The Eureka Machine: Recursive Superintelligence for Science

TL;DR

• The "Eureka Machine" is his personal version of going to Mars: the ultimate invention — a machine that automates all of humanity's future inventions. Getting there requires building recursive self-improving superintelligence (RSI) along the way.
• Four pillars: existing human knowledge (now carried by LLMs) → scientific measurement data → simulation → automated physical labs. On top of them runs an agent swarm with the equivalent of ~5,000 PhDs.
• Start with the science of AI itself. This became uniquely possible this year because AI is code and AI can now write code — over increasingly long time horizons.
• Receipts, not slideware: their RSI system beat years of accumulated community work on nanochat, the nanoGPT speedrun, and NVIDIA's SOL-ExecBench — and produced genuine inventions (hash tables inside transformers, new forms of momentum), not hyperparameter sweeps.
• We are astronomically far from the upper bounds of intelligence. AI decomposes into prediction (mathematically equivalent to compression) × actions × goals; even the simplest case, visual intelligence, has an upper bound of trillions of sensors spanning quantum uncertainty to gravitational waves.

Key Points

Why now: technology as the only perpetual source of growth (~00:09–00:13)

He opens on open-ended evolution, a process he thinks the broader AI community still underexplores. Biological evolution took over a billion years to invent the eye; technological evolution compressed the cycle to millennia; AI now runs it in weeks.

Citing Marc Andreessen, he argues technology is the only perpetual source of growth — and goes further: there are no material problems that cannot be solved with more technology (psychological and political problems being a different beast). On that view, not accelerating is the bigger danger, because it means fewer people get to flourish.

His framing line: our generation was too late to explore the earth, too early to explore the stars, but right on time to build superintelligence. And this jump will be faster than previous ones — in 1903 no human had achieved sustained powered flight, and 60 years later we landed on the moon; today the accelerating technology lives in software, so it compounds faster still.

The chain is simple: more science → more technology → more growth → more human flourishing. Hence goal number one is accelerating scientific discovery — the subject of his book, The Eureka Machine, out in September.

The four pillars and the agent swarm (~00:13–00:16)

1. Existing public knowledge, now largely encoded in LLMs. He addresses the copyright discomfort head-on: once these models are open source, they become a resource for humanity in the way the internet did.
2. Scientific measurement data, because human senses are severely limited and the ceiling on perceptual intelligence is unlocked by instruments, not eyes.
3. Simulation — "everything you can simulate, anything you can essentially verify, AI will obviously solve." He was never surprised that chess and Go fell: simulate, verify, generate unlimited training data.
4. Robotic process automation of physical labs, for everything that can't yet be simulated.

Above the pillars sits an agent swarm that ideally carries the equivalent of ~5,000 distinct PhDs before you ever ask it to run a physical experiment.

Recursive's angle: if you're automating science, the best science to start with is the science of AI itself. That means letting the system build a form of self-awareness about its own shortcomings across the whole pipeline, from pre-training through harnesses.

Auto research, measured: three benchmarks (~00:16–00:19)

He draws a clean line first: asking one AI to improve some other system is auto research, not true recursive self-improvement — but it's a useful milestone on the way.

Science, and AI research, is three steps: ideation, implementation, validation. Implementation keeps improving, validation still costs time, and the live question is how many genuinely novel ideas the system can generate.

They pointed their RSI system at three benchmarks thousands of people had already worked on, often with their own agents:

The point he stresses: this is not hyperparameter tuning. The system produced real inventions — incorporating hash tables into transformers, novel forms of momentum. And it matters commercially: kernel efficiency drives intelligence per dollar, and squeezing 5–10% more out of a multi-billion-dollar data center is a very significant end result.

The ceiling: we've barely started (~00:19–00:21)

He tried to define AI properly, found no good definition, and settled on three principal components: prediction (mathematically equivalent to compression) × actions × goals, from which he derives ten "spaces of intelligence."

Take the simplest one, visual intelligence. Computer vision still mostly lives inside the electromagnetic band of the human eye — and human eyes aren't that good (mantis shrimp have better ones). The actual upper bound isn't binocular vision at all: millions or trillions of sensors, spread as far apart as the light cone still permits them to talk to one central intelligence, seeing down to quantum uncertainty and up to gravitational waves, then fusing all of it across many more layers of object abstraction.

So when people say AI won't go much further: we are literally astronomically far from the upper bounds in many spaces of intelligence — which is exactly what makes the area worth researching.

Quotes

"Our generation was too late to explore the earth, too early to explore the stars, but we're right on time to build superintelligence." (~00:12)

"Everything you can simulate, anything you can essentially verify, AI will obviously solve." (~00:14)

Why board games fell early, and why "simulatable / verifiable" is the load-bearing pillar.

"It's not just like tuning hyperparameters — it makes some truly useful inventions, like incorporating hash tables into transformers." (~00:18)

"We're literally astronomically far away from the upper bounds across many of the different spaces of intelligence." (~00:21)

A direct rebuttal to the "AI is plateauing" story.