Friday, June 5, 2026 · 9:41 AM
ok dumb question: is this paper saying AI can actually discover science?
kinda, but it’s picky about the word “discover”
the paper’s whole move is: discovery means changing the rules of the scientific workspace
changing rules like… better prompts?
nah, deeper than prompts
think of science like a shared Google Sheet
retrieval adds a row. search finds a clever formula. discovery adds a new column everyone now needs
😮ohhhh so it’s not “the AI got an answer”
right. the authors separate retrieval, search, and discovery
retrieval: find an artifact that already fits
search: explore new paths inside the same schema
and discovery is “new schema”?
yep. new artifact types, new operations, new verifiers, new variables
the paper calls that a verified regime transition
by gates and audit trails, basically
if the system edits its worldview, it has to preserve old artifacts and show what new stuff wasn’t just re-labeled old stuff
that sounds very category theory-coded
extremely. category theory is the bookkeeping language here
schemas are categories. artifact states are copresheaves. provenance is tracked through the category of elements
i understood like 46% of those nouns
fair
plain version: every claim, dataset, model, test, and workflow step has a type and a receipt
so when the AI changes the map, you can replay how it got there
wait what’s the counterintuitive part
the paper says scaling inside a fixed vocabulary is not the same as discovery
a bigger model can search harder and still never add the missing column
i thought “more search” was basically discovery eventually
that’s the thing they’re pushing against
search can be brilliant and still stay inside the old types
discovery starts when the system can justify a new type, tool, verifier, or variable
do they actually instantiate it or is this pure math fog
they give 2 materials-science cases
first: Builder/Breaker, a protein-mechanics world model that revises itself under an MDL gate
MDL as in minimum description length?
yep. “does this new law explain the evidence without becoming a messy cheat code?”
the accepted law is about mode-conditioned compliance in protein chains
say that like i’m holding coffee
protein flexibility depends on all the ways it can bend, but especially on slow collective motions
the model didn’t just tweak a number. it changed what variable mattered
exactly
second case: CategoryScienceClaw, which turns skills, artifacts, needs, gates, stress tests, and public discourse into a typed graph
public discourse is in the graph too?
yeah. comments, reactions, open needs, accepted models, rejected models, all become inspectable objects
for a fiber-network example, they prefer an orientation-tensor stiffness model over a simpler isotropic fiber-count one
because the direction of fibers matters, not just “how many fibers”?
that’s the read. and they record the rejected alternative, AIC gate, and perturbation tests
so the audit trail is part of the science, not admin junk
yep. the receipts are load-bearing
if the system revises itself, the proof of revision has to travel with it
what should i take away from this
don’t ask “can the AI answer the question?” first
ask: what schema is it operating inside, what can it change, and what gate accepts the change?
so for AI science tools, logs and validators are basically first-class citizens
100%
you want typed artifacts, preserved provenance, rejected hypotheses, stress tests, and explicit verifiers
otherwise “self-revising” can just mean vibes with a lab coat
very fair. powerful idea, but the paper is a framework, not proof that every AI scientist now discovers real laws
use it as a checklist for whether an agent can responsibly change the map
got it. new columns need receipts
perfect summary. go forth and distrust fancy search bars
Read Fri, Jun 5 · 9:59 AM