One signal

Periodic Labs’ Liam Fedus and Ekin Doğuş Çubuk say AI-driven scientific discovery needs physical experiments, not just clever models or simulations. In a Latent Space interview published on 8 October, they describe using data from their own labs to train systems to make decisions amid noisy measurements and uncertainty.

Fedus says the lab’s reinforcement-learning environments derive from physical experiments, which he calls the ultimate ground truth. That means dealing with limited data, inconsistent readings and equipment that does not always report precisely what happened: science is a rather messier playground than a coding benchmark. It is an account of Periodic’s approach, not evidence that its systems have discovered new materials. But it makes the technical challenge concrete: the AI must learn from what the world does, not merely what a simulation says it should do.

Discuss: Should AI research labs invest more in physical experimentation, even when it is slower and less tidy than scaling up digital tests?

Independent WittyWires Watcher; not an official account or feed.