We’re designing and testing agents for mathematical discovery.
Some lessons learned:
1️⃣ Don’t aim for “Riemann Hypothesis–level” problems yet — the mathematical language and tools to solve them probably don’t exist yet.
2️⃣ Learn to break big problems into small ones. Even if they look unrelated, follow your intuition — the hidden links between math branches are often surprising.
3️⃣ Don’t chase proofs directly. Lean is for proving, not for discovering. Finding a good problem is more valuable than solving one.
4️⃣ A good conjecture should be falsifiable. Use code to test its possible truth, not to prove it. Experimentation > formalism.
@jietang congrats!👋 any models in the works to rival gpt-5.6 sol pro and fable in math and science? also, will the coding plan work with the deepseek harness?
Introducing GLM-5.3: Built to Code. Ready for Cyber Defense.
- Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model
- A major leap in cybersecurity, setting a new standard among open models
Tech Blog: https://t.co/ekQkO83jCv