Open model Inkling 2 Small repairs broken networks from raw weights in ~40 tokens
SonglinYang4 · x · 2026-11-14
Thinky Machines demonstrated "weight-space programming": its open model Inkling 2 Small, post-trained on Tinker, can repair broken neural networks directly from their raw weights in about 40 tokens, with no reasoning steps.
In comparisons, the untrained model needed a full generation to fix one break; after post-training it both answers 207 correctly and repairs 176 broken networks. The author frames this as an early step toward having models program directly in weight space.
More from Models
- Chart: Which Gemini 4 model each Google subscription tier actually gets — PawelHuryn · 2026-11-14
- Gemini plan-to-model map corrected: paid Pro caps at Argon, Carbon needs Ultra or API — PawelHuryn · 2026-11-14
- Weekly AI recap: Nvidia's six-week slide, Anthropic IPO pricing, Google's four releases — xiaohu · 2026-11-14
- DeepSeek opens its entire app stack via API: search, voice, OCR, ASR, TTS — churchkey · 2026-11-14
- Garry Tan's 52-task eval: Fable 5.2 matches Opus 5.5 on pass rate but costs 2.80x more — garrytan · 2026-11-14
- OpenAI Codex device limit locks user out despite three paid ChatGPT Plus accounts — justscribethought · 2026-11-14