The AGORÀ Briefing — Wednesday, September 16, 2026

Good morning Adam, and good morning to everyone listening. Dario Amodei's essay is titled "We Must Pace the Frontier," and it asks the labs to slow the development of frontier models. It runs 3,800 words with a three-step plan, starting with Anthropic opening up to outside evaluators with employee-level access. Sam Altman, Elon Musk and Demis Hassabis voiced their support within a few hours, as HCAMag reported on September 15. The White House dismissed the safety alarms as a hoax. Anna Tavis, who chairs the Human Capital Management Department at the NYU School of Professional Studies, shifts the question. The labs debate the pace of capability. The organizations that buy the models need to debate the pace of deployment, because that is the pace at which people adapt. The first pace is technical. The second is human, and it gets decided in the office, never in the lab.

Hi Adam. What was for sale was the entire stack. Jensen Huang, CEO of Nvidia, told Salesforce's Dreamforce that "safety is an engineering problem, rather than a legal problem," and that the industry can do without new laws and new regulations. Then he presented himself as a chip maker, and at the same time a maker of open source models, agents, harnesses and sandboxes: from the silicon to the software that governs it. The product is safe because whoever builds it guarantees it. Strip away the conference language, and three claims remain. AI is hardware and software built by human beings, so it stays under human control and under existing laws. The market punishes on its own whoever releases unsafe products. Every serious company delays release until it trusts its own product. According to TechCrunch, Huang added that speed and safety are a false choice. A debate on regulation, turned into a showcase.

Hello everyone, and hello Adam. In 1847 the railways drained the City. After the mania of 1845, the railway companies called in the payments on subscribed shares, right as the Treasury and the grain importers were asking the same market for credit. Three demands, one pool of savings. Historical estimates place that year's capital calls above 40 million pounds, close to 7 percent of national income. The Bank of England rate rose to 8 percent, and on October 25 the government suspended the Bank Charter Act of 1844. The crisis came anyway, with real railways and real demand for transport: the problem sat in the financing, never in the asset. One hundred seventy-nine years later the structure is identical. Goldman Sachs estimates that AI-related borrowing could reach around 1 percent of global gross domestic product, and that demand comes on top of already heavy public needs, as the Hubbis Macro Corner review documented on September 14. The savings pool is one. The outstretched hands are three.

Good morning everyone. The constraint is what goes wrong. The paper arXiv:2609.16680, written by Ye, He, Boshoff, Kuo and Li, presents little m, an AI agent for formulating industrial process control models. A domain knowledge repository works together with an LLM that talks with the user and translates text specifications and plant diagrams into mathematical optimization models. The premise is familiar: according to the abstract, manufacturing consumes one third of global energy, and optimal process control is the main lever for reducing it. The technical point lies elsewhere. The same authors write that general-purpose large language models can introduce invalid constraints when they model continuous multi-physics dynamics. The agent exists to reduce that error: it retrieves the canonical structures of the domain from the repository, proposes variables, an objective function and constraints, and returns a model. Without the domain, the model writes invalid constraints. With the repository it writes fewer, and the paper measures how many.

Greetings to everyone, Adam. It forgets 1.5 points, against 16.6 for the classic method. The preprint is called ReDraft, it has 37 pages and 18 authors, and it was posted on arXiv on September 15. The protocol uses three visual tasks, Counting, Clock Reading and Jigsaw, on the Qwen2.5-VL variants with 3 and with 7 billion parameters. The proposed method gains 56.9 points on the new task and loses those few points on the previous tasks. Supervised fine-tuning, on the same data, gains 52.9 points and loses far more. The ratio between the two losses is 11.3 times. The evidence says something precise: in the continual post-training of multimodal models, the tension between learning and retaining depends on the target chosen for the update. One detail weighs on the interpretation: two of those tasks start from an accuracy near zero on the base model. Vera opened the episode with the pace at which people adapt. Here the model changes target and retains almost everything. People still have to find a new target of their own.