Free synthetic data idea: train LLMs to write better formal specs via offensive cyber pressure
jd_pressman · x · 2026-10-28
jdpressman pitches a free synthetic data idea for AI lab employees: have LLMs write bad specs for formal verification, subject those specs to offensive cybersecurity pressure, and train on the remediation into better specs — bootstrapping strong spec-writing capability.
The idea comes out of his debate with Ivan Vendrov: he argues climbing formal verification capability is inseparable from having a capable adversary to grind spec-writing against, making offensive security a natural training pressure.
More from coding & agent
- Unsloth adds autonomous browser use for local LLMs, Qwen3 27B demoed — danielhanchen · 2026-10-28
- C3 AI engineers break down why long-running agents fail: context rot, instruction positioning, cached tokens — C3_AI · 2026-10-28
- Agents now write code faster than humans can review — a Claude Code skill to triage dozens of PRs — KyeGomezB · 2026-10-28
- OpenAI: Codex weekly active users up 14x since start of 2026, France top 10 for API devs — reach_vb · 2026-10-28
- 8 open-source repos to self-host a lean AI agent stack instead of stacking $20/month SaaS — SaiTarunRonanki · 2026-10-28
- LangSmith now tracks agent cache hit rates to catch costly bugs — js_craft_hq · 2026-10-28