Hacker News — 2026-07-20#

Top Story#

Human mathematicians are getting “outcounterexampled” by AI models. Recent breakthroughs by LLMs like Claude Fable and ChatGPT Sol have systematically dismantled decades-old math problems, including disproving the 100-year-old Jacobian Conjecture and solving questions posed by Erdős and Grothendieck. This marks a massive paradigm shift in formalization, where AI is auto-generating thousands of lines of verifiable Lean code to definitively solve open problems.

Front Page Highlights#

Stop Using OpenCode · wren.wtf This scathing teardown exposes the wildly popular open-source AI coding agent “OpenCode” as a security disaster masquerading as a productivity tool. The author highlights catastrophic flaws ranging from completely broken path validation for shell redirections to default-open HTTP servers that allow arbitrary RCEs. It is a stark reminder that slapping an LLM onto a stateful bash terminal without proper sandboxing is a recipe for having your machine compromised.