Dev says he built non-autoregressive decision models a year before frontier lab's Jev
JFPuget · x · 2026-09-20
Developer Nandakishor claims he published an arXiv paper on non-autoregressive decision models (arXiv:2503.23303) back in March 2025, open-sourced model weights on Hugging Face along with a SaaS sales-conversations dataset and a PyPI package, and built a sub-40ms open-weight System 1 decision engine (RLCD) guided by reinforcement learning (PPO over sequence representations) that outputs turn-by-turn conversion probabilities. A second paper (arXiv:2510.01237) formalized the schema-based decision + RL framework in September 2025.
He then watched TypeSafe AI — a frontier lab co-founded by ChatGPT co-inventor Diogo Almeida — launch Jev in September 2026, presenting the same non-autoregressive decision concept as a breakthrough, but with no papers, no open weights, and no open training data. The author says the moment feels both validating and frustrating, and notes Jev generalizes the parallel-sampling approach he pioneered.
More from Research
- Google's Physical-Engineering's Last Exam: all 14 frontier models score below 50% — dl_weekly · 2026-11-12
- CypherGuard fixes structurally wrong LLM Cypher queries, hitting 88.5% on CypherBench — JeremyCMorgan · 2026-11-12
- PufferLib 5.0 trains a Rubik's Cube solver at 8.8M SPS on a single 12GB GPU — haydendevs · 2026-11-11
- 10-line deterministic polynomial-time primality test relies on OpenAI's recent quasi-GRH result, making AKS obsolete — fortnow · 2026-11-11
- WSP Weight Edits Shown Equivalent to Constrained Activation Steering on Qwen — kastnerkyle · 2026-11-11
- codec.audio benchmarks neural audio codecs on noisy in-the-wild speech; MOSS tops — kastnerkyle · 2026-11-11