AI 2040 Isn't About Stopping AI — It's About Pushing Superintelligence to 2040
The real mechanism behind the anti-AI proposal is a compute cap: pause new frontier training, allow only inference, and make weight theft impossible with a 1 MB/s external link and a Faraday cage.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI 2040 doesn't stop AI, it delays it to 2040
Many people assume anti-AI sentiment is pushing to ‘stop AI’, but what the AI 2040 proposal actually says is: keep inference on existing models, keep doing capability research, and just push superintelligence from 2028 to 2040. The concrete mechanism is a compute cap — pause new frontier training, apply ‘inference-only verification’ to data centers above 10,000 H100-equivalents (roughly $100 million of equipment), and physically rip out the high-bandwidth east-west networking inside the data center so distributed training can't run but inference proceeds as normal. New R&D centers would have to be built from scratch inside a Faraday cage, with external connectivity capped at 1 MB/s — stealing the weights would take five years.
— Speaker 1Anti-AI mobilization is nowhere near anti-battery-chicken
Speaker 2 points out a contrast: the anti-flock (anti-battery-chicken) movement converts into real physical action by ordinary people, whereas on the anti-AI side you don't see a hundred-thousand-like video teaching you ‘how to cut power to your local data center’. Speaker 1 adds that the volume on the anti-AI side comes more from concrete self-interest — ‘I'm an artist, I don't like AI-generated images’ — than from x-risk arguments. Data centers sit in remote locations, behind fences, heavily fortified, far harder to get at than a chicken coop on the street.
— Speaker 2The compute cap is the valve easiest to turn
The most effective control lever in the proposal is judged to be the compute cap: chip controls plus slower data center construction. The goal is to make models stronger mainly by adding hardware rather than by new algorithms — because algorithmic breakthroughs can be hidden inside secret projects. Speaker 1 notes that Ilya's new lab recently got a large cluster from NVIDIA; even people taking an orthogonal research path have to stack compute, which suggests ‘scale is the precondition’ is basically consensus in this circle.
— Speaker 1Doomists don't bet, because they wouldn't live to collect
Tyler Cowen demands that pessimists name the corresponding market price to validate their belief, but Deep Dish pushes back: why should a short-term existential risk be reflected in market prices? ‘If everyone's dead, the contract that pays out is worthless to me.’ Speaker 5 twists the knife: in the nuclear-war era people built bunkers; nobody builds an AI bunker, because everyone thinks it's too totalizing for a bunker to help — so ‘you didn't build a bunker’ doesn't prove you don't believe. Paul Christiano was dug up as being 2x levered long and short 30-year US Treasuries, but that bet also makes money in the ‘good ending’, unlike a doomsday wager.
— Speaker 1The fruit fly agent moves the moral question forward
Google mapped every neuron of the fruit fly, and once the code was open-sourced people started messing with it in every way: someone put the fruit fly agent inside a Rabbit R1, and shaking the device makes its escape circuit light up; someone else made it play Doom. Speaker 1 thinks the real question is the next stop on this road — if you can simulate a human, hold a conversation, behave like a human, does it have moral status? His own judgment is not to overthink it, and not to be in the business of torturing anything. He also mentions that last night he had Astra play Balatro using computer use, and it felt like giving it a reward rather than torturing it.
— Speaker 1Once a reporter is in the room, he stops being escorted out
Alex Heath describes the core mechanism of his changed status: when a reporter is in the room, you're brought in and then immediately escorted out; now he gets to stay in the room and slowly marinate on it. So his newsletter perspective changes — it's not about landing a scoop, but about letting a large volume of off-the-record conversation settle into a judgment, like the personal agents piece he's writing. He says explicitly that confidentiality still matters, and he won't leak a memo a company is worried about again — ten years in, that's enough.
— Speaker 3A newcomer can hardly replicate his path now
Asked whether now is a good time to become the next Alex Heath, his answer is the opposite: the traditional media environment is structurally challenged, many places aren't hiring, traffic is falling, and they haven't shifted to direct models like subscriptions and streaming. The only viable path he offers is ‘maniacally focus on one thing and become the best in the world at that’ — he himself started by focusing on covering Snap, broke a pile of Snap news, and once noticed, migrated that capability to other companies. And the niche has to be the intersection of niche and matters, not just obscure.
— Speaker 3Fear of being left behind hands new players an opening
A host other than Mitesh offers a counterintuitive mechanism: because people fear AI, they're more willing to try new products — they don't want to be left behind, don't want to become a permanent underclass. The result is that people who have used various AI products for years are still willing to try new things, which keeps creating an advantage for second and third entrants. At the same time, once an AI brand gets big it starts accumulating baggage, and people become more willing to share and discuss something new rather than the one they use every day. That leaves a lot of room for startups.
— Speaker 2The endgame for personal agents is network effects
The discussion compares personal agents with the inference market for frontier models: every company will build a personal agent, and many already have. The key disagreement is whether this is a network-effects business — one side says yes, and that it can eventually take a cut of agentic commerce, e.g. you buy a car through it and it earns $500; Zuck has also said take rate is his business. The other side offers a counterexample: an agent isn't a person, you may not care whether it's in the network; but if some agent has never leaked, never crashed, never been hacked, you will indeed stay.
— Speaker 1The window to rebuild social networks is opening
One concrete judgment: now may be the best time to rebuild a social network, because you can have your agent open your Snap or LinkedIn, pull out every single person in there — they have no say in it — and add them back on the new network. The old arbitrage window of ‘export your contacts’ closed; now it's reopening. The immediate criticism: the old critique of social media was ‘we built social media for socializing, and it made people less social’; the new critique in the personal-agent era will become ‘nobody talks to people anymore, only agents talk to each other’ — even chatting with grandma about a road trip becomes ‘talk to my agent’.
— Speaker 1Chip companies are now asked whether they can build it
Mitesh says that walking into a meeting room at a hyperscaler or frontier lab, the question isn't performance but two things: first, can you build enough of it — is the supply chain and component base robust enough; second, can you deploy it — does it need liquid cooling, and if so they may not have a data center for it, because it's already allocated to GPUs or TPUs. The threshold number he gives is gigawatt plus; at that scale, even counting ASICs as cheaper, it's tens of billions of dollars. His own cadence is tape out this year, volume ramp in the second half of 2027, and a plan to reach several hundred megawatts in 2028.
— Speaker 4Using commodity memory to dodge the HBM and CoWoS bottleneck
Positron's technical route is to dodge the HBM and CoWoS queue — NVIDIA, TPU and AMD are all ahead of you. The approach is to use commodity memory, at the cost of being slow, so the technical innovation lies in how to solve the slowness. But commodity memory isn't freely available either; you can't just walk up to Samsung or Micron and ask for LPDDR5x, you still have to plan. He mentions the first-generation product was built on FPGA, with fewer than 20 people, delivering to the first customer in 15 months; Oracle now has 50 Atlas racks; the company just passed 100 people, more than 50 of whom joined in the last three months.
— Speaker 4The robot takes one to two hours to draw a picture
Thijs hooked Codex directly up to a Hugging Face open-source robotic arm to draw the TBPN logo, and one drawing takes one to two hours; he himself says it's ‘probably quite expensive’ and ‘definitely pretty slow’. His approach isn't to have the model look at an image and move once, but to first have the model write a plan in the form of code, execute it, then take a picture every few seconds to monitor in the background and fine-tune the plan. He judges that a high-frequency image–small action–image closed loop would perform best, but the current ‘plan first, then monitor’ is a speed compromise.
— Thijs SimonianRobotic arms went from thousands of dollars to two hundred
Thijs says traditional robot hardware costs thousands to tens of thousands of dollars, whereas the Hugging Face SO-100 behind him is open-source, fully 3D-printable, and only needs actuators bought separately, currently around $200 (he adds that this price is on the high side, caused by supply chain issues). He compares it to the 3D printing boom: before industrial-grade machines became widespread, people would first buy these cheap plastic arms, clamp them to a desk, plug in an off-the-shelf agent, and use them.
— Thijs SimonianGuy Oseary put nearly a billion dollars into OpenAI and Anthropic
Guy Oseary says that when he invested in Anthropic and OpenAI, they may have been the only fund betting that deep, and many people were confused at the time, but they were convinced this was a long-term foundational platform. He reveals they added more in both companies' later rounds, for a cumulative total approaching $1 billion. But he also says that telling what's real from what isn't is far harder today than it was then — ‘It's not as easy as it was for us to really decipher’.
— Guy OsearyListen to a pitch the way you listen to a song: find the chorus
Guy Oseary likens spotting founders to collecting demo tapes in clubs back in the day: at 17 he had 100 demos in hand, and the ones initially ranked top three had fallen out of the top 20 by the time he reached the hundredth. He says he listens to every pitch like a song or an album, and the test is ‘is there a chorus’ — ‘I'm always listening for the chorus’; if he can't hear a chorus, he figures the song doesn't work. Alanis Morissette played 30 to 40 seconds of Perfect in his office and he decided to sign her; Muse flew in from London, and after the first song he stopped them and said they could start.
— Guy OsearyAlex Heath spots faking by having seen the very best
Alex Heath says he can see through marketing at a glance because he has spent years around the world's top founders — he just did Zuckerberg, and before that Sam Altman. Once you've seen people at the apex, with high integrity, dominating on the field, you can quickly tell who's faking. He says he often has the instinct that ‘this looks like it will pop’, and it has come true often enough. Guy adds that this instinct came from having to move fast back then: if you didn't decide to sign an artist immediately, Jimmy Iovine would outbid you and take them.
— Alex HeathIn their own words · checked verbatim
So the whole goal, you can only inference the current models and you gotta verify your workloads with an independent auditor, probably the government.
Speaker 10:00
They want the r and d data center to be connected externally. If you want to communicate with it and and you want to tell it what to do, okay, train the next running or do whatever, they will have a bandwidth capped connection at one meg per second.
Speaker 10:00
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 taken the stuff out of the data center.
Speaker 10:00
But at the same time, I think we 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.
Speaker 10:00
when you're a journalist and you're in the room, you like get brought into the room and then escorted right back out. And now it's like I get to hang out in the room.
Speaker 30:00
No no free lunch in Silicon Land.
Speaker 40:00
I think these paintings took between one hour and, like, two hours depending on the methods exactly that that model, like, Toast to use here.
Thijs Simonian2:20
Similar to how there's this whole three d printing craze where everyone went and bought three d printers and ran software. Like, everyone's gonna buy these sort of cheap before we get, like, really industrial equipment for, like, personal use and personal product.
Thijs Simonian6:40
But music is the constant. I'm always listening for the chorus.
Guy Oseary22:10
So when you see like the people at the apex crushing it and who are at high integrity or beasts at the game on the field, you you can quickly see when someone is pretending.
Alex Heath28:45
Figures
| R&D data center external connection speed cap | 1 MB/s | 0:00 |
| AI 2040 target timeline | Top human-expert capability by 2035, superintelligence unlocked in 2040 | 0:00 |
| Size of the anti-AI protest in Chapel Hill | About 150 people | 0:00 |
| Price of the Hugging Face SO-100 robotic arm | About $200 | 6:40 |
| Time for one of Thijs's robot drawings | 1 to 2 hours | 2:20 |
| Guy Oseary's cumulative investment in OpenAI and Anthropic | Approaching $1 billion | 16:20 |
| Positron headcount | Just past 100 people, more than 50 of whom joined in the last three months | 0:00 |
| Positron first-generation product delivery cycle | First customer delivered in 15 months, built on FPGA, team under 20 people | 0:00 |
| Atlas racks deployed by Oracle | 50 | 0:00 |
Glossary
- AI 2040 / the AI 2040 proposal
- A regulatory plan that would pause new frontier training, allow only inference, and push superintelligence to 2040.
- SO-100 / SO-100 robotic arm
- Hugging Face's open-source, 3D-printable desktop robotic arm, around $200.
- commodity memory
- General-purpose memory such as LPDDR5x, used instead of HBM to dodge the CoWoS queue.
- CoWoS / CoWoS packaging
- TSMC's advanced packaging technology, required to integrate HBM with GPUs, and capacity-constrained.
- gigawatt plus
- A data center power threshold; at this scale even ASICs require tens of billions of dollars.
How to listen
Founders and investors watching AI regulation, robotics and the chip supply chain; practitioners who want to understand how VCs read people and where personal agents are heading.
If you don't care about music-industry people-reading stories, skip the Guy Oseary segment from 22:10 to 28:45.