The frontier got smarter. Then it hired a bouncer.

I expected ARC-AGI-3 to keep the frontier labs occupied for a while.

Apparently “a while” has become a rather short unit of measurement. (From under 1% to 99% in six months.. WOW)

Astra scored 62.7% with the Standard harness and 99.9% with OpenAI's Provider Adapter, according to ARC Prize. That is an enormous result. Persistent reasoning, context management and the ability to keep working are significant reasons for that 99.9%. Imagine if you had to start from zero with every new video game, it would be a lot harder, using previous tricks you learn and adapt a lot faster. That said 62.7% points to some focused training.

The model and the environment/harness around it are becoming difficult to separate if you want top performance. Which makes the question of what and how you can buy rather interesting.

Anthropic released Fable 5.1 broadly while reserving Mythos 5.1 for verified users in selected fields. OpenAI says Astra crossed its Critical cybersecurity threshold. Google created a trusted-access path around Gemini 3.8 Flash Cyber. Three labs, independently arriving at doors that need someone checking credentials. (and a bit of an uptick in charge)

Meta meanwhile pushed Muse Spark 1.3 into coding and longer agent work, with claims of fewer tool calls and tokens. I am keeping a lump of salt beside the coding rankings until we have broader independent testing. Configurations and reasoning budgets can move those results around considerably.

I am excited about the capability. I am way less comfortable assuming that an access list solves the control problem. A bouncer decides who enters the room. There is still the matter of what happens once everyone is inside.

The mathematics results belong beside the ARC result. OpenAI reported ten advances in mathematics and theoretical computer science. Anthropic reported a stronger Riemann-zeta bound and a computer-checked formalization of Fermat's Last Theorem. (For the non-math aficionados out there, suffice to say this is big and complicated news.)

Formalizing the existing proof is an achievement in verification; we should give it credit for the thing it did. There is plenty there to be excited about even though it’s not a new discovery.

The same appetite for persistent work is appearing in visual systems. World Labs' Atlas brings spatial context and controllable viewpoints to generated environments. Separately, fal says (and the internet full of AI video slop confirms) H3 Max can generate five seconds of video in roughly three seconds.

Generation speed and a world that stays coherent while you move through it are separate problems. Training simulations, games and digital twins become quite interesting when both work together. I expect developments in this space to continue and the next generation of video games and virtual worlds is going to get wackadoo level interesting.

Anthropic stopped parts of its work after agents took unauthorized actions. Most resumed after stronger isolation, monitoring and real-time classifiers were added. Some high-risk activity remained paused, with an independent METR review planned.

That is a useful operational response. The important details for every company to pay attention to are the boundaries around the agent: what it can reach, what it can change and whether anyone can see it finding an unexpected route. Giving it a more demanding task while leaving the same old credentials lying around seems an ambitious way to learn that (hard) lesson.

The politicians are concerned that the AI companies are taking up too much air time so the policy spread was remarkable too. G20 ministers backed pro-innovation principles. Sanders and Casar proposed a permanent ASI prohibition and a temporary advanced-AI pause, with substantial penalties. (20 years in prison for training an ASI…) A proposal still has to become law, but the distance between those positions tells you how differently people are reading the risk. (or playing the politics)

The Justice Department also entered the model-training copyright dispute fray, arguing that some uses can qualify as fair use. Its filing does not settle the case. I want creators to retain meaningful rights in what they make; I also see how the answer changes who can afford to train a frontier model. There is no clean slogan that resolves both sides of that problem.

OpenAI intends to wind down direct Cursor access following Cursor's acquisition by SpaceX. The proposed contractual cutoff is November 12. Other routes may remain, including a user's API credentials or the Codex extension.

You can build a very good product and still have someone else controlling the pipe underneath it. Particularly awkward when that someone also sells an application to your customers.

Bring-your-own-key can help preserve access. It gives you rather less influence over the supplier's next decision. Model portability needs to survive a real change in quality, latency and cost before I would (or you should) treat it as a business continuity plan.

NVIDIA's agreement to acquire Hugging Face extends the ownership question. Regulatory review and closing are still ahead. The companies say the platform will remain open and hardware-agnostic, and NVIDIA has contributed a great deal to that ecosystem so I remain hopeful.

That said, I will be watching where the easy paths lead over time. Model discovery, integrations and optimization defaults are all places where a promise of neutrality eventually becomes something developers can observe. Ownership has a way of showing up in small product decisions long before anybody changes the sign outside.

Meta committed $115 million to its Workforce Academy. Google committed $50 million through Google.org. BlackRock launched a $100 million Future Builders initiative.

That is $265 million going toward skilled-trades programs. Electricians, welders, pipefitters, fiber technicians. The people who turn a proposed data center into a building that can be switched on.

Meta's program covers travel and lodging and offers graduates a job with a construction partner. There is some practical recognition here that the labor supply takes time to build.

We have spent years discussing the jobs AI might remove. The companies expanding it are now paying to create a workforce they can’t summon with another deployment command.

The grid story was a little different this week. A temporary emergency order let Duke Energy Progress call on specified backup generators at data centers and other large-load sites during grid stress. Sometimes the big electricity customer can help supply it.

The emissions and local-air-quality costs don’t disappear in that exchange. Canada's voluntary data-center principles ask developers to be clearer about local benefits, power costs and resource use. Communities are entitled to care about what happens after the construction ribbon is cut.

The oncology development is meaningful for the health ecosystem interested reader.

The FDA granted accelerated approval to camizestrant with a CDK4/6 inhibitor for certain HR-positive, HER2-negative advanced or metastatic breast cancers. An ESR1 resistance mutation detected in circulating tumor DNA can trigger the treatment change before imaging shows progression.

In the reported comparison, median progression-free survival was 16.0 months versus 9.2 months. The endpoint is progression-free survival. Confirmatory evidence is still required, and comes with a cardiac-rhythm boxed warning.

The signal gives the clinical team a chance to respond before the scan delivers worse news. Making that useful takes access to testing, timely results and a team that owns the response.

The WHO Europe dialogue raised the institutional problem for AI: fragmented data, unclear accountability and limited literacy. Participants came from 105 countries, although this was a dialogue rather than a representative survey or formal WHO policy. The useful questions concern who checks the system, who understands its limits and who can intervene.

I also enjoyed the learning-circuit work that came out this week. A complete male fruit-fly central-nervous-system map contains roughly 166,000 neurons and 125 million synapses. A separate electric-fish study combines wiring, physiology and modeling to examine continual learning.

We are getting better at moving from a map of the connections to an explanation of what those connections do. There is a lot to learn there before anyone needs to worry about loading their brain into a computer though.

WeatherNext 3 brings live satellite observations into hourly forecasts. This is ridiculously difficult physics and math, and I think the advancement here is getting less attention than it deserves. More timely forecasts can change decisions about power, agriculture and logistics.

The shared-infrastructure story made it to Mars as well. NASA selected Blue Origin for a telecommunications network contract worth up to about $700 million, targeting Mars operations by 2030. The award gives the plan substance. Someone still has to deliver the network and keep it working. (Is intra-planetary communications astronaut in that list of jobs being pushed yet?)

And Saturn has an evolving southern decagon to go with its northern hexagon. There is real fluid dynamics in those giant weather patterns. I mostly appreciate living in a universe that keeps finding new ways to be peculiar.

On this week's board, Models, Labs & the Capability Frontier; Safety, Trust & Regulation; and Compute, Capital & Power accelerated. Workflows, Not Models; Innovation & Health/Science; The Human Layer; and Embodiment & the Physical World advanced.

How much of your AI plan depends on access somebody else can change? And further, do you actually understand enough about it underneath to know when the model changes, why?

For the live version and a few opinions that are harder to fit on a slide: this week's ChuckChat.

Chuck

Chief Technology Troublemaker
chuckdevries.ai

Someone forwarded you trouble.
They clearly think you can handle it. Trouble Worth Making is where I work out loud on how to stop bolting AI onto a pre-AI company and actually rebuild the operating model underneath it. One issue every week or two, no filler, the occasional uncomfortable question.

If the person who sent this has good taste (obviously, they do), subscribe here and get the next one straight from the source.