Researcher: LLMs moved faster than I expected, but tool use isn't a new paradigm
artetxem · x · 2026-09-30
Researcher Mikel Artetxe reflects: five years ago he believed in LLMs, yet didn't expect today's progress without a major paradigm shift — "I was wrong." But he pushes back on formerly vocal skeptics now claiming vindication because "today's models do far more than predict the next token."
Coding agents run terminal commands by predicting the next token, he notes; no one claimed LLMs would solve everything without environmental interaction. The layer on top is thin, and tool use isn't a new paradigm — "Giving BERT a Calculator" appeared seven years ago. Science means updating beliefs on evidence; admitting your view changed makes you a better scientist, while insisting "I was right all along" sets a terrible example for the public and young scientists.
More from AGI Musings
- AI's sneakier threat to work isn't replacement — it's forcing everyone to become an AI manager — random_walker · 2026-09-30
- Longevity Influencer: We're the Last Generation to Watch Our Parents Grow Old — rand_longevity · 2026-09-30
- Kevin Xu on Chinese open source: how China stumbled into open weights, from Linux floppies to DeepSeek — kevinsxu · 2026-09-30
- evilsocket: Non-Transformer 'Old School' Models Are Massively Under-Explored — evilsocket · 2026-09-30
- Allie Miller: Ultrafast AI Will Turn Meetings From Discussion Into Instant Decisions — alliekmiller · 2026-09-30
- From deer videos to a 4D time machine: the 'spatial memory palace' for physical reality — bilawalsidhu · 2026-09-30