1. GPT-5.6 Sol Pro Closes a 30-Year Gap in Convex Optimization
Researcher Phillipp Kerger used OpenAI GPT-5.6 Sol Pro to crack a convex optimization problem open for three decades. The model produced a full formal proof, verified in Lean, resolving a quadratic dimension dependence question.
After 148 minutes with a 10-page CDC-style prompt, Sol Pro returned a complete proof. Kerger said he would never have gotten this result on his own. The story hit 543 points on HN with fierce debate about human vs AI credit.
What This Means
- AI proofs verified in Lean are now a reality
- CDC-style prompting could become standard for theorem proving
- Expect an explosion of similar results
2. Claude Fable 5 vs GPT-5.6 Sol: The NP-Hard Showdown
Charles Azam benchmarked Claude Fable 5 against GPT-5.6 Sol on an NP-hard network design problem. Fable 5 was the overall winner with remarkable consistency. The /goal command won 4 of 6 trials but made averages worse due to catastrophic regressions.
Key Takeaways
- Fable 5 plain mean: 32,386 with tight variance
- Sol plain mean: 34,261 with wider spread
- /goal is not a generic try harder switch
3. AI Mania and the Evisceration of Decision-Making
A viral essay argues organizations are deferring strategic judgment to LLMs without understanding limitations, accelerating as models get more fluent.
Comments