One forward pass is one token — the lock-in happens in the autoregressive loop
gerardsans · x · 2026-09-09
Gerard Sans untangles a common conflation: a single forward pass (one token, residual stream carry-over within that pass) is a different level from the autoregressive loop. The lock-in shows up in the loop — each emitted token feeds back as the next prior, so repeating within a session gives the original task 'mass', and later instructions arrive after the truck is already moving. His summary: the pass computes, the loop commits.
Related event: 'Momentum Prior': Agents Locked Onto Initial Tasks Resist New Instructions(2 posts)→
More from Research
- ValsAI Launches Tax Agent Bench: 391 Expert Questions to Test LLMs on Corporate Tax — JenniferHli · 2026-09-09
- AI Intern reproduces ICML best paper "How much do LLMs memorize" for under $8 — _akhaliq · 2026-09-09
- Buckmaster & Alpöge publish Euler (112pp), Boussinesq (76pp), porous media papers with Lean proofs — xennygrimmato_ · 2026-09-09
- Boltz Lab is hiring as partners weekly solve problems traditional methods failed — GabriCorso · 2026-09-09
- Boltz creator Gabriele Corso named to MIT TR's 2026 Innovators Under 35 as models near the clinic — GabriCorso · 2026-09-09
- Boltz models keep improving as partners solve previously failed drug-discovery challenges weekly — GabriCorso · 2026-09-09