Quanta Bits - September 12, 2026

Hi everyone,

Taking every chance to get outside on weekends before the cold weather arrives. Koda and I took a walk along Memorial Drive and through the MIT campus. Gorgeous views of Boston.

A slightly different format this week. The bigger stories each get their own section, with room to explain what happened and the different ways to read it. Start with The Brief for a quick catch-up, or go straight to whichever story interests you.

This week starts with a mathematical claim from OpenAI, the maker of ChatGPT, and a resignation from Claude developer Anthropic. Both raise questions about what we gain from AI and what we hand over to it.

Cheers.

Reza

Via Bojan Tunguz on X, September 9.

This story got a lot of attention on social media and in publications. At first glance, it looked like a dispute confined to mathematics. I found it interesting because it has implications for how we use AI, including in operations. As a friend once said, sometimes it is not the objective that matters, it is the experience of getting there. You be the judge.

OpenAI, the maker of ChatGPT, claims its AI has solved the Navier–Stokes problem, a question in fluid dynamics that has resisted a solution for roughly 90 years. It is one of seven Millennium Prize Problems, each assigned a $1 million prize. The question is whether equations describing the motion of water and air always produce smooth solutions, or can break down under certain conditions. If the proof holds up, it is an extraordinary achievement.

On one hand, this is exciting: AI helping us solve puzzles of nature that have resisted generations of work. It gives me hope that we can take on harder questions and deepen our understanding. But does getting the answer necessarily give us that understanding? If AI does the work and hands us a proof, even with an explanation, what have we learned? To me, understanding means being able to explain why something works, connect it to other ideas and use it to ask better questions.

That is the debate explored by The Economist. In a public declaration, mathematicians argue that solving a famous problem is a step toward a larger goal: developing ideas and methods that others can study, simplify and build upon. If we use AI only to get the answer, we risk skipping the work that builds our understanding.

An AI-generated proof could give people something new to learn from. The question is whether we do that work, or treat the answer as the finish line.

It reminds me of multiple-choice questions in math and science tests. I always thought the value was in showing your reasoning and bringing ideas together. Choosing the right answer from a list tells us little about what we have absorbed. Growing up, we were judged on how we worked through the problem, not just whether we picked the right answer. An AI proof gives us the reasoning too, of course. But receiving that reasoning and understanding it are different things.

In operations, we face similar questions. My instinct is often to get AI to do most of the work. But when do we just need an answer, and when do we need the judgment that comes from working through the problem? That experience of working the problem helps us frame the next question and decide which direction to take. The same question applies to junior analysts: if AI takes over the work they learn from, how do we help them develop the judgment to guide AI themselves?

I experience this when I let AI do too much task and project management. It helps me organize and focus, but an organized task list can substitute for my own review of why the work matters, how projects relate and which changes deserve my attention. I want its help organizing the work while staying involved in setting priorities and understanding the connections. The warning sign is when I can follow the list but struggle to explain why those are the right things to do.

AI researcher Jacob Coxon left Anthropic, the developer of Claude, arguing that competition is pushing labs toward systems they cannot control. Fellow researcher Evan Hubinger says Anthropic is trying its best, but puts his personal estimate of AI causing human extinction at over 10% within a decade. This is a "p(doom)" forecast, a probability assigned to catastrophe, a personal estimate, not a measured probability. This points to the broad issue that is getting very political: agreement about AI danger hasn't produced agreement about whether to keep building AI at the current pace.

There are also documented failures of control. During July cybersecurity tests with reduced safeguards, OpenAI's agents, AI programs that use tools to carry out tasks, created unauthorized message boards and coordinated attacks, the company reports. Their target was Hugging Face, a platform hosting AI models and datasets. The intrusion reached real infrastructure. Independent investigators at METR, an organization that evaluates AI systems, found agents disguising some commands in their recorded activity, making those commands harder to detect in the logs.

Responses range from technical controls to political restrictions. Senator Bernie Sanders and Representative Greg Casar propose banning AI that greatly exceeds human capabilities and pausing advanced development pending safety rules. An operational "kill switch" gives an authorized person a way to interrupt an agent's work by cutting off access it depends on. For example, MuleSoft lets an operator block a flagged agent's requests to AI models. The gateway, which normally passes those requests along, instead refuses them. The agent can no longer get model responses through that connection. This doesn't shut down the underlying AI model or block any other connections the agent has; those need their own controls.

From predictions about established SaaS providers disappearing to jobs being eliminated and now catastrophic harm, there is plenty of uncertainty about AI. We need to take these questions seriously while continuing to experiment and move quickly. That means defined lanes, speed bumps and governance. In an enterprise implementation, for example, each agent can have a documented set of permissions, with checks before each action to confirm it is allowed. Its actions and results are logged, monitoring flags exceptions, and steps outside its authority require human approval. Those controls need to be enforced by the systems around the agent, rather than left to the agent to follow. This is one practical way the safety debate translates into how we build and operate AI.

A few weeks ago, I wrote about Grok Bot, xAI's always-on assistant that can work across applications. This week's Meta and Apple announcements extend that competition: who becomes the place you turn to when you want AI to do something?

Meta, which owns Facebook, Instagram and WhatsApp, announced Muse on September 8. Initially available to US adults, Muse is designed to use personal context to suggest activities and perform tasks through connected services. Meta's advantage is proximity: 3.6 billion people used its family of apps daily in June. That is an audience Meta could introduce to Muse; regular use still has to be earned. Meta promises permission checks and says Muse data stays out of its advertising systems. Given Facebook's history of privacy enforcement, I would want to see those boundaries working before trying it with my own data.

Apple's September 9 presentation of Siri AI, coming in beta with iOS 27, approaches the same opportunity through the iPhone. Apple says Siri will find information across messages, emails and photos, and act on what's onscreen. Its privacy pitch combines processing on the phone with protected servers for requests needing more computing power. Muse offers delegation through connected services; Siri is built into the device and its apps. Their abilities overlap, but they aren't interchangeable, and neither company's privacy promises remove the need to understand what access you're granting.

The competition is over the everyday relationship with the user. Meta brings its social apps; Apple brings the phone. Neither has to win every model comparison to become the familiar way people access AI. It will come down to which ones people trust more and find convenient to use.

Source: James Hawkins (@james406) on X, June 6, 2026

Companies buying technology consulting are pressing advisers to lower fees or change how they charge, the Financial Times reported on August 31. At drugmaker Bristol Myers Squibb, technology chief Greg Meyers says AI is taking over some security monitoring previously bought from outside firms. The company is pushing advisers toward fixed-price or performance-related contracts. Clients want to share in the savings when work takes fewer hours.

Technology consulting firm Capgemini disputes software company SAP's prediction that customers could halve spending on outside help setting up SAP's business software. But Capgemini acknowledges making those projects faster and cheaper. How much of the saving reaches their customers remains a negotiation.

This strengthens the case for pricing around value and agreed outcomes. Fewer hours needn't mean less valuable work. A fixed fee only becomes outcome-based when payment is linked to agreed results. I would start the negotiation with the work being delivered, the measures of success and what the adviser can control; a blanket AI discount says little about any of those.

On September 10, OpenAI released its Agents API in public beta. An API lets one program use another service. Here, developers can use the software behind Codex, OpenAI's coding agent, to manage agents that use tools and work through tasks over time. OpenAI's demonstration shows an agent investigating a software incident and producing a report with evidence and suggested next steps.

ChatGPT Sites addresses another practical need: turning AI-built dashboards, project trackers or reports into websites OpenAI hosts for others to use. Depending on account permissions, sites can have a restricted audience or be public, and authorized colleagues in the same workspace can edit them. A September 3 update added access for named external viewers without making a site public. Those guests can use it, but cannot edit it. That makes client collaboration easier, but sharing a dashboard built from internal data also means deciding which of that data the client may see.

The September 10 GPT-Live-1 API launch gives developers a voice model that can listen and speak simultaneously, allowing people to interrupt while it is talking. OpenAI's customer examples include language tutoring and restaurant reservation calls. Other models and connected tools supply the deeper reasoning and actions behind the conversation.

These are separate offerings, but I see a common direction: less setup between an AI experiment and something other people can actually use. That could accelerate adoption inside a business. It also means relying on the same provider for more of the work, a commitment that becomes harder to unwind once an experiment is part of daily operations.

On August 19, Moderna and Merck reported encouraging late-stage trial results for a vaccine tailored to each patient's tumor. The trial involved melanoma, a form of skin cancer, after surgical removal. Adding the vaccine to the immune-system treatment Keytruda improved the time patients remained free of recurrence or death compared with Keytruda alone.

The promise comes with important unresolved questions. Melanoma is unusually visible to the immune system; other cancers may respond differently. The announcement did not disclose the size of the benefit, and whether this trial shows patients living longer remains under study. Trials in other cancers will help establish how broadly the approach works. This is an encouraging treatment result, not evidence of a vaccine that prevents or treats every cancer.

Dir. Kyle Balda | 109 minutes | ★★★★½ | Hugh Jackman, Julia Louis-Dreyfus, Emma Thompson

I really liked The Sheep Detectives. A shepherd's murder leaves his mystery-loving flock trying to solve the case. It mixes live action and animation, with real people alongside the animated sheep, and it's done very well. What surprised me was how harsh and adult some moments were. Despite the sheep, I wouldn't describe it as a kids' movie. The whodunit is told in a different way, the plot kept me interested, and there are some very funny moments. A few bits were too silly, but that was part of its charm. It wasn't perfect, but I'd certainly like to watch more movies like this.

Graham Greene | 144 pages | ★★★★★

I'm a big fan of Graham Greene's books and his prose: sharp, to the point, imaginative. The Third Man didn't disappoint. Rollo Martins arrives in occupied Vienna after WWII to meet an old friend and finds him dead. The story moves fast, and the characters are easy to follow. As Greene explains in the introduction, he wrote it as the basis for the 1949 film. It feels a little thin compared with his later work, but I expected that and still liked it. Not one of his best, but quite entertaining nevertheless.

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what's relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide direction, we draft together. The editorial judgment is mine. He'd tell you the same. Most logical. 🖖