Latent Space

Liam Fedus & Ekin Dogus Cubuk on why AI still needs the lab

Liam Fedus & Ekin Dogus Cubuk· Co-founders at Periodic Labs
·~85 min·English·Latent Space
AI CompanyTrainingReasoningAgentsRobotics
TL;DR

The Periodic Labs founders argue intelligence alone can't create new knowledge, so they built autonomous labs that pair AI and simulation with physical experiments and train models on the process of doing science, not just its published results.

01Core Thesis

Intelligence Is Necessary but Not Sufficient

You can't think your way to a discovery — new knowledge exists only once an idea is tested against reality, which is why the lab, not the model, is where truth gets made.

You can't just think your way to a solution. The universe is so complicated that in order to actually push the frontier of knowledge and to make progress, you need to create these conjectures and then actually see whether or not it holds.

— Liam Fedus, Latent Space
Key Insight
The thesis quietly rejects the “scale compute until the model solves science” view. If knowledge only exists once an idea is checked against reality, no amount of pretraining substitutes for an experiment — the lab isn't a chore bolted onto the model, it's the only place truth is manufactured.
02A Different Intelligence

Most Real Intelligence Looks Like Science, Not Code

Math and code are easy because the whole problem is on the page, while science is noisy, under-sampled, and missing context — and most useful intelligence is science-shaped.

a lot of the current improvements focus on math and coding and theoretical computer science because it's easier for LLMs. But in real life, most things that require intelligence are actually more like science.

— Ekin Dogus Cubuk, Latent Space
Key Insight
This reframes the whole AI race. The benchmarks everyone chases — math, competitive coding — are the easy case precisely because the full problem is written down. Betting on science-shaped intelligence is a bet that today's leaderboard measures the wrong thing for most of the real world.
03The Training Bet

Train on the Process, Not the Paper

Their reward signal comes from the lab, not from papers, so the model trains on the full lineage of doing science instead of memorizing published answers.

rather than training on the final output of science, you're training on the process of doing science.

— Liam Fedus, Latent Space
Key Insight
Training on the process closes a quiet cheating loophole: a model that already memorized a published result can reach the right answer without doing the physics, and RL then rewards reasoning that won't generalize. Owning the full lab lineage — data that exists nowhere else — is what keeps the reward signal honest.
04The Real Bottleneck

Trying Things Is Easy; Knowing What You Made Is Hard

Mixing powders to try something is cheap; characterization is the bottleneck — the scarce skill is reading a noisy experiment well enough to decide what to try next.

it's not that hard to mix powders to get to try stuff but if you can't characterize and analyze it and then decide what the next step should be intelligently you don't really benefit much from mixing powders randomly.

— Ekin Dogus Cubuk, Latent Space
Key Insight
It inverts the intuition that running the experiment is the hard part. Mixing powders to try something is cheap; knowing what you actually made is expensive. That's why they put a “140 IQ” on every instrument — the scarce resource is interpretation, so they spend their intelligence budget on reading the experiment, not running it.
05The Data Moat

Negative Results Are the Moat

Materials papers usually report what worked, not what failed, so Periodic's own failed experiments supply the negative examples a classifier needs — data competitors can't easily buy.

this is a particularly bad problem in material science because people usually publish crystals they could synthesize but they usually don't publish if they fail to synthesize a crystal.

— Ekin Dogus Cubuk, Latent Space
Key Insight
The published literature is a survivorship-biased dataset — almost all positives — so anyone training on papers alone is missing half the labels a classifier needs. Running their own failed experiments turns Periodic's biggest cost, the negative results, into the one asset competitors literally cannot purchase.
06Why the Lab Wins

Even Better Models Still Have to Run Experiments

ML is good at what it was trained on, but discovery is what hasn't been, so even a far better model still has to run experiments to reach the frontier.

machine learning is really good at what's been trained on. But scientific discovery is almost by definition what you haven't been trained on. And that's why we're building these labs so that whether open models or closed models can use these labs to tinker with the universe.

— Liam Fedus, Latent Space
Key Insight
The claim doubles as a rebuttal to “inference-time reasoning is all you need.” If that were true, the frontier labs would have frozen pretraining at GPT-4. Compressing knowledge into weights still buys capability — but weights can only compress what reality has already revealed, which is why the lab comes first.
07The North Star

Synthesis Superintelligence Widens the Surface Area for Luck

The hard part isn't imagining new materials but making them, so automating synthesis could compress a career of 30,000 tries into a single month.

we tried the 30,000 the Takito group tried in over his career in a month. We just really increase the surface area for luck.

— Ekin Dogus Cubuk, Latent Space
Key Insight
It relocates the bottleneck to throughput. Magnesium diboride's superconductivity wasn't deduced from theory — the compound had sat on shelves as a precursor for decades, and was found only when someone tried it. Automating the slow synthesis step doesn't out-think the search space; it runs far more attempts, buying what Cubuk calls more surface area for luck.
08The Flywheel

Make Physics Profitable, and the Talent Follows

Prove the science pays, and the people come — just as ChatGPT made CS majors popular, Periodic wants commercial wins to pull talent back into solid-state physics.

our best contribution to solid state physics and science could be if we made these tools and topics profitable.

— Ekin Dogus Cubuk, Latent Space
Key Insight
The flywheel is a talent strategy disguised as a business model. ChatGPT didn't just make money, it made CS cool and pulled a generation into the field. Periodic is betting that commercial proof — not government grants — is what reroutes smart young people back into solid-state physics, the way Bell Labs once kept world-class people doing hands-on work.