AI4SLEARNING LIBRARY

A Latent Space interview with Radical AI CEO Joseph Krause on the unique challenges AI faces in structural-alloy research—and how self-driving labs could shorten the long journey from discovery to production.

• Multidimensional physical constraints rule out one-click generation：Material performance depends not only on elemental composition, but also on microstructure, processing routes such as casting and 3D printing, and real-world supply-chain and cost constraints.
• Close the loop through physical experiments：Move beyond human-planned semi-automation so AI agents can run the full high-throughput loop: generate hypotheses, direct robotic synthesis and characterization, then revise from experimental feedback.
• The real moat is experimental data, not the model：As foundation models trend toward open source, the durable commercial advantage is the ability to continuously generate scarce, real-world closed-loop experimental data with automated physical infrastructure.

A Latent Space interview with Lila Sciences CTO Andy Beam and Chief Science Officer Rafa Gomez-Bombarelli on using real physical laboratories as reinforcement-learning verifiers to address AI for Science's shortage of high-quality data.

• Scientific data scarcity and a new scaling law：Large language models have consumed much of the internet's human text—what Andy calls AI's fossil fuel. In biomedicine and materials science, high-quality real-world data is far scarcer, so Lila treats experimental data generation as a new scaling dimension.
• An AI science factory：Lila positions itself neither as a traditional biotech nor merely a lab-automation company. Its goal is a model-controlled experimental platform whose instruments share a physical transport layer analogous to a computer's PCI bus, including 96-well plates moving on magnetic tracks.
• Closed-loop RL: nature as the verifier：Code and mathematics offer clear verification rules. For scientific discovery, high-throughput physical experiment results can become the reward signal: AI proposes a hypothesis, writes experimental code, the lab executes it, and both successes and failures feed back into the model.
• General scientific reasoning across domains：Rather than building a separate vertical model for protein folding or battery materials, Lila says it has assembled roughly ten trillion experimentally verified reasoning tokens across biology, chemistry and materials for foundation-model post-training; the scale and transfer gains remain company-reported.

An interview by Emergence Lab with Stanford professor James Zou on how multi-agent systems are reshaping scientific discovery and their latest work using virtual labs to simulate new-drug development.

• The social dynamics of AI agents：Virtual labs assign roles such as AI professor and AI student. Who speaks first and the personalities assigned to each agent can change the trajectory of scientific reasoning and the novelty of the final proposal.
• Running a ten-thousand-person pharma company in the cloud：The team built a virtual biotech company in which tens of thousands of agents simulate cross-functional work from target discovery and molecular design to clinical research. It produced a lung-cancer ADC strategy closely aligned with Merck's independent experimental results.
• Human advantage shifts toward scientific taste：As foundation models advance, researchers may be able to create millions of AI scientist agents. Zou argues that the scientist's durable advantage will shift from basic execution to taste and judgment in choosing scientific questions.