Honnibal: LLM weight-editing paper abstract reads like stage magic — each result uses a different method
honnibal · x · 2026-11-03
Explosion (spacy) founder Matthew Honnibal argues research abstracts now resemble stage magic: readers assume all experiments share one method, but each result uses different tricks. Citing an abstract claiming "LLMs can natively read & write neural network weights" (identify a network's purpose from raw weights, write small game-playing nets from scratch, fix Qwen3.5-0.8B's math by editing 0.016% of weights), he notes "just by looking at the raw weights" is narrowly scoped to one result while another requires actually running the network. By the academic standards he trained under, he would reject such an abstract — and he wonders if that makes him outdated.
Related event: spaCy founder dissects LLM weight read-write research(2 posts)→
More from AGI Musings
- "People who build things with AI will never exceed 10%": the metacognition argument sparks debate — IndraVahan · 2026-11-03
- AI safety researcher: technical 'AI Safety' work has been product work empowering big labs — j_foerst · 2026-11-03
- Alan's AGI Countdown Shows Just 1% Remaining, Sparking Reddit Debate — deles_dota · 2026-11-03
- HCI researcher warns AI use is addicting like TikTok, EMNLP shows coming deskilling — IanArawjo · 2026-11-03
- Claude Code team keeps a framed 'Bitter Lesson' on the wall: 'never bet against the model' — Hesamation · 2026-11-03
- Ben: AI should make 99% of people economic actors, not just hedge funds — LinusEkenstam · 2026-11-03