ianand's thread: Jev, an Encoder-based AI model from a parallel universe
ianand · x · 2026-09-20
ianand published a thread imagining 'Jev, an AI model from a parallel universe':
- History: the 2017 Transformer did one job (English-to-French translation) with an Encoder analyzing text and a Decoder producing output.
- Our universe: GPT-2 took the decoder half and showed 'all you really need' is the decoder, a single model for multiple tasks, prompting instead of task fine-tuning, and scale — giving rise to ChatGPT, Claude, Gemini and all popular LLMs.
- The parallel universe: no GPT-2 moment; the same innovation went into the Encoder, producing Jev.
- Jev's traits: faster, more predictable, more reliable — but can't code or chat, only make decisions. A 'decision AI' rather than generative AI.
- Key takeaway: Decoder self-supervised training scales easier, so maybe we were destined to build ChatGPT first; but Jev gives the Encoder a ChatGPT-equivalent product surface, and devs are rediscovering use cases for this 'decision AI'.
Related event: ianand's Thought Experiment: An AI World Built on the Encoder(9 posts)→
More from AGI Musings
- Fleuret: if AI designs drugs as fast as it makes dumb videos, regulation will buckle — francoisfleuret · 2026-11-12
- Newsweek op-ed: US should fund hospitals and universities in federated AI learning — Saul_Loveman · 2026-11-12
- 10-line deterministic polynomial-time primality test relies on OpenAI's recent quasi-GRH result, making AKS obsolete — fortnow · 2026-11-11
- AI2027 Author Kokotajlo Shifts AGI Median Timeline to Around 2030 — JacquesThibs · 2026-11-11
- Ron Bodkin: automated research taste advances faster than expected, but the evidence is weak — ronbodkin · 2026-11-11
- Should devs still hand-write code? Author proposes 1,000-hour baseline before AI agents — bendee983 · 2026-11-11