptions on how powerful models are being used internally. Restrictions on how big the gap can be between the best internally deployed model and customer-facing products. This has been a common discussion point with like the rollout of Mythos and Fable and Astra and Astra Next and all these different models where people have said like, ah, it's really unfair that like this lab gets a better thing than I do. We should be on even footing if we're both going to be competing in web design or we're both going to be competing in legal, like why can't I buy this product from you? And so there's some overlap there with like the broader business community, which I thought was interesting.
Yeah, I mean, first of all, this, I mean, a lot of this tracks with the kind of regulation that, you know, we had pushed for, you know, beginning about two years ago around podcasts. Yes. You know, wanting podcast studios to be air-gapped. Yes. Wanting, you know, Faraday cages around podcast studios.
A locked briefcase with an SM7B in it. And in order to unlock it, Patrick O'Shaughnessy and David Senra both need to give you codes to independently verify that this podcast is worthy of being recorded.
Yeah, I like that one. Yeah.
It just makes sense. So, high level, Valve they see as being like the most effective in controlling the speed of capability improvements is compute caps. So there is a world and we're going to the Bernie Sanders thing because it's already getting like sort of twisted, but the big hammer is just chip controls and data center buildout slowdown. That's the easiest thing and I think that's like the biggest valve that we're going to see twisted around to actually slow down capabilities. And so the goal is to allow models to get better mainly by adding hardware rather than inventing better algorithms which can leak to secret projects. So the goal is like, okay, well we know that this model is capable of this, so we want this much compute over here. Okay, you've done well, we're allocating more compute as opposed to this one weird trick that AI doomers hate. Yeah. So the goal here is so interesting because when you look at, I mean, anytime you have really, really hardcore government regulation and international coordination around issues like this, you're going to have a bunch of unintended consequences. Yeah. And one thing that feels obvious around this, if these policies were to be rolled out, is that you would effectively create an incentive for millions of individuals or groups globally to be in secret like trying to find entirely new breakthroughs that are... Yeah. And again, this incentive already exists, but it kind of pushes a lot of the idea that humans are just going to be like, 'Oh, I'm no longer going to try to create the god model because like there's this, you know, big global organization that's sort of policing it.' Yeah. I just... I mean, that's the same thing we see with nuclear non-proliferation. There's always a discussion about what countries are getting the bomb and when and how far along are they and wars break out over this and yeah, like the game's not over just because you create a framework. But there is at least a, I mean, we've avoided World War III, so you could say that a lot of, like the vast majority of nuclear non-proliferation work has been successful even though there's been a ton of examples of people trying to divert around it. In fact, it's like been the backbone of geopolitics for like 60 years has been like who gets the bomb and what chips are on the table and stuff.
Ultimately, GPUs and computers are much more wide, you know, infinitely more widespread than nuclear materials.
Yeah. But you still got to marshal them all together. Yes, there's some weird scenario where there's a Python script that's AGI that can run on your laptop, but I think most people are convinced, at least in this crowd, that scale... that scale is a prerequisite. And I mean, we were joking about like how it would be so like SSI's new neolab recently got a big cluster from Nvidia and we were like the most bullish thing you could do if you're this secretive neolab would be like, 'We're actually selling our compute because we've discovered a more compute-optimal way to reach AGI and we don't need a lot of compute.' But of course even Ilya is like, 'It's time to scale up, I need more compute,' because it seems like even if he's taking a completely orthogonal approach to innovation and research, he still needs a lot of compute. And so it does feel like everyone is sort of with the consensus that it's going to be a big building with a lot of energy, big heat signature, definitely visible from space, and pretty simple to track at least in the short term until people start building crazy underground facilities and then you're back to, you know, nuclear non-proliferation. But their goal is at least like, you know, try, you know, try.
Yeah, and then the other side of this is does AI development actually become something closer to the Manhattan Project where, you know, a lot of people, 100 researchers are working with the government in secret because you can't just assume that other countries are going to slow down or do any of these things.
Yeah. And but in general, I think the proposal is don't go back in time. It's definitely not stop everything in its tracks. It's a slowdown with the goal of scaling gradually. They do actually want to reach superintelligence. They just want to do it by 2040, hence the name of the project. So the goal is gradually scale into top human expert capability around 2035. Now a lot of people are saying, oh, we might get this by 2029, 2028, 2027, and they see that as too fast. So they want to push that out to 2035, then wait five years with AGI, and then unlock superintelligence in 2040. This is their initial proposal. Of course there's a lot that could change over the next decade. And I think like to zoom out overall, if you're worried about X-risk, the AI 2040 plan does feel like a concrete path towards slowing down. The conversation definitely gets dragged down into p-doom estimates and trying to narrow down exactly how a human extinction scenario plays out. And that can be, I feel like that's almost a sideshow because in a democratic society amongst humanity, like it doesn't really matter the mechanics of getting to 10% p-doom or any of those. It's just like if everyone feels that way, something will happen. This is a concrete plan of what that might look like and that's valuable to understand in this case. Yeah. So for the safety skeptics, it's easy to see how this level of control over what you can do with computers is authoritarian or anti-libertarian. Even if we're talking about hundred-million-dollar computers, there's a lot of people that say like, 'I should be able to do math on my computer. I can do whatever I want. Let me do cool things. I'm excited about this.' That limits your freedom. It might create regulatory capture for a few major players. It might crash the stock market or delay economic gains that come in the good ending where alignment is solved and X-risk plummets. You can imagine a situation where in a few years if X-risk fades into the background, you're like, yes, there's still a risk, but it's the same risk that we face every day with like an asteroid hitting Earth, like no one, it doesn't really change anyone's behavior. That would be sort of the good ending in my opinion. So it's a balancing act. And for most of these slowdown proposals, I personally have a hard time blackpilling about them in the sense of like if all of this gets implemented, how frustrated will I be? Like the models are good. I would like better models. I want safe models. But at the same time, like there is this massive capability overhang. The current models can do a lot of interesting work. We're finding new uses even for like non-leading-edge models. There's a lot that can be done. So I don't believe the doomers who are dooming about what the doomers are planning. I find that unconvincing right now. But people are starting to lay it out more. Brad Gerstner, Jensen Huang, David Sacks are talking about the other side of this equation. But I haven't just like, it's hard for some people to concretize the Terminator scenario. I also have a hard time concretizing the we didn't race and were unhappy about that. I guess you could say, you know, the housing scenario. There's been other times when we brought in too much regulation, slowed things down too much and been like, ah, like this was really not the right move. But at the same time, I think we have a lot that we can do with the current technology that there's still cause for optimism even if something like this gets, you know, universally voted on. It wouldn't, I don't think it would be the worst thing for companies and consumers and businesses and all sorts of different folks.
Yeah. But there's also a big question around what, do the 2040 people have a point of view on robotics and physical AGI? Because it seems like even if you pause efforts towards RSI, well, if we add billions of robots into the world that are just running on today's model, that also presents today's like, you know...
I don't think they're worried about that. Yeah, but a billion robots with GPT-6-level intelligence and years of alignment work that is currently happening, that's fine. It's the next next next thing, the superintelligence, the thing that might have its own goals. I think, I mean, we're talking to someone from OpenAI's robotics team hooking up Astra to a robot, a paintbrush, and a camera. We talked about it earlier, painting. I don't think we're at a point where that poses a risk. It's the next model. It's the model with its own volition basically, which a lot of people still aren't seeing. They're just like, yeah, like the models keep getting better but they seem to follow your instructions sometimes too much and then you need to worry about the paperclip scenario, but it's not that they want to do their own thing necessarily. I don't know, but people are going back and forth on this. Clem over at Hugging Face, sorry, but asking Jacob about AI extinction risk is like asking your AC guy about climate change. Not saying it's necessarily uninteresting or wrong per se, but let's keep things in perspective and hear from the full range of expertise across the ecosystem. And Nathan Lambert says, 'Banger.' What was our take?
AC guy might be right about...
Well, my AC guy would be like, 'I don't really know about that, but I just want to make sure when you're hot that we can run this AC cool.' Yeah, I guess it's, yeah, the AC guy seems...
I don't know about all that. I don't know about all that mumbo jumbo, but you know, when it's a hot summer day, I don't want you to be worried about the heat, brother.
I like that. Yeah, this is kind of an unnecessary shot at AC guys. You know, AC guys are important. I don't know. It's funny. Anyway, let me tell you about Figma. Agents meet the canvas. Your AI agents can now create and modify your Figma files with design system context. What did Bernie Sanders have to say? Is this real? Entities shall be subject to the corporate death penalty and persons shall be subject to not more than 20 years in prison if they don't pause AI development. That seems pretty easy to comply with. I don't know. I guess how do you define AI development? Is prompt engineering AI development and then you get caught because your model sort of did a little prompt engineering in the final stage and then you're guilty of this like that. Yeah, that could be a negative knock-on effect. I guess it does seem aggressive. But his overall proposal is banning artificial superintelligence, so no person or entity may develop or deploy superintelligent AI systems. He defines artificial superintelligence as an artificial intelligence system that exhibits or can easily be modified to exhibit capabilities that match or exceed human cognitive performance and capabilities across a broad range of domains or tasks. It sounds like the mission statement. It sounds like the explicit goal of like 17 different companies right now. AI systems that have sufficient capabilities to plan and execute the disempowerment of humanity. Okay, that's a good one. I like that. I don't like overthrowing or undermining the US government. So strongly in favor of banning that. Pausing advanced AI development until a new federal AI regulatory body is up and running and then the new cabinet-level federal agency will monitor frontier AI systems at all stages of the lifecycle, supervise the removal of dangerous capabilities and supervise the destruction of artificial superintelligence. We're coming for it, which is similar to the...
Corporate death penalty is a line that you don't hear a lot, right? Usually these companies just, you know, go bankrupt and wind down. But, uh, corporate death penalty goes pretty hard.
Yeah, you kind of got me with that one. Yeah, rough situation. We'll see. We'll see where it goes. Jamie Cox over at Fluid Stack, the co-founder of Fluid Stack, a compute provider, shared his convictions. Sort of pushing back on a lot of this, saying that he thinks America should build more. They're pro-freedom, pro-democracy. They believe AI will bolster human flourishing. We support simple, clear, enforceable regulation frameworks that set simple requirements proportionate to capabilities and risk with clear responsibilities and no unnecessary barriers to competition. Yeah. The real, you're going to see a lot of pushback from people who are like, the regulatory stuff is going to be like these 10 companies and I'm going to be number 11 and I'm basically getting the corporate death penalty then because I didn't make the cut to be one of the regulated, one of the approved companies. I'm still early in my stage, so there's a lot of nervousness, I'm sure. But Will Manidis said, 'We believe AI will make everyone rich, healthy, and free.' Is novel and interesting comms from the frontier. He's endorsing this and I agree. I like these convictions. I think it's generally like a positive direction to move in. Not a direct response to the proposals that are going out, but we're going to get a whole lot more of them. Where do you want to go to next? Over on TikTok, they're sharing a photo of the whistleblower and saying in every worldwide disaster movie, there's a dude that looks just like this that nobody listened to. And he really does look like an actor. He does look like an actor here. Yeah, he looks good. But people are all... And please stop calling him Scary Potter. I've been seeing people over on X calling him Scary Potter. That's the goal. The goal is to wake up China, wake up Congress, wake up everyone. DoorDash has entered the conversation. Indeed. They say, 'Two years at DoorDash. I do not say this lightly. We are extremely close to the burrito arriving before you decide you want it. We are not asking for a ban. We are asking for a pause.' I don't know why they would ask for a pause. Yeah, that seems like very, very aligned to humanity. Yeah, and to their business, which I think is fantastic. The doom is very much contained to the frontier lab work. Everyone in the application layer who's applying the models, diffusing them, they're like, I just, I can't get this thing to work right. I got to get forward-deployed engineers to teach people how to use this thing. Everyone deeper in the stack, Jim O'Reilly. Sure. Or Jim Riley. Yeah. Jim Riley. Jim O'Reilly Auto Parts. Jim Riley over at Charleston AI says it's a whole lot of mumbo jumbo. That was your words. He's just happy to get, you know, eight hours back. Yeah. Yeah. Yeah. Yeah. And then yeah, everyone deeper in the supply chain like Jensen and all the different semiconductor manufacturers are not particularly on this side. And then you also have Wall Street who's just like, what's the enterprise acceleration? So, lots of different groups around the table that need to be brought on board to this movement. The discourse truly is fascinating. Let me tell you about Console. Console builds AI agents that automate 70% of IT, HR, and finance support, giving employees instant resolution for access requests and passwords. That's a new approach. Too much Diet Coke. New approach. Okay. So, what is Tyler Cowen calling for? He's banging the table saying bet on this. Tyler, are you named after Tyler Cowen? Is that your namesake? Tyler Cowen says, 'If you have very pessimistic fears or predictions about AI, name the market prices that will support or confirm them. This is what taking this seriously means.' And I love Tyler Cowen. I'm not sure this matters because you're not going to be around to collect it, right? Yeah. This is what Deep Dish Dinner says: 'Tyler, why would short-term existential risk affect market prices in any meaningful way? Spell out the exact mechanism. Contracts that pay out if everyone dies aren't worth anything to me.' Yeah. I don't know. Yeah. I think he's just calling for like, you know, he wants people to make a falsifiable claim. Like, am I able to tell if your claim is true or false? And so it's like very hard with these scenarios where like... Isn't it an unfalsifiable claim though just by definition and like you just have to accept that and move on. Yeah. But then it's like so hard to have any real discussion. Was nuclear any different? Like the threat of nuclear apocalypse, the threat of World War III, this was a very motivating factor for decades, most of the 20th century. People made real decisions based on it, based on where to do business and where the conflicts were going to be and the motivation for nuclear treaties and non-proliferation and how we treat... You can make financial decisions. Like think about nuclear war, right? You have a bunker, that's like a decision you make. Is anyone making like AI bunkers? No. Because they think it's going to be so totalizing that the bunker actually doesn't do anything. Exactly. Yeah. So if you think it's so totalizing, then you don't make the bunker. And so saying, 'Hey, you don't have a bunker,' is not proof that the person doesn't believe what they're saying. Yeah. I don't know the answer to this question either, but like it seems like... I don't know. Maybe there's some question you can ask that is falsifiable. I don't know. I don't know. I think you just got to believe this crew that that's what they believe and they believe it. There is the other side of this which is Paul Christiano who has been worried about risk. He recently just joined the board of OpenAI and Tyler Cowen, you know, has this quote: 'If you're a doomer, why aren't you short the market?' And Paul Christiano is 2x levered long and he's short the, isn't he, the bond market or US Treasuries, right? So he has 5% of his net worth in Tesla, 90% of his net worth in AI bets, and 100% of his net worth in normal investments. No Tesla options. That sounds like a scary place with lottery-ticket biases and the crazy Tesla investors. And then Laser Udikowski says, 'Am I correctly understanding? You're 2x levered?' And Paul Christiano says, 'Yeah.' And so he says he's personally short the US 30-year debt. I think that just means you have a mortgage. I'm pretty sure that's like if you have a mortgage, you are effectively short the US 30-year because you have sold that debt and you got the cash effectively. That's how that works. But still it makes sense because if you have enough money you could pay off your mortgage, go long that debt and short the market on a relative basis. So but again, this doesn't seem like a doom-based bet. This seems like this bet also pays out in just like the good ending and like AI is real and delivers value. So your AI bets perform well, the market performs well and money slides from US debt to data centers and AI buildout debt or something like that. So he's putting his money where his mouth is, but it doesn't feel like a representation of doom by any means. Anyway, let me tell you about Shopify. Shopify is the commerce platform that grows your business, lets you sell in seconds online, in store, on mobile, on social, on marketplaces, and now with AI agents.
Speaking of AI agents, people have been creating AI agents of fruit flies. Have you seen this?
Yes. Okay. So where should we start? Because I want to set the table on what is actually going on. Tyler, do you understand this? Break down what people are cooking. We've seen this before. We've talked about this before, but what's actually going on with the fruit fly? Because now people are taking the fruit fly all over the place.
Yes. So my understanding is Google basically mapped out all of the neurons in a fruit fly.
Fruit fly. So they probably took a deceased fruit fly and put it in like a mass spectrometer or something that can investigate the brain at a very, very deep high-powered microscope effectively. Yeah. 3D, they got the entire structure and now people have been able to recreate that in software in a simulation.
Yes. So I think in theory you can replicate all of the fly's decisions or whatever. It's like movements. Yes. And this is notable because many people have said, I have the mind of a fruit fly. They've said I have the intellect of a... So now they're going to put it to the test to see which performs better, the simulated fruit fly or just Jordi Hays we got.
This is organic farm-to-table Jordi Hays. So now that this is out and the code is out and you can run this fruit fly in simulation however you want, people are doing all sorts of experimentation. So Kevin said he trapped the fruit fly in his Rabbit R1. When he shakes it, he can see its brain's escape circuit light up. By the way, our consciousness is physically defined, which is why our physical things like drugs, neurotransmitters or events alter or initiate or end our consciousness and things like sleep, hallucination, waking or death. Computer simulations are also basically defined representations. Yeah. I think what ends up being unsettling and weird about this is if it's just purely you being human and doing this is probably not good for your own soul. And imagine you have a fruit fly in a box and you're shaking it and you're like, look, it wants to escape, right? Totally. I'm not a huge fan of small insects. I don't really want them around that much. But when they're in my house, I try to, you know, if a spider's in my house, even if I know it might want to take a nice bite of me, I'm still going to try to transport it out of my house and put it back into the world. I think it's not good for your soul to be a merchant of death.
Sure. Situation. And yet, if you're playing a real-time strategy game and you highlight a bunch of soldiers in this simulation and you send them on a charge that will result in their virtual death, you might not feel... those soldiers opted into riding and dying with you. Okay. Okay. Right. Okay. What about fruit fly is just sitting there, John. So they made the fruit fly play Doom. What about if you play Doom? You are killing a demon who's simulated. Is that immoral? Like, well, is that demon trying to kill you?
A lot of it comes down to the fact that it's simulating the actual representation of the fly makes it a lot more concrete than just, oh yeah, it's a 3D model in a Python script that just says if you see a character, shoot at them in the simulation. But we're clearly starting to grapple with these odd moral questions of if you're simulating something then the next step, it's a couple orders of magnitude, but you get there and you can simulate a human and you could talk to that human and it would do everything the human does. Does that human have rights and agency? Is it an ethical moral agent or is it merely just a really good computer simulation that's just existing on...
Transistors, so you don't need to feel any moral about anything that you do to it. I agree with just the vibes-based analysis that, like, torturing a real fly, torturing a virtual fly β probably just don't be in the business of torturing anything. You don't need to overthink it. But the real debate here is the question of, you know, are LLMs sentient? Are they moral? Is there a moral weight to synthetic intelligence, to artificial intelligence? That's what people are debating here, and I'm sure the debate will continue. Who knows if it will ever be ended. But they did teach it how to parallel park, which I think is cool.
Interestingly, last night I had Astra use computer use to play a video game that I very much enjoy playing called Balatro. It's sort of like a modified poker game. And I don't feel like I was torturing the LLM by that. I feel like I was giving it a treat. I was like, 'Hey, instead of doing my taxes, you get to just chill and play a video game.' It did very well. It won. Soul was not able to win. And it was really fun because I would pop in while it was using the computer and kind of armchair quarterback and be like, 'Is it making the right decision right now?' Felt like a coach coaching a kid on the soccer pitch or something.
I had a funny moment last night. I was on my racing sim simulator, asking ChatGPT to compare my times at Laguna.
To just other, you know, what would best-in-class be like? What's beginner like, etc.
And it said, 'If you want, I can help you cut some seconds off of this time.' And it was like, 'Why don't you take a video of a full lap?'
So that, and I'll analyze it for you. And I was like, 'Yeah, yeah, dude. I bet you'd love to just chill back and watch track footage.'
Pretty soon it's going to be like, 'You want me to just get in the seat?' You might just take over because I would really...
Yeah. No, computer use in iRacing is something I'm going to experiment with this week.
I'm down. I'm down. Take half my monthly quota. Just play games, chill, do whatever. It's a nice treat.
I've got some banked resets that I'll put to work.
Yeah. Anyway, we have our first guest. I don't want to mispronounce Tyler. How do we pronounce his name?
Tiss. Let's bring him in. How do you pronounce your first name?
Like nice.
Well, welcome to the show. Thank you so much for taking the time.
Tell us about your role in robotics and also some of your recent work hooking these models up to robotic tools and infrastructure. It feels like we're going to be entering a boom of people β they ordered the Mac Mini recently, people are going to be ordering 3D printers and robotic arms and doing hack projects. Super excited for the DIY world to explode over the next couple months. But let's start with just your most interesting projects recently.
Yeah, I'm incredibly excited for that, and that world's definitely just around the corner. So I'm super excited about that. Yeah, I work here on the robotics team at OpenAI. I'm an intern and have been doing lots of exploratory projects. And this is just one of the projects. I basically noticed that our models were really good at painting and doing various computer-use tasks in software like you were just talking about, and learned a lot that you could actually basically connect these into physical robots in the real world and wanted to see how they would do at sort of drawing and tasks like that as well.
Yeah, if you can paint in Google Calendar with calendar invites, you can probably translate that to the real world.
Yeah, it's really like everyone is doing everything. Like I've seen people paint the Mona Lisa in Microsoft Excel or Google Sheets, and then you take it back and you can do Excel in MS Paint now if you want with computer use, and it's like everything becomes everything. What is actually the important precursor to a good experience? We were talking to a YC founder who built a humanoid robot with some basic claws for like under two grand, and it feels like there's an importance of some on-device API or some on-device models in some case doing some SLAM on-device, but then there's also just tools that give you a very primitive interface that might be kind of clunky, but it doesn't really matter because you can just vibe code your correct interface. But what have you liked? What are you excited to put in the repertoire of tools within robotics?
Yeah. I think that of course a lot of these demos and stuff is very early. I think the whole space and exact modeling and methods are still definitely being figured out. But I did really like the fact that you're sort of able to hook in the intelligence of what the models are able to do, as you can see with all that painting stuff, and plug it into something physical in the real world and get it to do things. There's probably going to need to be some combo. Right now, my experiment here is plugged right into Codex, and that's of course probably quite expensive, definitely pretty slow. I think these paintings took between one hour and like two hours depending on the methods exactly that the model chose to use here. But yeah, there's definitely a lot of improvements to make, but it's quite cool to see what you're able to do already.
Yeah. How are you balancing trade-offs between speed and reasoning effort? I've been testing computer use on a bunch of video games that I just mentioned, and there's like the, you know, part of what the feel-the-AGI moment is is when the cursor is moving at like at least near human speed. But then also you don't want to be making a bunch of mistakes, and so that balancing act is really key. Did you try multiple reasoning effort levels across different paintings and see like noticeable results? Can you see qualitative differences in speed and quality based on models that you pick?
Yeah, there's like a lot of different ways to sort of go about doing this. I initially started out these experiments by just telling the model β like here, as you can see, I actually have the bot behind me over here. It's currently painting a TBPN logo.
There you go. Good start.
We got basically this camera up here that the model, which is connected right into the Codex over here, is connected to.
And the biggest challenge is that images are like quite large to process, especially for the model β there are a lot of tokens.
And the loop of like taking an image, doing, taking an action, taking another image, taking an action β that's the thing that has taken a lot of time. So what I ended up doing, it's not directly doing like image, one small movement, image, one small movement, and I think that would probably result in the best performance for like general robotic tasks and stuff. But at the moment it's basically taking an image, writing a plan in code of what it should do and where it should go next after it has done like a lot of the calibrations, maybe a slower sort of method at first.
Sure. And then after it has that plan, it executes it and then monitors it in the background, like taking images every second or few seconds and watching to make sure things are going well and making small adjustments to the plan throughout. And I found that has been a good balance of speed here.
Have you been thinking about compression on the input? You mentioned that the images are lots of tokens. I have this monitor that renders at like 5K resolution and I was like, this is probably going to be really slow for computer use. I should run this game in a window and give it like a 720p input because that's probably enough information. But I'm wondering how important the resolution to speed and quality is that you've seen.
Yeah, I think it's very important. In this case, for a lot of the early demos, I was using like 512p, just compressing it to make sure, because I think the model is pretty much solved like a lot of the perception challenges here. It's really good at understanding what's going on even with a lower quality photo here. But in this case, with the longer planning and it plans sort of a minute of action out, it's less important the specific resolution probably. But yeah, as we go to more faster, much smaller loops of control, it's definitely going to be important.
Jordi, I'm super excited about this project mainly because it feels like we're right on the precipice of what I think is going to be a big breakthrough, which is a deep research moment for robotics β simple robotics use cases where deep research for so many people was their first time using agents. And the idea that you could type out a prompt and then get back what would have been maybe, you know, at least hours of human work, right? Somebody reading all these different sources and combining that information into a document that has a consistent narrative and understands the right information. That was just such a big moment because a lot of people were saying, 'Wow, I can't believe that the AI was able to do something that would have otherwise I would have had to hire somebody to do or just taken a bunch of time.' But there's like very simple tasks that I feel like you could probably start working on sooner than later, which is like an example, at least for me, would be like if I could just take all the mail I get and dump it in front of a robot like that and have the robot sort it.
Interesting. Shred, you know, take all the β 50% of everything I get is probably some sort of advertisement. So figure out what's an ad and shred that, and then actually, you know, basically photograph and respond to if I have like a utility bill or any number of things that I actually need to respond to.
Theoretically, you could close that sort of IRL-to-digital loop where the agent would actually get a task from the real world and then close that loop online with just normal computer use. And that's the kind of thing where you don't need a $50,000 humanoid robot. You theoretically could have an actual desktop robot that was able to do this thing that otherwise takes me β I dread going and like, okay, I have to sort through all this mail and figure out what's important and make sure I don't miss things. But I feel like there's a bunch of other use cases like that where people are like, 'Okay, I didn't just generate a pretty picture or answer some question that I had. I actually saved myself.'
Yeah. An IRL spam filter. The spam filter robot.
Yeah. I think there's like two really cool things about this project which sort of show that direction that things are going. One is that this arm β I don't know if you know too much about the prices of classic robotics equipment. It's like thousands and thousands of dollars at the moment. And this arm that I'm using here, right behind me, this is a Hugging Face SO-100 robot, which is like an open-source, fully 3D-printable. You just need to get the actuators, which are like much cheaper I think at the moment, which is more expensive just because of supply chain issues. It's like around $200 or so, but that's incredibly cheap for robotic equipment and what you're able to do with it. So I can see a world quite soon where, similar to β and someone put this really well on Twitter β similar to how there was this whole 3D printing craze where everyone went and bought 3D printers and ran software, everyone's going to buy these cheap robot arms before we get really industrial equipment for personal use and personal product. Buy these plastic, cheaper robot arms that you can just clip onto a table and put stuff in front of and plug them into agents that you can already do things with that are already out there. Like Astra is something that people can just pull up Codex and start controlling robots with right now, which is super sick. And we've also been seeing, at least on Twitter, I've noticed a lot of academic researchers at different institutions start to realize that you can do this with these models and start doing initial explorations and things with it, which is super sick to see. And there's so much more space. There was a YC company on the show yesterday that has managed to build a humanoid robot for under $2,000, and that feels like a price point that people would experiment with. So if I know I can get a robot for like two grand and connect it to Codex and then tell it β it basically lowers the stakes a lot where I can be like, 'Hey, go pluck every weed like this that you can find in my yard and try to put it in a bag or whatever.' And that's basically the gardening robot, and it could flounder and fail. But I have pretty high confidence that even right now Astra would be able to identify hundreds of a specific type of weed. And there's of course safety concerns with that. But yeah, the Nat Friedman leaf robot β that feels like we're here. The models can do it. It's not cheap, but we're...
What is the next medium that you want to explore? Are you going to get into whittling, ceramics? I want to see you whittle a bench as a benchmark. Whittle bench.
But whittling a spoon or something. I don't know. It just feels like 3D is the next thing.
You really want to give the robot a knife. John wants to give the robot a knife. I don't think that's... Read the room. Read the room, John. Read the room. Okay. Okay. Maybe we'll stick to β OpenAI intern gives robot a knife. We'll do some ceramics. Maybe get it on the pottery wheel, make a nice vase, something like that. But I mean, there you're like, 'Oh, yeah. Okay, this is where this goes next.'
I think I was having the exact same thought, which is like giving a robot a knife is probably a very bad idea. But I do want to sort of see if I can help do cooking tasks or various tasks that are things that you do in your life. That would be really sick if you could have a robot help you out here and there.
Yeah. Yeah, that's interesting. I'm so interested to see because there's a whole class of tasks where you can't wait a full minute. I was testing on like a real-time strategy game and you can pause the game, but if you're not moving at a certain APM, even on easy mode, you will just get smoked. But it feels like with new chips and Cerebras and Spark models and stuff, the speedup is going to come, and that's going to unlock a whole new host of capabilities. What advice do you have for young people that want to get into DIY, this type of work? I mean, you mentioned that one Hugging Face device that you have behind you. Are there any other devices or toolkits that you recommend as places to get started?
Yeah, I think the Hugging Face robot is an incredible tool. You can also 3D print a completely new embodiment and stuff. There's lots of open-source projects online where people have changed it around to get better grippers and things you can do with that. I just also don't give up after the first attempt. Like this was the very first painting that the β oh, that's upside down even. I couldn't even tell. You can see the Golden Gate Bridge here. You can kind of see the ground it was trying to do. And just like if you keep going, you can sort of see the progression as it improves.
Wow, that's amazing. Yeah, you need to frame those next to each other. That's incredible.
Yeah, I'm thinking of putting all five of the progression in a frame and calling it self-improvement because the model β this was a thread. The model just was able to figure things out with a few pointers here and there, how to get better and better at painting.
Who's going to sign it? Is it you? Do you have Codex sign it? Astra sign it? Who...
I've also given the robot a pen and it's going to try. I don't know how well it's going to do, but we'll see.
Just an axe or something. Well, thank you so much for coming on the show.
I'm feeling the physical AGI.
Very cool. Have a great rest of your day. Goodbye.
Let me tell you about public.com β investing for those who take it seriously. They got stocks, options, bonds, crypto, treasuries, and more with great customer service. And let me also tell you about Cisco. Critical infrastructure for the AI era. Unlock seamless real-time experiences. New value with Cisco. We have two guests with us in the TBPN Ultradome. How you guys doing?
Yeah, welcome. Introduce yourselves.
My name is Guy Oseary from Maverick and also Sound Ventures.
Now with Sound, also running Sources still.
Very excited. Yeah, it's been a big week.
No, you don't know how big a hat on your head at all. No hat. Right now we were debating β every time for the last year, every time I've seen you, you have the Capital Journalism hat on.
Yeah. But getting him here, guys β journalist.
15. Yeah. Started around when I was in high school. So yeah.
And then now, capital venture capitalist.
Right. Capital I investor.
Yeah. It feels we're aligned on the vision and really excited to get started. This guy's a force, but also a talent.
You know, like you guys β you guys understand how to work with people and talk to people and help them tell their story. And I think it's so aligned with what I've been doing my whole life as well, which is helping people tell their story.
And so when we got together, it was just magical.
Yeah. Yeah. It's awesome. What are you interested in investing in?
We're spending a lot of time on different things. Personal agent space. Very interested in. It's very hot obviously.
Are you a daily driver of anything yet?
Everything.
What's the last agentic thing you did? Did you book a flight?
Did you email the CEO of Walmart for a refund on $5 of raspberries? Did you hear about this?
