Dev reframes LLMs as "universal translators" translating between representations
brandon_xyzw · x · 2026-11-09
A developer argues LLMs are best understood as "universal translators": the core operation is converting between representations — prompts into code, pixels into another representation and back.
He extends the framing to agentic workflows: translation into task lists, where simple callbacks spawn more translators (threads), which call themselves (loops) and stack (hierarchy), each converting one representation into another. He suggests most knowledge work is really learning different "languages" and translating between them — and that natural human language is itself many interacting kinds of language.
More from AGI Musings
- AI has broken science's main evaluation proxy, argues computational biologist — anshulkundaje · 2026-11-09
- Two major cardiology trials fail, a reality check for 'AI will cure all disease' — anshulkundaje · 2026-11-09
- Toby Ord: a small share of impossible tasks would crush the 80% time horizon — tobyordoxford · 2026-11-09
- Learning is compression: a long-form essay on manifold hypothesis and why intelligence is possible at all — burny_tech · 2026-11-09
- 'AI democratizes Jevons paradox itself': efficiency fuels even more usage — cocktailpeanut · 2026-11-09
- Toby Ord: share of impossible tasks could skew METR's AI time horizon — tobyordoxford · 2026-11-09