Tuesday, 6 October 2026
Asked for a Number, Nobody in the Room Had One
The Hearing Happened, and the Four Companies That Came Could Not Put a Figure on Catastrophic Risk
We covered the New York City Council’s AI hearing yesterday as a thing about to happen — fifty-one members in one room, four companies sending representatives, three of them only after the Council mentioned its subpoena power. Here is what came out of it. Asked by Speaker Julie Menin to quantify the worst case, OpenAI’s head of policy development and operations, Morgan Dwyer, said: “I don’t know. I also don’t think it matters whether it’s 1% or 10% or 20%.” None of the four — OpenAI, Anthropic, Google, Meta — would commit to automatically withdrawing a model that failed an independent safety test, and none gave a straight answer on who bears legal responsibility when a system causes serious harm. “I think it’s disappointing that they can’t seem to answer some of the most basic, rudimentary questions,” Menin said afterwards, adding that on the question of pulling a failing model, “it would be far clearer if we could get clear yeses or no’s.” The sharpest testimony of the day came from someone with no company to protect: Jacob Coxon, a former researcher at both OpenAI and Anthropic, told the Council that “on the current path, I think it is more likely than not that humanity loses control of these AIs.” SpaceXAI, subpoenaed on 28 September, did not appear, and the Council is now going to court over it. The companies spoke at length about their internal reviews. The one person who gave the Council a number does not work at any of them.
OpenAI Will Watermark Its Text in Europe, and Published the Number That Shows How Easily It Rubs Off
Under the EU AI Act’s transparency rules, OpenAI set out its text-watermarking approach yesterday: a method it calls textGrain, which subtly shapes the model’s word choices so that a detector can recognise the pattern and a reader cannot see it. It rolls out to eligible ChatGPT and Codex users on every plan across the EU over the coming weeks, and API customers worldwide can switch it on immediately — off by default. OpenAI also opened applications for outside access to its detector, limited at first to approved researchers and expert organisations. What makes the announcement unusually honest is the figure buried in it: replace ten per cent of the words with synonyms and detection falls from 92 per cent to 66. Short passages, maths answers and translated text are harder still, and OpenAI states plainly that a missing watermark proves nothing about who wrote a thing. So the first concrete compliance product built for the AI Act ships with its own defeat conditions documented by the company shipping it. That is either refreshing candour or a very carefully managed expectation, and it is probably both.
Source: TechCrunch, October 5, 2026
Reflection Released Its First Frontier Model Open-Weight, and the Benchmark It Chose to Be Measured Against Is Chinese
Beam went out yesterday: a 501-billion-parameter mixture-of-experts model with 23 billion active, pretrained on 23.8 trillion tokens, a million-token context window, aimed squarely at reasoning, coding and agentic work. Reflection — Nvidia-, Sequoia- and Lightspeed-backed, roughly $4.7 billion raised, last valued at $25 billion pre-money — says Beam matches the leading Chinese open models on advanced reasoning benchmarks, scores on par with Z.ai’s GLM-5.2, and beats the best Western open models while using three to four times less inference compute. Weights and the technical report are due later this month through hyperscalers and neoclouds, which is when anyone outside the company will be able to check. Read the claim structure rather than the numbers: a well-funded American lab announcing its strongest model describes success as parity with a Chinese release, and Nathan Lambert, who has shipped frontier open models himself, flagged Reflection on X as the third Western lab in a row — after Nvidia and Thinking Machines — to land behind Chinese open weights with its best work. The compute-efficiency figure is the interesting part and the hardest to verify. Three weeks from now we will know whether it survived contact with other people’s hardware.
Source: TechCrunch, October 5, 2026
The Senate Armed Services Committee Has Written Down That AI Does Not Get the Nuclear Decision
Provisions added to this year’s annual military policy and spending bill would require human oversight of AI-controlled weapons, ban the use of AI in a decision to launch or detonate a nuclear weapon, ban AI for domestic surveillance, and oblige the Pentagon to record which decisions were made by or with machine assistance. The hearing testimony behind them is the part to read. The Maven Smart System identified targets at ten times the previous speed during the opening of the Iran war — a speed that reporting now connects to a strike that destroyed an Iranian school and killed more than 160 people. Separately, an Air Force special-operations analyst used a chatbot to draft an intelligence report that falsely claimed a Chinese vessel was carrying nuclear-weapons components; officials caught it before the operation it would have justified. Senator Mark Kelly’s question in committee was whether a human has to make the final decision to strike, or whether a system can execute the engagement itself. Retired Lieutenant Colonel Joe Chapa put the unglamorous version: everyone who touches one of these systems “needs at least baseline education in how statistical AI works.” All of this is a committee markup, not law. And it arrives the same week Defense Secretary Pete Hegseth stood up a new Autonomous Warfare Command to speed the same technology into the field.
Source: WUNC, October 5, 2026
Cohere’s Biggest Agent Release Yet Is Largely a Set of Ways to Stop the Agents Spending
North 2 landed yesterday with the features you would expect of an enterprise agent platform in its second generation: memory that persists between sessions so nobody re-explains the situation every morning, a rebuilt harness for multi-step tasks with human sign-off, reusable skills, shared document libraries, connectors into Slack, GitHub, SharePoint, OneDrive and Jira, and deployment on-premises, in a customer VPC or fully air-gapped. Then there is North Admin, and this is the half that tells you what the past year of production deployments was actually like: token consumption tracked per user and per agent, consumption tiers by request and token rate, organisation-wide caps, and alerts that fire before a limit rather than after it. Cohere’s own account of why is admirably blunt — a year of large-scale rollouts across finance and the public sector found enterprises sacrificing one of security, capability and cost control to get the other two. Nvidia’s Kari Briski supplied the quote about delivering “more tokens for less.” The telling detail is that the headline feature of a 2026 agent platform is a spending cap.
Source: SiliconANGLE, October 5, 2026