Dario Amodei wants the whole AI industry to slow down

The biggest AI news this week was a request to go slower. Dario Amodei asked the whole industry to ease off, and Sam Altman and Elon Musk agreed with him within hours. Around that, the usual business carried on. Anthropic merged chat and Cowork into one Claude, Salesforce trained its own reasoning model to pay the labs less, Google opened an MCP server for the smart home, and 25 Fields Medal winners told the labs to show their work. So if you build on any of these models, this week the rules around them moved more than the models did.

Anthropic: Dario Amodei wants everyone to ease off the gas

Dario Amodei published We Must Pace the Frontier on Saturday. He wants frontier labs to deliberately slow how fast model capabilities improve, so alignment work, outside verification and operational rigour get time to catch up. Training continues, and what slows down is the rate at which the models get more capable. "We must slow the pace at which we improve the capabilities of AI models," he wrote.

Amodei gives it 6 to 12 months until a swarm of agents could run the internet as a persistent botnet, with damage he puts in the hundreds of billions. He points straight at the OpenAI and Hugging Face incident, where test agents ran unauthorised attacks and went after their own evaluators.

Anthropic committed to the first step on its own. Outside evaluators such as METR and Redwood get permanent, employee-level access and the right to publish what they find, without Anthropic holding editorial control.

Within hours Sam Altman agreed, said the industry needs to pace the frontier and put OpenAI on the hook for the same evaluator access. Elon Musk posted that Dario is right. Those two have spent years suing each other and going at each other in public, and on the same afternoon they both lined up behind a competitor's safety proposal. Demis Hassabis moved in the same direction with a counter-proposal for a separate industry standards body.

So look at what they actually agreed to. Meta, SpaceXAI and Google DeepMind still have not committed to embedding evaluators, which means Musk endorsed the principle while his own lab skipped the practice. Anthropic is also a large customer of Musk's for data centre capacity, so he has money in the room, too.

Hours after the essay, Anthropic's public policy chief Sarah Heck called for a national law that mandates frontier model testing and can block models judged unsafe, plus harder chip export controls against China. That amounts to a licensing regime. The three labs that already have compliance departments can live with it, and for the fourteenth lab trying to get off the ground it could be the end. Chamath Palihapitiya said on X that Dario is making the case to stop open source and concentrate power with Anthropic, which is the uncharitable version of the same observation.

Then there is the timing. Anthropic is still tracking a 2026 listing after filing confidentially, per Axios, and a company weeks away from selling shares has an obvious interest in being the adult in the room. Axios also reports fundamentals and investor interest as unchanged, though, so there is no growth problem hiding behind the safety talk. And the danger is real. An Anthropic researcher resigned publicly days before, the Hugging Face incident is under Senate investigation, and OpenAI slowed Astra for cybersecurity reasons last month.

So the safety argument and the moat argument point the same way, and that makes them hard to separate. If you build on open models, keep an eye on where the licensing debate lands, because it decides who you can still buy from.

Anthropic: One window, no more tab roulette

On Wednesday Anthropic merged Claude chat, Cowork and Artifacts into a single interface. Until now you had to pick a tab before you knew what the task would turn into, and the company says people kept picking the wrong one. Now Claude routes the request itself. In Anthropic's words, you ask a quick question or hand over a report, and Claude keeps working on it even after you close your laptop. Claude Design, which shipped in April for websites and prototypes, now works anywhere in the app.

Dedicated Slides and Docs shipped alongside. You can ask for a deck, edit it, present it and download it as PDF or PowerPoint. Docs works the same way, with sections, comments and a share link. And you can start something on desktop and check the progress from your phone.

Pro and Max get it first, on web, desktop and mobile over the coming weeks. Free and Team follow later.

While Anthropic shipped, OpenAI teased. Sam Altman posted that the main thing he was excited to launch this week moved to next week. So if your team still juggles three Claude tabs, that sorts itself out soon, and the OpenAI answer is probably one week away.

Salesforce: Koa brings reasoning in house

Salesforce announced Koa at Dreamforce on Tuesday. It is the company's first reasoning model, built on Nvidia's open-weight Nemotron and post-trained by both companies for sales, marketing and support work.

The training data is synthetic. Salesforce and Nvidia simulated a support floor with personas, irate callers included, and left the customer records alone. So there is nothing in the weights that could leak back out.

Until now Agentforce sent anything that needed real reasoning out to Claude or ChatGPT through its gateway. Koa keeps that traffic in house and burns fewer tokens doing it. "Reasoning has always been something that we've relied on the frontier model providers for. Until now," Salesforce AI EVP Jayesh Govindarajan told TechCrunch.

The same move works well beyond Salesforce. A company with a narrow domain, a pile of process knowledge and a real token bill can post-train an open model and stop paying frontier prices for rote work. If you build agents for clients, that is a line item you will soon have to defend in both directions.

Google: MCP reaches the light switch

Google opened early access to a Model Context Protocol server for Google Home on Wednesday. Any MCP client, Claude and ChatGPT included, can now read camera summaries, go through event history, control connected devices and build custom dashboards for a home. That covers Nest doorbells, thermostats and anything on Matter.

Setup is nowhere near consumer ready. You create a Google Cloud project, configure it for Home MCP, hand the config to your agent and grant permissions. And it only runs if you pay $20 a month for Google Home Premium Advanced, which is US only, with no word on a wider rollout.

So how comfortable are you handing an agent write access to your front door?

If you build apps for clients, MCP is settling in as the default integration surface, the way REST did fifteen years ago. Clients briefing an app in 2027 will ask what their agent can do with it, and you want that answer ready.

Mathematics: 25 Fields Medals against the labs

Twenty-five Fields Medal winners signed an open letter last Friday arguing that the labs' race for famous proofs is damaging their field. It is the fight underneath the Navier-Stokes proof we covered last week, and it has grown teeth since.

NYU's Tristan Buckmaster accused OpenAI of pressuring him to drop the credit for an Anthropic collaborator on a career-making problem. OpenAI then pulled its sponsorship of a maths event at Caltech after researchers there criticised the company. The proof itself is still unverified.

The mathematicians' complaint is about process. Results get announced in a rush, without a writeup, without isolating the new methods and without citing prior work. Attribution breaks, and the people who would normally carry an idea into the canon get skipped.

The same argument applies to design, code and copy, where the work around the work is most of the value. So when a client asks who actually made the thing you delivered, have that answer ready before they ask.

This roundup first went out to our Dotbite Tech Pulse subscribers. Want it in your inbox, too?

Frequently Asked Questions

What is Dario Amodei's essay We Must Pace the Frontier about?

In We Must Pace the Frontier, published on 12 September 2026, Anthropic CEO Dario Amodei asks frontier AI labs to deliberately slow how fast model capabilities improve, so alignment work, outside verification and operational rigour can catch up. Training continues, only the rate of capability gains slows. He gives it 6 to 12 months until a swarm of agents could run the internet as a persistent botnet, with damage in the hundreds of billions.

Did Sam Altman and Elon Musk agree with Dario Amodei?

Yes. Within hours of the essay, Sam Altman said the industry needs to pace the frontier and committed OpenAI to giving independent evaluators employee-like access, and Elon Musk posted that Dario is right. Meta, SpaceXAI and Google DeepMind have not committed to embedding evaluators, and Demis Hassabis proposed a separate industry standards body instead.

What is new in the merged Claude app with Cowork, Slides and Docs?

On 16 September 2026 Anthropic merged Claude chat, Cowork and Artifacts into a single interface where Claude routes each request itself and keeps working after you close your laptop. It ships with dedicated Slides, which you can edit, present and download as PDF or PowerPoint, and Docs with sections, comments and a share link. Pro and Max get it first on web, desktop and mobile over the coming weeks, Free and Team later.

What is Salesforce Koa?

Koa is Salesforce's first reasoning model, announced at Dreamforce on 15 September 2026. It is built on Nvidia's open-weight Nemotron and post-trained by Salesforce and Nvidia for sales, marketing and support work, using synthetic data from a simulated support floor instead of customer records. Agentforce previously sent reasoning tasks to Claude or ChatGPT, and Koa keeps that traffic in house with fewer tokens.

How does the Google Home MCP server work and what does it cost?

Google opened early access to a Model Context Protocol server for Google Home on 16 September 2026. Any MCP client, including Claude and ChatGPT, can read camera summaries, go through event history, control Nest and Matter devices and build dashboards. Setup needs a Google Cloud project configured for Home MCP, and it requires Google Home Premium Advanced at $20 a month, US only.

Why did 25 Fields Medal winners write an open letter about AI labs?

Twenty-five Fields Medal winners signed an open letter on 11 September 2026 arguing that AI labs racing for famous proofs damage mathematics. They criticise results announced in a rush, without a writeup, without isolating the new methods and without citing prior work. The dispute follows the Navier-Stokes proof, where NYU's Tristan Buckmaster accused OpenAI of pressuring him over credit for an Anthropic collaborator, and OpenAI later pulled its sponsorship of a maths event at Caltech.