After a day with Jev: a fast general classifier, not an AGI leap

jiayuan_jy · x · 2026-09-19

After a day testing Jev (launched by ChatGPT co-inventor, claiming 20-200x faster and 40-400x cheaper than LLMs), the author concludes the novelty fades fast: it behaves like a faster general classifier/decision-maker that regular LLMs can replicate, and its world knowledge may be thinner given unknown (likely small) parameter scale.
Good fits: finite action spaces with low latency and formal output guarantees — most notably Computer Use (DOM list → Jev action → new DOM list loop), plus game AI, robotics, and autonomous driving.
Hands-on: swapping parts of his Pi Agent to Jev worked for classification-style tasks like Compaction, but not tool use — each call carries unbounded parameters (e.g., line numbers) that require real inference.
He's building a Poker AI demo pitting Jev against other models; no benchmarks yet — reasoning likely below frontier models, but the speed is striking.

More from coding & agent