Stripping cyber data to handicap models hurts general coding too, warns Lentils80 citing K3.1 Pro
lyraxana · x · 2026-10-24
Lentils80 argues that deliberately removing cybersecurity-related training data from open-weight releases — as shown in a screenshot describing K3.1 Pro's data removal at both mid-training and RL stages — will generalize beyond cyber: security work is largely coding, so the handicap degrades real-world coding tasks and weakens defensive security too, citing the poorly received Opus 4.7 as precedent. He accepts the safety rationale but insists there must be a better approach.
More from Models
- Dev notes model fixed overthinking with fewer reasoning tokens, suspects benchmaxxing — cedric_chee · 2026-10-24
- Public benchmarks show Kimi K3 beating all OpenAI models, including Astra — bindureddy · 2026-10-24
- First Kimi K3.1 Pro benchmarks compare it against Opus 5.5, Sonnet 5.5 and Gemini 4 — arena · 2026-10-24
- Recursive's 40B Flint Hits ~200 tok/s on M3/M5 Ultra via MLX-VLM with Fused Metal Kernels — timshi_ai · 2026-10-24
- Richard Socher launches Flint, an open-weights local coding model hitting 125+ tokens/s on a MacBook — RichardSocher · 2026-10-24
- Anthropic's unannounced "Claude Duet" app spotted testing on iOS at version 1.999999 — kevinakwok · 2026-10-24