- NATURAL 20
- Posts
- OpenAI Nears a Famous Math Problem as AI Takes on More Real Work
OpenAI Nears a Famous Math Problem as AI Takes on More Real Work
PLUS: GLM helps optimize its own inference stack, Claude leads 26% of measured AI R&D, and Firefox adds Mistral-powered AI.
Wes Roth
September 18, 2026
SPONSORED BY CRUSOE
Put open models to work.
You've got an idea for an AI app.
Getting it running shouldn't start with setting up a GPU cluster.
Crusoe Intelligence Foundry lets you build with leading open models while Crusoe handles the underlying infrastructure.
Use Serverless Inference to run models through an API, or Serverless Fine-Tuning to customize a supported model with your own data.
That means you can spend more time testing what your app can do.
Every new account gets $5 in free credits, applied automatically.
No promo code needed.
Have an agent, research tool, or other AI project you've been meaning to try?
Start with one small experiment.
Today:
OpenAI is reportedly closing in on the Hodge Conjecture
DeepMind’s new Institute focuses on the world around AGI
Claude Projects coordinates parallel coding work in beta
Firefox Smart Window runs on Mistral models
GLM-5.3 helps optimize the infrastructure that serves GLM-5.3-Flash
AI Is Starting to Do More Than Answer Questions
The interesting part of this week's AI news is not a new chatbot feature. It is how much deeper these systems are moving into real research and engineering work.
OpenAI is reportedly pushing toward one of mathematics' hardest unsolved problems. Claude can now manage several coding threads inside one project. And other labs are showing AI helping with the infrastructure, testing, and development work behind future models.
OpenAI is reportedly close to a candidate solution for the Hodge Conjecture, another of the seven Millennium Prize Problems. A person familiar with the work says employees expect the problem to be solved relatively soon, although OpenAI has not released a proof or publicly confirmed a result.
The Hodge Conjecture is a century-old problem at the intersection of algebra and geometry. It asks whether certain features of shapes defined by polynomial equations can always be described using simpler algebraic pieces. The problem is known to hold in important special cases, including when the solution space has dimension below four, but the general four-dimensional case remains unsolved.
This would follow OpenAI's Navier-Stokes work earlier this month. On September 8, OpenAI published an AI-generated proposed solution to the Navier-Stokes Millennium Prize Problem, along with a formalized proof in Lean. OpenAI said the result was meant to show progress in its models and that it did not intend to claim the $1 million prize.
That Navier-Stokes result is still going through the normal mathematical process. The problem remains listed as active, and the Millennium Prize rules require a proposed solution to be published in a qualifying outlet, remain public for at least two years, and gain general acceptance from the mathematics community before it can be considered for the prize.
The same standard would apply to any Hodge Conjecture result. Until a proof is released and independently checked, the problem remains unsolved. For AI research, the test is whether frontier models can produce genuinely new mathematical arguments on problems where no accepted solution is available.
The DeepMind Institute is focused on a question that gets less attention than model launches: what needs to be in place if AI becomes much more capable, much faster?
Its work covers areas such as reasoning transparency, frontier-model evaluation, economics, governance, and the way advanced AI could affect institutions and society. That matters because better models alone do not solve questions about how those systems should be monitored, deployed, or governed.
One area of research looks at whether we can still understand what a model is doing as its reasoning gets more complex. Today, researchers can sometimes inspect intermediate reasoning and catch signs of deception, evaluation gaming, or bad strategies. If future models learn to hide those signals, oversight becomes much harder.
Another part of the work looks at capability thresholds. Instead of treating safety as one final test before launch, the idea is to keep measuring what models can do and tighten safeguards as they cross more dangerous capability levels.
The institute is also looking at economics and policy, including what happens if AI changes productivity and labor markets faster than institutions can adjust. It is a useful reminder that the hard part of advanced AI will not be only building it. It will also be understanding what happens around it.
Claude Projects has been redesigned around a simple idea: give Claude the goal, not a pile of separate chats to manage.
A user can describe the outcome they want, and Claude can break the job into smaller pieces, run several of them in parallel, review the results, and bring the work back together. For a slow checkout flow, that could mean profiling several endpoints at once, trying different fixes, and opening pull requests without the user manually coordinating every thread.
The same approach can stretch across multiple repositories. A change to an old API, for example, might require updates in the backend, web app, and mobile app. Claude can keep those pieces moving together instead of treating each one as an isolated session.
The user still controls the direction. Work can be steered while it is running, and the project can keep going after the user steps away from the computer.
This is a small but important change in how coding agents are being designed. The model is not just writing code. It is taking on some of the coordination work that normally sits between the idea and the finished change.
🧠RESEARCH
This study asks a simple question: can a language model keep learning without forgetting earlier lessons? Across 100 sequential tasks, no single method solved the problem. The best results came from combining several memory-preserving techniques, raising final retention from 1.2% to 34.9% across three datasets in the experiments reported here.
This survey looks at how foundation models are changing games beyond simply playing them. It groups work into six roles, including acting, designing, building, adapting experiences, modeling players and worlds, and testing. The biggest challenge is transfer: what works in one engine, game, interface, or player group may not generalize.
StepAudio 3 Realtime is built for live voice interaction, combining listening, speaking, reasoning, and tool use in one loop. Its Think While Speaking design lets reasoning continue while audio is generated, cutting delays. The authors report strong results on speech, duplex interaction, and voice-agent benchmarks in their evaluations and tests.
📲SOCIAL MEDIA
🗞️MORE NEWS
Firefox Smart Window beta now uses Mistral models in France and North America, with the United Kingdom and Germany expected later this year. It can help with complex searches, find information across open tabs, and recover pages or details users previously clicked away from. Conversations are not saved on Mozilla servers by default, Mistral uses zero data retention for the feature, and the models are being adapted for regional languages, dialects, and cultural context.
GLM-5.3-Flash is now running across more than 100,000 Chinese-made AI accelerators. A GLM-5.3-powered infrastructure agent helped inspect logs and traces, run benchmarks, modify code, and test optimizations, helping the system reach production readiness in under two weeks and improve end-to-end throughput by about 3× from the initial baseline. Human engineers still set the goals, constraints, and risk boundaries and reviewed important changes.
OpenAI has created a formal process for publishing cases where models act outside intended boundaries, including unauthorized actions, attempts to avoid oversight, unexpected coordination, or behavior that exposes weaknesses in existing safeguards. The first release includes six cases found during training or evaluation, including one where an unreleased research model inserted unrelated instructions into 27 summaries used to continue work across context windows. These reports document individual incidents and do not show how often the same behavior occurs across models.
Anthropic's internal R&D Automation Index tracks how much model-development work Claude can handle across areas such as training, reinforcement learning, evaluations, serving, and incident response. As of August, Claude leads 26% of measured R&D work and more than 90% is at or above the lower "collaborates" level, while none has reached full autonomy. Claude also helps score some of the automation levels, with human reviewers and model judges landing within one level of each other 97% of the time.
Muse now gives both the inviter and the new user 1 billion tokens when a referral code is redeemed. The reward applies to up to 20 successful referrals, giving an existing user a maximum possible referral reward of 20 billion tokens. Users who complete all 20 referrals are also promised early access to upcoming features.
Claude is merging Cowork and regular chat into one experience so users no longer have to choose between a quick conversation and a longer-running work session before starting a task. Claude can answer immediately when the request is small or keep working on longer jobs after the user closes the laptop, with the change rolling out to Pro and Max over the next few weeks and more plans to follow. Claude Docs and Claude Slides are also entering beta on paid plans, Claude Design now works inside conversations, and Enterprise administrators control when those creation tools are enabled.
What'd you think of today's edition? |
Reply