AI researcher fires at Anthropic: LLMs are state machines, anthropomorphism is alchemy
gerardsans · x · 2026-09-19
Developer gerardsans posted a long thread criticizing Anthropic and the broader safety/alignment community for projecting human traits onto "matrix multiplication," calling the AI industry intellectually corrupt with barely any pushback.
His core arguments:
- An automaton is just states and transitions mapping inputs to outputs — a program, a state machine, not a self or a mind.
- Most ML/AI practitioners lack software engineering backgrounds, leading to "alchemy and anthropomorphic fever dreams" — a repeat of the Eliza effect.
- Steering such a machine is not magic; it's next-token prediction over a DAG.
- His one-line conclusion: "Sampling ≠ agency. Soft programs ≠ personhood. Stop projecting a mind onto the sampler."
This is a notable pushback against Anthropic's model-welfare research direction.
Related event: Engineers Argue LLMs Are Just State Machines, Not Minds(2 posts)→
More from AGI Musings
- XBOW's Brendan Dolan-Gavitt: LLMs killed security through obscurity — moyix · 2026-09-19
- Schmidhuber: I published the first concrete RSI algorithms back in 1987 — SchmidhuberAI · 2026-09-19
- NBER field experiment: AI boosts patent lawyers' skill, but gains accrue only to senior lawyers — avicgoldfarb · 2026-09-19
- Philosopher Piccinini: phenomenal consciousness outstrips computation, so AI won't be conscious via compute alone — pwlot · 2026-09-19
- Philosopher Jeff Sebo rebuts Suleyman's 'model welfare' essay: 'No evidence' of AI consciousness is too strong — connoraxiotes · 2026-09-19
- UnpromptedAU security conf: AI makes reverse engineering trivial, killing security through obscurity — moyix · 2026-09-19