中文

AI FOR SCIENCE · OPEN QUESTIONS

Questions I’m tracking

Questions the site follows over time. Each has a current answer that is revised as evidence changes.

5 questions

  1. CLOSED LOOP · PHYSICAL WORLD

    Can AI move from proposing candidates to closed-loop learning in the physical world?

    Radical AI · Lila Sciences · DP Technology

    Standardized workflows can improve throughput and comparability. Few public systems trace how an observation changes the next decision and show that the change helps.

  2. REPRESENTATION · INFORMATION FIDELITY

    Is information lost during measurement, extraction, or encoding a hard ceiling on scientific AI, or can scale buy it back?

    MinerU.Chem · OCSR · Smirk · AlphaGenome Atlas

    Scale can improve learning from common patterns and lower inference and search costs, but it cannot reliably reconstruct information that was deleted, collapsed or never measured. MinerU.Chem, OCSR and Smirk expose consecutive representation gates; AlphaGenome Atlas shows that nine billion precomputations also propagate the base model's biological blind spots across the map.

  3. AI SCIENTIST · EXPERIMENT ECONOMICS

    Does the AI Scientist win through intelligence, or a new economics of experimentation?

    Radical AI · Lila Sciences · FutureHouse

    Adaptive experiment selection can work in narrow physical campaigns. Whether it becomes a reliable, economical and transferable operating system remains unresolved.

  4. DISCOVERY · DEPLOYMENT

    Does faster discovery actually produce faster deployment?

    Radical AI · Recursion · Isomorphic Labs · Microsoft

    Local discovery and validation cycles can be compressed, but public evidence does not yet show deployment accelerating proportionally. As a candidate approaches real use, timelines are increasingly governed by manufacturability, batch consistency, safety, qualification, workflow integration and organizational decisions rather than model inference.

  5. DATA · DURABLE MOAT

    Can experimental data become a more durable moat than the model?

    Radical AI · Recursion · Lila Sciences

    Current evidence supports the scientific and operational value of high-quality experimental data, but does not yet show that the data itself forms a durable commercial moat. Replicability, exclusivity and attributable impact on final outcomes still need separate validation.