The reading room
Goodfire's Eric Bigelow argues in-context learning covers nearly everything a model does to adapt as it reads. He expects context, memory and harnesses, not per-user fine-tuning, to drive personalization, but hasn't researched continual learning.
Goodfire researcher Eric Bigelow resampled a GPT-3.5 model's reasoning at every token and saw answers collapse at single tokens, some arbitrary. He argues the "decision" happens in sampling; newer reasoning models often fork more smoothly.
Goodfire researcher Eric Bigelow said training that rewards only final answers gives models no reason to keep their reasoning readable. He said his optimism about monitoring that reasoning has faded sharply over about a year.
Google DeepMind's Pushmeet Kohli said scientists need models with calibrated confidence and known limits more than human-readable inner workings. Biohub's Sal Candido said protein models hold knowledge not yet unlocked.
Biohub's Sal Candido and Google DeepMind's Pushmeet Kohli argued on a panel that scaling laws in biology must be found, not assumed, and that the problem should decide whether to build better models or gather new data.
Google DeepMind's Pushmeet Kohli said AlphaFold reproduces structures deposited in the Protein Data Bank, not proteins' shifting shapes. He sees raw cryo-EM images as one possible route to that motion, though his own attempt needs more work.
Liam Fedus said Periodic Labs is taking AI tools built for its own materials research to semiconductor partners through onsite engineers. He said the company could later charge for reaching goals, not only for assisting researchers.
On Latent Space, Periodic Labs' Liam Fedus and Ekin Doğuş Çubuk discussed why experiments remain essential in materials discovery. The conversation covered how the standard simulation method, density functional theory, usually models perfect crystals and can't easily predict properties like superconducting temperature.
Periodic Labs' Liam Fedus says the startup puts AI on lab instruments to capture richer data and automates bottlenecks one at a time. He calls full autonomy a non-goal and says humanoid robots would be a slower path to its aims.
Periodic Labs' Liam Fedus says its records of failed experiments, and of the whole process behind each result, give its AI data that published research lacks. He argues even stronger future models will still need physical experiments.
Liam Fedus of Periodic Labs says rewarding AI for finding a superconductor is too slow and noisy. The startup instead rewards it for identifying what X-ray data show a lab actually made, though ambiguity remains.
Palisade Research's Jeffrey Ladish argues that making superintelligent AI care about people is a hard but possible scientific problem that nobody yet knows how to solve. Host Steven Bartlett counters that humans can't even align each other.
Browse by tag