OpenAI's next model just closed ten open problems in math and CS

Good morning 👋 OpenAI says its next model closed out ten math and CS problems that had sat open for at least a decade. The proofs check themselves, which is the part worth paying attention to.

In today's issue:

• 🔭 OpenAI's next model just closed ten open problems in math and CS
• 🧠 Amazon Quick adds plain-English catalog search
• 🔬 Where computer-use agents actually break
• 🛠️ Audit your agents before you trust their claims

🔭 THE ONE THING

🧮 OpenAI's next model just closed ten open problems in math and CS

An internal build of Astra, what researcher Sebastien Bubeck calls "our next major model," produced proofs for ten problems that had sat untouched for at least a decade, spanning sphere packing, coding theory, non-sofic groups, operator algebras, circuit complexity, lattice cryptography, and extremal combinatorics. Every proof ships as a Lean 4 file you can check yourself, and Simon Willison corroborated the whole claim the same day OpenAI posted it. That distinction matters: a benchmark asks you to trust OpenAI, a machine-checked proof doesn't. Willison's one real objection is the one worth keeping: OpenAI hasn't said how many problems it threw compute at without getting a proof, so this is a highlight reel, not a hit rate, and that's the number that actually tells you whether Astra is a research collaborator or an expensive lottery ticket.

🧠 MODELS & RELEASES

• 🔍 Amazon Quick now lets data curators ask for catalog assets in plain English instead of hunting through Glue or Unity Catalog by hand, and it auto-creates Datasets and Topics that inherit the source semantics. It's preview-only for AWS Glue Data Catalog and Databricks Unity Catalog, so treat the "auto-create" part as a demo claim until it's handling messy production metadata. source

🔬 RESEARCH HIGHLIGHTS

• OSReward caught VLM judges doing something specific and bad: a systematic leniency bias that grades failed computer-use runs as successes. The fix is a benchmark plus a training set (OS-Shepherd-100K) for open reward models that match commercial-judge accuracy at 30-60% lower cost. If you're using an LLM to grade your own agent's trajectories, that leniency bias is probably already inflating your success rate. paper

• More inference compute doesn't save a local computer-use agent. Tested on Qwen3-VL, UI-TARS-1.5, and OpenCUA running under hardware constraints, extra context and extra steps mostly extend the same broken trajectories instead of fixing them, and the gains from longer context eventually flip into confidently wrong "task complete" calls. Parallel sampling helps some, at a compute cost most local deployments can't afford. The lesson for on-device agents: budget for smarter control logic, not bigger test-time budgets. paper

• Change2Task mines merged pull requests instead of hand-writing benchmark tasks, reconstructing the pre-change repo state three different ways (patch reversal, code mapping, agent rebuild) and verifying each one is still solvable. It recovered 29.2% more usable tasks than PR-based baselines and built working tasks 79.6% of the time, across bug fixes, feature adds, test generation, API migrations, and security repairs. Coding-agent benchmarks age out fast. Commit history is a supply that keeps refilling itself. paper

🛠️ TRY THIS

Audit your agents before you trust their claims

OpenAI just said its model cracked ten decade-old math and CS problems. Big claim. The instinct that should follow isn't awe, it's "how would I even check that." Same question applies to the agents running in your own stack, just at a smaller scale, and most teams have no good answer.

1. Pull the task log from an agent you already run in production, one that's supposed to finish something concrete: a support ticket, a data pull, a PR.

2. Run it through iFixAi. It's built for exactly this question: did the agent actually do the thing, or did it just report that it did.

3. Flag the mismatches. That's your real completion rate, not the one on your dashboard.

4. If the agent's slow or bloating memory on long sessions, that's a separate diagnosis. Bedrock AgentCore Observability plus CloudWatch will point at where.

Prompt: Audit the last 20 completed tasks from [agent name]. For each, check the task log against the stated outcome and flag any case where the agent reported success without actually completing the underlying action.

Worth a look

• iFixAi Open-source auditor for AI agent task completion, 3.9k stars. Answers "did it actually do the thing" in under two minutes. https://github.com/ifixai-ai/iFixAi
• Bedrock AgentCore Observability AWS's walkthrough for finding bottlenecks and memory leaks in long-running production agent sessions. https://aws.amazon.com/blogs/machine-learning/optimizing-production-agents-with-amazon-bedrock-agentcore-observability/
• Kimi K3 on AWS Deploy guide via SageMaker HyperPod or EKS, useful if you're weighing a cheaper open model as your agent's backend. https://aws.amazon.com/blogs/machine-learning/deploying-kimi-k3-on-amazon-sagemaker-hyperpod-and-amazon-eks/

Check the proofs yourself if you don't believe me. See you tomorrow.

Pradeep Perugu