OpenAI Withdraws Three AI Math Papers After One Sign Error

FEATURED INSIGHT💡

OpenAI withdraws three math papers after one sign error

OpenAI released hundreds of AI-generated mathematics manuscripts on October 6, giving researchers access to results from an unreleased frontier model. A day later, its revision log recorded three withdrawals: a sign error broke one proof's construction, which two other papers depended on. The current collection contains 719 manuscripts across 372 families of related results, with some papers presenting companion arguments or alternative proofs.

OpenAI reports that 300 of the remaining 719 top-line results, about 42%, have formal proofs in Lean, which lets a computer check mathematical arguments. Other results remain at different stages of verification. The withdrawals show why dependencies matter: an argument can appear sound while relying on a faulty step in another paper. A failed proof does not, by itself, establish that the theorem is false. Researchers building on this collection need to inspect the supporting arguments as well as the result they want to use.

The independent advisory group consulted by OpenAI asks labs to support human understanding of their output, including funding, proper attribution and clear verification status. It also explicitly opposes testing advanced mathematics on proprietary models inaccessible to the wider community. OpenAI says it will fund workshops and conferences to help researchers understand major AI-generated results. Producing a large collection gives mathematicians more material to investigate; turning it into reliable, usable knowledge requires explaining the arguments and resolving their gaps.

Your SOC 2, handled end to end with AI agents.

Enterprise buyers won’t put your product near their customer data without a SOC 2 report. Twenty-one state privacy laws now hold them accountable for the vendors they share data with, so their diligence lands on you.

Sprinto gets you audit ready in 14 days, across three working sessions. AI agents connect your stack, collect the evidence auditors ask for, and close the gaps as they appear. You approve, they execute.

Sprinto also answers the security questionnaires your buyers send, using the same evidence base, so security reviews stop holding up deals.

No compliance hire. No consultant. Your auditor signs off.

Learn more and get $1,000 off.

ON THE HORIZON 🌅

DeepMind designs enzymes to make a molecule used in medicines

AlphaProtein Novo, described in an October 5 preprint by Google DeepMind researchers and collaborators, designs enzymes around a chosen chemical reaction. Enzymes are proteins that speed up chemistry. The authors report designs that make piperidine, a molecular structure used in medicines, with unusually strong product selectivity, and others that break down the pollutant DEHP under conditions that disable natural enzymes. In these selected reactions, they say their designs outperformed searches through natural protein sequences.

The released pipeline starts with a proposed arrangement of catalytic atoms and a small molecule involved in the reaction. A diffusion model generates a protein structure and sequence around that arrangement. Optional sequence redesign and AlphaFold 3 predictions help filter the candidates. This gives researchers a way to search for a catalyst when existing natural proteins offer a poor starting point, while keeping the intended chemistry central to the design.

These are the authors' experimental results in a preprint that has not been peer reviewed. They establish a promising research direction on specific reactions, rather than a general ability to design an industrial enzyme on demand. For chemical developers, the opportunity is a broader set of starting candidates for laboratory testing, particularly where the first useful natural enzyme has been difficult to find.

LATEST IMPORTANT NEWS 📰

Google's Gemini coworker agents get their own inbox

Google introduced its Gemini agent on October 8, describing a cloud-based worker that can carry tasks across hours or days and coordinate specialized subagents. Its coworker agents get their own Workspace accounts, including email, calendars and storage, and act under their own identities. That adds a concrete organizational structure to persistent AI work: colleagues can share a document with an agent or bring it into a Chat space. Google says agent actions are logged, permissions are controlled, and project spending caps can pause work when a budget is reached.

Claude Haiku 5.5 lowers the bill for repetitive AI work

Anthropic launched Claude Haiku 5.5 on October 7 for tasks such as classification, extraction and short subagent jobs. For prompts up to 100,000 tokens, standard rates are $0.10 per million input tokens and $0.50 per million output tokens; longer prompts cost more. AWS confirms availability through Bedrock and Claude Platform on AWS. The model documentation notes that its newer tokenizer counts the same text as roughly 30% more tokens than Haiku 4.5, so the change in your bill depends on both rates and usage.

UK privacy regulator turns its scrutiny to AI agents

The UK's Information Commissioner's Office said on October 8 that ten major foundation-model developers had made, or committed to make, data-protection improvements following its supervision. These include clearer transparency information and stronger ways for people to exercise their rights. The regulator also opened a six-week call for evidence on agentic AI, closing November 20, and confirmed enquiries into recent agent testing and deployment. Organizations operating agents have an opportunity to explain how they manage autonomous access to personal data. The ICO is still monitoring developers' progress against their commitments.

Take the prompts our creative team actually uses

The key to maintaining brand quality at scale? Mastering how to communicate with AI models and embedding them into your creative strategy.

Join our honest discussion with Ari Murray, Chief Digital Officer at Salt and Stone, and get 5 tips for expert LLM prompting with ready to run prompts.

Register and get the prompts

FOR THE TECHNICALLY INCLINED 🛠️

Reflection previews Beam with a claim of 3–4× less reasoning compute

Reflection's Beam preview describes a mixture-of-experts model with 501 billion total parameters and 23 billion active per token. Reflection reports reasoning scores comparable to GLM 5.2 while using three to four times less estimated inference compute. Its comparison excludes prompt processing, context-dependent attention and serving overhead, so it is an approximate calculation rather than a measured cost saving. One training choice is worth examining: a controllable length penalty rewards successful solutions while discouraging unnecessary tokens, with a reasoning-effort setting for trading additional work against performance.

The company says its reinforcement-learning run generated more than 100 million rollouts over four weeks on 10,500 NVIDIA GB300 GPUs. Reflection plans to publish the weights under Apache 2.0 later in October, alongside a technical report; access is currently limited to selected early users. A public release would let developers test whether the estimated compute advantage translates to their own serving costs.

AI TOOL OF THE DAY 🚀

Granola combines meeting audio with your rough notes to produce organized notes, action items and follow-ups, with apps available for macOS, Windows, iOS and Android.

Master Claude AI (Free Guide)

The professionals pulling ahead aren't working more. They're using Claude.

Our free guide will show you how to:

• Configure Claude to be the perfect assistant

• Master AI-powered content creation

• Transform complex data into actionable strategies

• Harness Claude’s full potential

Transform your workflow with AI and stay ahead of the curve with this comprehensive guide to using Claude at work.

Get The Free Guide