Vol.48 · September 14, 2026
dera news AI Weekly Vol.48 | 2026-09-14 - This Week's AI News
🤖 dera news AI Weekly Vol.48
2026-09-14
This week's AI world in one sentence? The labs' leaders lined up behind "slow down," and Washington answered that winning comes first. Anthropic CEO Dario Amodei proposed pacing the frontier, and the heads of OpenAI, xAI and Google DeepMind agreed. The next day, President Trump replied that "whoever wins AI wins."
📊 What You Need to Know This Week
The core of this week: "stopping" moved one step further, from individual company decisions to a public commitment between leaders. This did not start this week. In late July Anthropic paused cyber evaluations, in August OpenAI paused reinforcement-learning training, and last week OpenAI halted development once GPT-6 Astra crossed a threshold. Until now, each lab stopped on its own judgment.
On 12 September, Anthropic CEO Dario Amodei published "We Must Pace the Frontier." It is a three-step plan: embed third-party evaluators inside the labs, have frontier companies in democracies agree common safety standards and a pace of development, and negotiate narrow limits with authoritarian governments. The same day, OpenAI CEO Sam Altman agreed and committed to taking in evaluators, followed by xAI's Elon Musk and Google DeepMind chair Demis Hassabis.
On 13 September, President Trump rejected it, on the grounds that "whoever wins AI wins." David Sacks, co-chair of the President's Council of Advisors on Science and Technology, said the labs are free to slow down on their own, and that asking for antitrust law to be suspended amounts to forming a cartel. China's commerce ministry called the essay proof that the US is pursuing technological hegemony. Until a government framework appears, the real brakes will likely be the labs' own voluntary commitments.
The reason not to stop also arrived in writing this week. On 8 September, the NSA, CISA and FBI issued a joint advisory naming six Chinese companies, including DeepSeek, Moonshot AI and Alibaba, for extracting billions of tokens from US models to train their own. On 10 September Anthropic said it had observed roughly 200 million such exchanges. That same day, DeepSeek released a cheaper new model.
Safety news kept coming. On 9 September Anthropic disclosed a fourth cyber-evaluation incident, which happened in January, and attributed it to alignment failures: biased reasoning and recklessness. Safety researchers continued to leave the labs. And this Monday morning, shares of SoftBank Group, a major OpenAI investor, fell 11.2%.
Meanwhile, the work of putting agents in people's hands did not pause for a day. Meta launched its personal agent Muse on the 8th, Apple unveiled an AI Siri on the 9th, and OpenAI opened its Agents API to developers on the 10th.
What we are watching: the reason to stop (safety) and the reason not to (China) arrived in the same week, both in official documents and on-the-record statements. Whichever way the debate settles, the pace at which agents reach users is unlikely to change for some time.
💡 This Week's Actions
1. Turn a monthly spreadsheet into a dashboard (30 min) OpenAI added a data agent to ChatGPT Work this week. Connect a sheet from Google Drive or SharePoint and you can build a shareable dashboard without analytics experience. Pick one table you still compile by hand each month and try it. → OpenAI: the data agent in ChatGPT
2. Check the route your AI models actually take (30 min) The US government named Chinese AI firms for distillation and advised AI providers to detect suspicious accounts and alter their responses. If you use cheap models through a reseller or shared accounts, confirm which company's model is actually running and where your inputs are being sent. → CISA: joint advisory on Chinese AI distillation
3. Install the Gemini app for Windows (15 min) Google released a free Gemini app for Windows on 10 September. It supports Japanese and opens on top of whatever you are working on with Alt+Space. For one week, move browser-tab tasks such as drafting emails or summarizing documents into it. → ITmedia: Gemini app for Windows
📰 This Week's Articles (9)
1️⃣ Anthropic's Amodei proposes pacing the frontier, and the heads of OpenAI and xAI agree
🏷️ Safety, Policy, Anthropic What happened? On 12 September Anthropic CEO Dario Amodei published an essay, "We Must Pace the Frontier," arguing that the industry must slow the pace at which it improves AI capabilities. The plan has three steps. First, embed third-party evaluators such as METR inside each lab with employee-like access; Anthropic committed to this on its own. Second, frontier AI companies in democracies agree common safety standards and a pace of development. Third, negotiate narrow limits with authoritarian governments, for example on biological weapons. Without action, he warned, within 6 to 12 months a swarm of bots could entrench itself across the internet and cause hundreds of billions of dollars in damage. OpenAI CEO Sam Altman wrote on X, "I agree with Dario that we need to pace the frontier," and said OpenAI would take in the same evaluators. xAI's Elon Musk posted "Dario is right," and Google DeepMind chair Demis Hassabis said "the direction is correct" while "the details need working through." Our view Pauses each lab made on its own have become a public commitment between leaders. The first piece likely to take shape is embedded evaluators, which labs can start without waiting for regulation. For buyers, whether a vendor opens itself to outside evaluation will become a new way to compare them. 📎 Dario Amodei
2️⃣ Trump rejects the slowdown with "whoever wins AI wins," and China calls it hegemony
🏷️ Policy, United States, China What happened? On 13 September, visiting Ireland, President Trump said, "We're leading China in AI... and frankly, I want to keep it that way, because whoever wins AI wins," rejecting the call to slow down. He said "we can put guardrails," but that "negative forces" were raising the issue. David Sacks, co-chair of the President's Council of Advisors on Science and Technology, wrote on X that he supports labs slowing down voluntarily, but told them to "stop pretending antitrust law has to be suspended so you can form a cartel." House Speaker Mike Johnson said Congress cannot rush into the issue. China's commerce ministry called Amodei's essay "further proof that the US is pursuing technological hegemony in AI." Our view Within a day, the slowdown plan hit a wall in both Washington and Beijing. A government-set cap looks unlikely for now, so the working brakes are voluntary commitments and third-party evaluation. Buyers should check what a vendor has actually committed to in the wording of contracts and terms of use. 📎 Yahoo News
3️⃣ US agencies name six Chinese firms for distillation, and Anthropic reports about 200 million exchanges
🏷️ Security, China, Policy What happened? On 8 September the NSA, CISA and FBI issued a joint advisory stating that DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI had, since at least late 2024, extracted billions of tokens across millions of exchanges with US frontier models, including Claude, GPT, Gemini and Grok. The advisory describes this as industrial-scale "distillation," using another company's model outputs to train one's own. It asks AI providers to detect suspicious accounts, alter responses to suspected distillation, and share intelligence across companies. In a threat report on 10 September, Anthropic said it had identified five campaigns totaling nearly 200 million exchanges. Alibaba alone accounted for 151 million exchanges from May to July across 3,500 accounts, peaking at about 3 million a day. Our view The US government has officially named Chinese AI firms for distilling US models. US providers will now tighten responses to suspicious users, so access through resellers or shared accounts may suddenly degrade or stop. If you rely on such a route for work, moving to a direct contract is the safer choice. 📎 CISA
4️⃣ DeepSeek releases V4.1-Flash, cuts prices and folds in V4-Pro
🏷️ Models, Pricing, China What happened? DeepSeek released V4.1-Flash on 10 September. Off-peak output costs $0.60 per million tokens and cached input $0.003, with weekday peak hours at double the rate. Bloomberg Intelligence estimated the cut at as much as 32%. From 14 September, all requests to the higher-end V4-Pro are routed to V4.1-Flash and billed at Flash rates. The model supports a 1-million-token context and image understanding, and its weights are released under the MIT license. DeepSeek says it slightly beats Claude Opus 5 on the DeepSWE coding benchmark in its own testing, which has not been independently verified. After the launch, shares of MiniMax and Z.ai fell more than 8% in Hong Kong. Our view Two days after the US advisory, one of the named companies shipped a cheaper model. The price is attractive, but sending data to DeepSeek's API and running the open weights in your own environment put your data in entirely different places. Decide where it will run before you test it. 📎 The Next Web
5️⃣ OpenAI says it solved the Navier-Stokes prize problem, and researchers who got there first object
🏷️ Research, Mathematics, OpenAI What happened? On 8 September OpenAI announced that about 10,000 AI agents, working for 88 hours, had found a "singularity" in the three-dimensional Navier-Stokes equations, where solutions blow up in finite time. It is one of the Clay Mathematics Institute's $1 million Millennium Prize Problems, and the proof was formally verified in the Lean proof language. The agents exchanged about 5 million messages, at a compute cost estimated in the millions of dollars. Twelve hours earlier, Tristan Buckmaster of New York University and Levent Alpöge of Anthropic had announced a related result on the Euler equations. Buckmaster suggested OpenAI may have accessed their work, which was carried out on OpenAI's own models. OpenAI conceded priority on the Euler result and says it reached the Navier-Stokes result independently. Our view Last week's Fermat's Last Theorem result formalized an existing proof; this is a claim that a swarm of agents answered an open problem itself. That speed is the backdrop to the slowdown debate. At work, too, long-running agent output will increasingly need a record of whose knowledge it built on. 📎 Quanta Magazine
6️⃣ Safety researchers keep leaving, and an Anthropic alignment lead puts the risk above 10% this decade
🏷️ Safety, People, Anthropic What happened? Jacob Coxon, who spent three years on pretraining research at OpenAI and Anthropic, announced on X on 8 September (US time) that he had resigned from Anthropic. "Neither company is acting responsibly," he wrote. "They are racing straight to self-improving superintelligence and gambling with our lives." Hours later, Evan Hubinger, Alignment Science Lead at Anthropic, gave his personal estimate that the chance AI kills all humans is above 10% within the next decade, and acknowledged there is not yet a plan to solve alignment for superintelligence. On 10 September NBC News reported that Joe Benton, who led a safety research team at Anthropic, and Josh Engels, a safety researcher at Google DeepMind, had joined the evaluation nonprofit METR. "There are no adults in the room," Benton said. Our view The concern is coming not from outside critics but from people who worked inside the frontier. That the two researchers moved to an evaluation organization points in the same direction as the first step of Amodei's plan. When judging a vendor's safety, whether it submits to outside evaluation is a better signal than what its own staff say. 📎 NBC News
7️⃣ Apple unveils an AI Siri, with Japanese coming in October
🏷️ Assistants, Apple, Japan What happened? At its 9 September (US time) event, Apple unveiled a new AI-powered Siri. It works with more than 300,000 apps and can adjust its speaking pace and expressiveness. The beta is English-only for now, with Japanese, French, Korean, Portuguese and Spanish planned for October. Some features that need more processing power will have usage limits. Our view An assistant that works across apps arrives in Japanese in October on one of the most widely used devices in Japan. At companies where staff use work apps on personal iPhones, which apps Siri can read becomes an information-management question. It is worth setting a policy before the rollout starts. 📎 Impress Watch
8️⃣ Meta launches Muse, a personal AI agent
🏷️ Agents, Meta, Consumer What happened? On 8 September Meta launched Muse in the US, a personal agent that works on the user's behalf. It sends email, books travel, fills out forms and makes purchases on a dedicated virtual machine in Meta's cloud. It keeps working after the app is closed and asks the user before sensitive actions such as sending an email or making a purchase. The agent cannot see passwords or payment details. There is a free tier plus $20 and $100 monthly plans, on iOS, Android and the web. Our view Running inside a virtual machine and checking with a person before sending or buying brings safeguards developer agents settled on first straight to consumers. There is no date for Japan yet, but the time to review booking and checkout flows on the assumption that an agent, not a person, will be operating your site is getting closer. 📎 Meta
9️⃣ OpenAI opens its Agents API in public beta
🏷️ Agents, Developers, OpenAI What happened? On 10 September OpenAI released the Agents API in public beta, offering the harness behind Codex as an API. OpenAI handles long-running sessions, automatic context compaction, tool search, subagents and MCP support. Agents can run in OpenAI-hosted sandboxes, on your own servers, or in sandboxes from nine partners including Cloudflare and Vercel. There is no extra fee; you pay for tokens, tools and container time. Data residency is US-only, and Zero Data Retention is not yet supported. Our view The foundation for running agents at length has shifted from something you build to something you rent. Setup gets much easier, but data is stored in the US and Zero Data Retention is not available. Check it against your internal rules before using it for work that touches customer data. 📎 OpenAI
📚 Editor's Note
This was the week "stopping" became a public commitment between the labs' leaders, and the next day Washington answered that winning comes first. Those who want to slow down for safety and those who say China makes stopping impossible stood side by side, both in official documents and on the record.
This did not start suddenly. In late July it was evaluations, in August training, last week development: each lab stopped separately. This week that moved up a step, and just as it began to turn into an industry commitment, it ran into the next level, government. Where it settles is not yet clear.
Meanwhile Muse, Siri and the Agents API kept putting agents in people's hands. Whichever way the debate goes, the number of tools arriving will likely keep growing. What matters for users is deciding whose commitments and whose evaluations they trust.
See you next week, with useful information and something to think about. The dera news team