
Sep 11, 2026 Β· 1h 46m
AI risk meets robot arms, agent investing and chip competition
Fruitfly Hard Takeoff, Washington on AI Risk, π Timeline Reactions | Thijs Simonian, Alex Heath & Guy Oseary, Mitesh Agrawal
The episode connects abstract arguments about AI safety to hands-on robotics, emerging agent businesses and the hardware race challenging NVIDIA.
- 1AI risk proposals hinge on whether doomer predictions can produce testable market signals.
- 2A low-cost robot arm shows how models trade reasoning quality against latency in physical tasks.
- 3Agent investing and inference-chip competition reveal how AIβs next phase depends on distribution, financing and infrastructure.
Don't miss
Thijs Simonian explains how Codex controls a camera-equipped robot arm to paint while balancing image costs, latency and action planning.
The brief
The hosts open with a 2001 Wall Street Journal time capsule before asking whether public reactions to AI and data centers resemble earlier moments of technological fear.
A debate over compute caps and AI doomerism turns on falsifiability: if existential-risk predictions are serious, what observable market prices or actions should support them?
The fruit-fly connectome sharpens the ethical question: if an organism can be simulated, does interacting with its digital counterpart create new obligations around consciousness and harm?
OpenAI robotics engineer Thijs Simonian demonstrates a camera-equipped arm that paints, exposing the practical trade-off between model reasoning, image-token costs and real-time control.
Guy Oseary and Alex Heath discuss personal agents and venture pattern recognition, while Positronβs Mitesh Agrawal explains the financing, software and scale challenges behind an NVIDIA alternative.
The episodeβs throughline is AI becoming physical and commercial: safety claims meet experiments, agents meet markets, and inference chips meet the demands of deployment.
What was said on this episode
54 statements Β· 37 positive Β· 8 negative Β· 1 mixed Β· 8 neutral
Large-scale compute is considered necessary for AGI
βScale. That scale is a prerequisite.β
Listen at 16:28
AI slowdown proposals might cause regulatory capture, a stock-market crash, or delayed gains.
βIt might create regulatory capture for a few major players. It might crash the stock market or delay economic gainsβ
Listen at 19:45
AI slowdown controls might create regulatory capture
βIt might create regulatory capture for a few major players.β
Listen at 19:45
AI slowdown controls might crash markets or delay economic gains
βIt might crash the stock market or delay economic gainsβ
Listen at 19:49
Universal AI slowdown regulation would not be especially harmful
βI don't think it would be the worst thing for companies and consumers and businesses and all sorts of different folks.β
Listen at 21:37
AI 2040 is not concerned about a billion aligned GPT-6-level robots.
βI don't think they're worried about a billion robots with GPT-6 level intelligence and years of alignment work that is currently happening. That's fine.β
Listen at 22:12
The main AI risk concerns models with independent goals
βIt's the model with its own volition, basicallyβ
Listen at 22:51
Paul Christiano has expressed concern about AI risk
βPaul Cristiano, uh, who, uh, is, has been worried about risk.β
Listen at 32:29
Paul Cristiano is two-times leveraged long.
βPaul Cristiano is, uh, two X levered. longβ
Listen at 32:43
Paul Cristiano reportedly allocates 5% to Tesla and 90% to AI bets.
βhe has 5% of his net worth in Tesla, 90% of his net worth in AI betsβ
Listen at 32:57
The fruit-fly simulation can theoretically replicate fly decisions and movements
βin theory you can replicate all of the fly's decisions or whatever, it's like movements.β
Listen at 35:23
Consciousness is physically defined
βour consciousness is physically definedβ
Listen at 36:10
A sufficiently advanced simulation could reproduce human behavior and conversation.
βyou can simulate a human and you could talk to that human and it would do everything the human does.β
Listen at 38:24
The moral status and sentience of LLMs remain debated
βthe real debate here is like 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.β
Listen at 38:57
- Thijs Simonianon DIY roboticsPositive41:53
DIY robotics is expected to expand soon
βthat world's definitely, like, just around the corner.β
Listen at 41:53
- Thijs Simonianon OpenAI models and physical robotsPositive42:17
OpenAI models can be connected to physical robots for drawing tasks.
βyou could actually basically connect these into physical robots in the real world and wanted to see how they would do at drawing and tasks like that as well.β
Listen at 42:17
Planning then monitoring provides a good robotics speed balance
βI found that has been a good balance of speed here.β
Listen at 46:22
The model has largely solved perception challenges in this robotics task
βthe model has pretty much solved a lot of the perception challenges here.β
Listen at 47:02
- Thijs Simonianon Hugging Face SO100Positive49:59
The Hugging Face SO100 is open-source and fully 3D printable.
βthis is a Hugging Face SO100 robot, which is open source, fully 3D printable.β
Listen at 49:59
- Thijs Simonianon Hugging Face SO100Positive50:11
The Hugging Face SO100 costs approximately $200.
βit's around $200 or so.β
Listen at 50:11
- Thijs Simonianon consumer robot armsPositive50:27
Consumers will buy inexpensive robot arms for personal use before industrial equipment dominates.
βEveryone's going to buy these sort of cheap, before we get really industrial equipment, for personal use and personal product.β
Listen at 50:27
- Thijs Simonianon Robotic knife useNegative52:50
Giving a robot a knife is probably unsafe
βgiving a robot a knife is like probably a very bad idea.β
Listen at 52:50
- Thijs Simonianon robot assistancePositive52:54
Thijs wants to explore robots assisting with cooking and everyday tasks.
βI do want to like sort of see if I can help do like cooking tasks or do various tasks that are like things that you do in your life that would be really sick if you could have a robot help you out here and there.β
Listen at 52:54
- Thijs Simonianon Robot painting modelPositive54:40
The model improved its painting through limited guidance
βThe model just was able to figure things out with a few pointers here and there, how to get better and better at painting.β
Listen at 54:40
The personal-agent sector is currently highly attractive for investment.
βPersonal agent space. Okay. Very interested in. It's very hot, obviously.β
Listen at 56:59
Guy Oseary viewed OpenAI and Anthropic as enduring foundational platforms
βthese were the companies. These were going to be longstanding foundational platforms.β
Listen at 59:14
Guy Oseary believed OpenAI and Anthropic would become longstanding foundational platforms.
βThese were going to be longstanding foundational platforms.β
Listen at 59:15
Sound Ventures has nearly $1 billion invested in Anthropic and OpenAI.
βwe have close to a billion dollars worth of money invested into those two companies.β
Listen at 1:01:26
Experience with top founders helps identify people who are pretending.
βwhen you see the people at the Apex who are crushing it and who are at high integrity, are beasts at the game, on the field, you can quickly see when someone is pretending.β
Listen at 1:13:31
Experience with top founders helps identify people who are pretending
βyou can quickly see when someone is pretending.β
Listen at 1:13:38
Alex Heath expects to interview the Magnificent Seven CEOs
βThe mag seven, I feel pretty good about. Yeah. You're on that track for sure.β
Listen at 1:15:40
Inference is currently highly important in the AI ecosystem.
βinference is just so important right now.β
Listen at 1:16:26
Aspiring technology journalists should focus intensely on one important niche
βYou have to maniacally focus on one thing that matters.β
Listen at 1:22:41
Aspiring journalists should focus intensely on a niche that matters.
βYou have to maniacally focus on one thing that matters. But it's the intersection of niche and matters.β
Listen at 1:22:41
Alex Heath would not be surprised if Instinct exceeds Snap's valuation
βI wouldn't be surprised.β
Listen at 1:23:00
Agent commerce with network effects and transaction fees could be highly profitable
βIf you get to network effect and you get to some sort of take rate on agenda commerce, like you buy your car through it and they make 500 bucks, like that's a very, very good business.β
Listen at 1:25:12
Instinct, Town, and Meta are pursuing network effects for agents.
βAnd Instinct's doing it, Town is doing it, Meta's going to do it.β
Listen at 1:25:30
Agent networks may become the next major network effect
βmaybe that is the next network effect.β
Listen at 1:25:36
Memory will become increasingly important for AI inference
βMemory is going to get a big part of the story for inference.β
Listen at 1:29:23
NVIDIA leads the market while releasing a new chip selection every 12 months.
βnew chip every 12 months. NVIDIA is the absolute king.β
Listen at 1:30:21
Nvidia's annual chip releases pressure competitors to match its pace
βif they're coming out with a new selection every 12 months, you better get in that gameβ
Listen at 1:30:24
Positron shipped its first customer product within 15 months.
βwe got our first shipment to a customer in 15 months.β
Listen at 1:31:24
Inference hardware combining low cost and interactivity should attract demand
βif you can do and or, you're going to bound to have get demand.β
Listen at 1:33:48
AI-chip startups should reach 10% of NVIDIAβs revenue before claiming to rival it.
βno one is rivaling NVIDIA. Like, get to at least 10% of their revenue before putting the tagline onβ
Listen at 1:36:05
New chip companies are not currently rivaling Nvidia
βno one is rivaling NVIDIA.β
Listen at 1:36:05
Hyperscalers and frontier labs require AI hardware proposals exceeding one gigawatt.
βfor hyperscalers and Frontier Labs, honestly, they will literally say gigawatt plusβ
Listen at 1:38:26
Hyperscalers and frontier labs require proposals exceeding one gigawatt
βthey will literally say gigawatt plus, come up to us with a proposal of a gigawatt plusβ
Listen at 1:38:30
Positron expects to plan for hundreds of megawatts by 2028
βIn 2028, we better have a plan of how do we get to, like, hundreds.β
Listen at 1:39:00
Positron could grow revenue through alternative customers to $200 million, $250 million, $1 billion, or $2 billion.
βyou can grow the company, you know, get 200 million revenue for 250 million, a billion, 2 billion through this other kind of channels as wellβ
Listen at 1:39:48
Positron's first-generation system can run new model drops within hours
βwithin our team, you know, on our ad, let's first-gen could get it up and running within hoursβ
Listen at 1:42:49
AI automation will probably handle model-drop integration
βModel drops, comes in, can probably do it.β
Listen at 1:43:04
AI agents will soon automate bringing new model releases into production.
βModel drops, comes in, can probably do it. And that's a very near future of it.β
Listen at 1:43:04
Positron will build its software stack around open source
βThe software stack is going to be built around open source.β
Listen at 1:43:55
Positron raised $875 million.
βYou raised $875 millionβ
Listen at 1:44:32
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.