Kushal Chakrabarti

(trending #3 on alphaXiv) Why does CLAUDE.md just keep growing?

Appending to CLAUDE.md costs O(1); deleting without risking a regression costs O(2ⁿ), so it grows without bound.

The fix? (Prompt) comments! -99.3% prompt bloat, +23.1% instruction-following.

Midway through NeurIPS rebuttals, I'm seeing a new failure mode: authors using reviewers as verifiers.

The pattern: a provocative thesis, N claims whose evidence just doesn't support them. Scrutinize any one and there's no defense — just an instant, gracious concession, plus N′ fresh claims in the reply.

Post-Fable, generating a plausible claim is (basically) free; verifying one still costs a reviewer-hour. Add in A* variance and a rational strategy falls out: Don't write claims that survive verification; sample more claims than the verifier's budget covers and let selection effects do the rest.

That is, peer review just became RLVR — and the verifier is a human with six papers, a day job, and a rebuttal window measured in days. Concession is a penalty that costs nothing to absorb; resampling costs nothing to run. Weak verifiers don't get satisfied; they get reward-hacked.

Is anyone else seeing anything like this?

#ICML hot-take #3: publish-or-perish incentives now also apply to industry (+ academia of course) -> simplest + sufficient explanation of previous two

#ICML hot-take #2 (from fancy VC event last night): increasing consensus that even papers w/signal are becoming “safe and boring" (AKA incremental).

#ICML hot-take #1 (flying back from Korea RN): the signal is growing exponentially, but the noise is growing super-exponentially.