← @bgSponge

Claude-shaped science

Summary

Matthew Schwartz says current language models are good at science but are not scientists, so he stopped treating Claude as the collaborator he wanted and looked for problems that fit it.

  1. The mismatch is the method.

    Models know a lot and code fast, but Schwartz says they still need hand-holding to produce work a scientist would value.

  2. Hunt Claude-shaped problems.

    BootLoops is his harness for checkable quantitative work, starting with the semi-numerical S-matrix bootstrap rather than deep conceptual questions.

    The current generation of AI tools is not capable of solving most problems in science.

  3. Correct is not the same as interesting.

    Outside his field, experts often shrugged at the first result and then redirected the question.

  4. The loop does not collapse.

    Data still has to be acquired and checked. He sees no reason to replace the scientific method, and he wants humans credited for direction rather than typing.

    I don’t see any evidence or need to revisit the scientific method.

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →