rasbt: Jev's breakthrough is real generalization, and the secret sauce is likely in the data
multiply_matrix · x · 2026-09-20
- Sebastian Raschka (rasbt) pushes back on dismissing Jev as "just a classifier": from years of training encoder-style models, he notes such models were usually special-purpose and limited, while Jev actually generalizes well.
- He believes the secret sauce is likely more in the data than in the training algorithm, plus a nice API design on top.
- Former Microsoft/LinkedIn executive Mikhail Parakhin agrees, saying he's been brainstorming for two days on how to collect data correctly so it generalizes as well as Jev does — calling the data part the most interesting aspect of the approach.
More from Models
- Reddit user: GPT 6 reasoning feels worse than 5.6 Terra, like GPT 3 flashbacks — FigInternational4873 · 2026-09-21
- Google launches Gemini 3.8 Flash and Flash Cyber: 54.9% HLE, 2.6x better vuln patches — AlienX__0 · 2026-09-21
- Open-source model Jev called out for not crediting Lava, its 8192-context prior work — nischay_twt · 2026-09-21
- Devin makes SWE-2 free across Cloud, CLI and Desktop until October 8 — HankYeomans · 2026-09-21
- Encoder-style classification gets hot again: one multimodal BERT solved 1,000+ classification tasks — cwolferesearch · 2026-09-21
- Deployers told: benchmarks are a directional signal, not a substitute for your use case — evijit · 2026-09-21