Video · 2026-06-28 · 1h 3m · 6 moments

The Thermodynamic AI Chip · Thomas Ahle

✦ AI generated

timeline · colored by role

Data

Commercial hardware verification/simulation licenses cost around $10,000 per seat for a single CPU kernel, so scaling agentic AI use to a data center with a million agents would cost on the order of $10 billion in licensing alone — a major reason AI models aren't well-trained for hardware workloads.

Thomas Ahle notes that scaling AI agents to hardware workloads would cost roughly $10 billion in commercial EDA licensing fees, which he says is part of why AI models remain undertrained for chip design.

transcript

Thomas Ahle: Like say you want to scale this up in a data center with a million agents running. Like you're going to like I mean comput is already expensive but not that expensive. Uh what yeah what is that $10 billion or something right? Just for for those licenses.

Claim

What matters in agentic coding is not whether you end up with passing tests, but how you got there and how structured the process is — agentic coding tends to produce a working 'spaghetti monster' of code that no one has actually read or understood.

Tim Scarfe argues the real risk of agentic coding isn't whether tests pass, but that it produces sprawling 'spaghetti monster' codebases (like the 500,000-line Verilog simulator) that no one has actually read or understood.

transcript

Tim Scarfe: my contention with this is I think that it's not about where you end up. It's not about the functions and the tests passing. Um, it's about how you got there and how structured it is. So there is this tendency with agentic coding to build a spaghetti monster which seems to work.

Claim

Benchmarks that only report the percentage of tests an AI-generated program passes are misleading, because a program that passes just 70-80% of tests is probably not actually correct — what matters is whether it got any test fully right, not the aggregate score.

Thomas Ahle argues that reporting benchmark pass-rates (like '70-80% of tests correct') is misleading, since a program that fails even some tests is likely not truly correct overall.

transcript

Thomas Ahle: They always post this like oh the percentage of tests that it got right or like even like the Fable it was like oh yeah it got like 70 80% tests correct and it's like yeah but I talked with the people who made the benchmark and like yeah but it did it get any of them actually right? you know, like did it because it's if if the program only passes 70% of the tests, it's probably not right.

Mechanism

In training AlphaProof, it didn't actually matter whether the auto-formalization of a problem was correct, because asking the model to prove or disprove the statement worked either way — an incorrectly formalized (false) statement simply gets 'proved false,' so it remains usable as training data.

Thomas Ahle explains a key trick behind AlphaProof: since the model is trained to prove OR disprove a formalized statement, an incorrect auto-formalization still yields usable training signal, so correctness only needed to be checked by hand for the final IMO submissions.

transcript

Thomas Ahle: I think it's actually really trick in alpha proof was that when they did the formalization of the proof, it didn't really matter if they got it right or wrong because um if they just asked the model to provide a proof or a disproof. So if they if they got it wrong and it was no longer true then it would just prove that it was not true or like it would they would like you you could still use it.

Mechanism◆

Rather than fighting to eliminate noise the way chip manufacturers normally do, it makes sense to build a chip that is inherently random, since infusing controlled noise lets the physical circuit itself behave like a stochastic differential equation that can compute useful quantities.

Thomas Ahle explains the core idea behind Normal Computing's thermodynamic chip: instead of eliminating noise like conventional manufacturers do, they inject randomness so the chip's own physics performs the computation.

transcript

Thomas Ahle: But then, but then it's funny because you have these chips and like the chip manufacturers, they spend so much time like getting out every single little piece of noise out of their systems and like having these extremely sharp margins for everything. Like so much, you know, precision. It's probably like most precise business in the world. And then what do we do with them? We just like add randomness everywhere. Um and um yeah, so why not try and build a chip that's just inherently random?

Claim◆

It's not just that AI is getting smarter — humans are getting dumber too, because relying on AI to explain papers and do the thinking makes people lazy about actually understanding things themselves.

Thomas Ahle warns that AI risk isn't only about AI capability increasing — it's compounded by humans getting lazier and less knowledgeable, e.g. no longer reading papers themselves and just asking AI to explain them.

transcript

Thomas Ahle: it's not like it's not just that it's getting smarter. It's also that humans are getting dumber like we no longer um like know and like we get we get lazy in terms of understanding stuff. We don't read the papers. You just put them in AI and be like, "Oh, explain this paper to me or something."

Highlight slides

Chips Are Built to Kill Noise✦ from: Rather than fighting to eliminate noise the way chip manufacturers normally do, it makes sense to build a chip that is inherently random, since infusing controlled noise lets the physical circuit itself behave like a stochastic differential equation that can compute useful quantities.Normal Computing Inverts the Approach✦ from: Rather than fighting to eliminate noise the way chip manufacturers normally do, it makes sense to build a chip that is inherently random, since infusing controlled noise lets the physical circuit itself behave like a stochastic differential equation that can compute useful quantities.Noise as a Computational Resource✦ from: Rather than fighting to eliminate noise the way chip manufacturers normally do, it makes sense to build a chip that is inherently random, since infusing controlled noise lets the physical circuit itself behave like a stochastic differential equation that can compute useful quantities.The Real AI Risk: A Two-Sided Problem✦ from: It's not just that AI is getting smarter — humans are getting dumber too, because relying on AI to explain papers and do the thinking makes people lazy about actually understanding things themselves.How Humans Are Getting Lazier✦ from: It's not just that AI is getting smarter — humans are getting dumber too, because relying on AI to explain papers and do the thinking makes people lazy about actually understanding things themselves.

Related episodes

Article · Import AI

What matters in agentic coding is not whether you end up with passing tests, but how you got there and how structured the process is — agentic coding tends to produce a working 'spaghetti monster' of code that no one has actually read or understood.

Benchmarks that only report the percentage of tests an AI-generated program passes are misleading, because a program that passes just 70-80% of tests is probably not actually correct — what matters is whether it got any test fully right, not the aggregate score.

In training AlphaProof, it didn't actually matter whether the auto-formalization of a problem was correct, because asking the model to prove or disprove the statement worked either way — an incorrectly formalized (false) statement simply gets 'proved false,' so it remains usable as training data.

Rather than fighting to eliminate noise the way chip manufacturers normally do, it makes sense to build a chip that is inherently random, since infusing controlled noise lets the physical circuit itself behave like a stochastic differential equation that can compute useful quantities.

It's not just that AI is getting smarter — humans are getting dumber too, because relying on AI to explain papers and do the thinking makes people lazy about actually understanding things themselves.

Import AI 464: Fables writes GPU kernels; AI automation; and analog computation

Import AI 464: Fable writes GPU kernels; AI automation; and analog computation

[AINews] OpenAI GPT-5.6 Sol / Terra / Luna — restricted to trusted partners