The Signal — September 14, 2026
Dario Amodei asked the industry to slow down. Within two days every major lab had said some version of yes, the White House and the Speaker of the House had said no, and nobody had signed anything. Below that argument, two quieter items: the NSA is rebuilding its org chart around an AI mission center, and a group of Korean researchers found that the steps a model writes out in its reasoning have identifiable shapes inside it.
Four lab leaders agreed to be watched, and Washington told them to relax
We covered the first round of this on September 11, when Sam Altman told OpenAI staff the company was open to slowing down and President Trump said he was not worried. Amodei's essay, We Must Pace the Frontier, went up on September 12 and turned a set of private positions into public ones. He argued that the rate of capability improvement has to come down so risk work can keep up, and he committed Anthropic unilaterally to giving third-party evaluators permanent, employee-level access.
The responses came fast and they were not identical. Altman wrote that pacing the frontier "has been a primary topic of discussions we've had at OpenAI in recent weeks," called employee-like access for independent evaluators a good idea, and said OpenAI would do the same, with more to share soon. Elon Musk wrote that Dario is right. Demis Hassabis endorsed the direction and said the details still needed work. Satya Nadella went furthest into specifics, welcoming "the research, focus, and deliberate pacing needed to get alignment right," naming embedded evaluators as a mechanism worth building, and saying Microsoft would publish the Code of Conduct underlying its first-party MAI models for public consultation. He attached a condition none of the others did, which is that oversight "cannot be controlled by a handful of entities" and needs representation across countries, fields and academia.
Then Washington answered. Speaking to reporters at his golf resort in Doonbeg, Ireland, President Trump said the warnings were exaggerated and that "negative forces" were raising things that should not be raised. "We're leading China in AI, we're the most sophisticated country in the world, and frankly, I want to keep it that way because whoever wins AI, wins." He allowed that guardrails were possible without saying which ones. House Speaker Mike Johnson told CNN's State of the Union that an emergency congressional session on AI would lose the race to China, and that he would rather tech companies led on safety themselves. David Sacks, co-chair of the President's Council of Advisors on Science and Technology, took a third position: go ahead and slow down if your unreleased models scare you, but stop pretending you need permission, and stop pretending the motivation is purely altruistic when you are carrying product-liability exposure for any cyberattack your systems enable.
Research had its own dissent. Google's Peyman Milanfar argued that recursive self-improvement is a naive premise because the feedback loops are unstable, that reliable evidence comes from deployment rather than benchmarks, and that any system which actually improves itself will end up slow and heavily damped on its own. Stability is the speed limit, he wrote, so an imposed one is unnecessary.
What exists after all of this is four public statements and one company's unilateral commitment. Anthropic's evaluator access is self-described and self-administered. Altman's "we will do the same" has no published mechanism attached to it yet, and Musk endorsed a sentence. Microsoft's MAI Code of Conduct is a draft opened for comment; Nadella said it would appear the day after his post, and it had not when we checked. Four CEOs agreeing that someone should watch them changes what is sayable out loud inside these companies. It is not yet a rule, a law, or anything a person outside them can enforce.
Sources: The Verge · NPR · The Decoder · Sam Altman on X · Satya Nadella on X · Dario Amodei
The NSA is rebuilding its org chart around five mission centers, one of them for AI
The National Security Agency is dissolving its existing directorates and replacing them with five mission centers at Fort Meade, covering artificial intelligence, China, cybersecurity, combat support, and global intelligence. Army Gen. Joshua Rudd, who runs both the NSA and U.S. Cyber Command, shared the plan with the workforce earlier this month and started a 30-day clock. Each center gets a newly elevated mission director carrying deputy-director authorities. That places the head of AI near the top of the agency's hierarchy with budget and hiring power attached, which a working group does not have. Some of the chiefs have been chosen, including the head of AI, and the five bodies are expected to reach full operational capability by January.
The agency has tried this before. NSA21, roughly a decade ago, merged the offensive and defensive organizations and is still described by current and former officials as a failure that mostly generated bureaucratic confusion. A second plan called EVO25, started around the beginning of the second Trump administration, was widely read inside the building as a way to pre-empt a restructuring imposed by DOGE, and it was abandoned. Rudd told his confirmation hearing in January that his first priority is speed, and both a current and a former official told The Record that he and Deputy Director Tim Kosiba know the pace will break things in a slow-moving bureaucracy and plan to fix them in flight. Several large questions are still open, including where the Cybersecurity Collaboration Center ends up and how any of this changes the shared arrangement with Cyber Command.
The Washington Post broke the story and The Record confirmed it through its own sources, all of whom spoke anonymously. The NSA did not respond to a request for comment and has announced nothing on the record, so the claim is that officials describe this plan, not that the agency has confirmed it. An AI mission center is also a management decision about where authority sits rather than evidence of new capability. The agency's last public AI move was joining CISA and the FBI on the advisory naming six Chinese labs it accused of distilling American models, which we covered on September 9. Building a permanent directorate around that work is how an agency signals it expects the work to be constant.
Sources: The Record · The Verge · The Washington Post
The reasoning steps a model writes out have distinct shapes inside it
We flagged this paper in yesterday's Editor's Desk and held it for more room. Researchers at KAIST and Naver AI Lab asked whether anything inside a model corresponds to the reasoning steps it writes down. They defined eight recurring operations, among them extraction, decomposition, formula recall, deduction and computation, had Qwen2.5-7B, Qwen3-8B and Gemma4-31B work through math problems, segmented the solution paths, and used GPT-5 to label each segment with an operation. The operations turned out to be reliably separable in the models' hidden states, and the separation peaks in the middle layers.
The controls rule out the obvious alternative explanations. A classifier looking only at the tokens did worse than one reading internal representations, and position in the solution path did not account for the effect either. Common function words show this clearly, since "a," "is" and "the" sit jumbled together in the early layers and then separate by surrounding operation further up, so the same token carries a different internal representation depending on which reasoning step it belongs to. When the researchers blocked attention to the preceding 30 tokens, the signal for that operation weakened, which says a step is assembled from its context rather than springing up on its own. Even on problems the models got wrong, the operation type stayed legible, and a botched computation still looked internally like a computation. The result replicated on Llama-3-8B, and for Qwen3-8B the trained classifiers transferred to GPQA-Diamond and MATH-500. The paper is accepted to EMNLP 2026 and the code is public.
Reading the chain of thought is one of the few oversight tools anyone has that does not require retraining anything, and it has been getting less trustworthy. Anthropic has published work finding that models mention the hints they actually used only 25 to 39 percent of the time, and OpenAI's Astra moves part of its reasoning into internal representations instead of text. If written steps map onto identifiable internal structure, that is at least a foothold for asking whether the visible trace matches what the model did.
This is author-run preprint work, not independent replication, on math tasks and a handful of open models. Showing that operation types are separable is not the same as showing that the written trace is faithful to the computation. The authors list error detection and mid-generation steering as future work, which is another way of saying nobody has done it yet.
Sources: arXiv (Jeong et al.) · The Decoder · Naver AI Lab code
On the Editor's Desk
Two open-weight releases came close and did not make it. AllSpark's Iris search agents have published weights and an evaluation harness, but the paper went up on September 3 and every headline number is author-run against competitors' self-reported scores under different setups, with the training pipeline still unreleased. Cognition's SWE-2 is more interesting than its press cycle, since the 50 percent score is on FrontierCode, Cognition's own benchmark, and the company's own table shows it at 27 percent on Terminal-Bench 4 against 56 percent for Fable 5.1. We would rather wait for someone other than the vendor to run the numbers.
We also skipped Xi Jinping's proposal for a BRICS AI open-source community, which is well documented but was not adopted by the summit that heard it, and a two-year study on classroom AI bans that is generating strong headlines off 66 and 164 students in a single law course, originally published in 2025 and resurfaced this month because it went open access. Held for a better version: Samsung and SK Hynix turned down KEPCO's request to prepay roughly $18.7 billion in electricity bills to fund grid buildout for Korea's chip clusters. Somebody has to pay for the power those clusters run on, and the question of who deserves better sourcing than one anonymous official behind a paywall.
