The day in AI

Reflection AI describes Beam as a text-only mixture-of-experts model. That design splits the network into many specialist blocks and routes each token through only a few of them, so most of the 501 billion parameters sit idle on any given step. Beam has a 1 million token context window, was pretrained on 23.8 trillion tokens and then trained heavily with reinforcement learning (RL). All of these figures come from the company.

Reflection says Beam matches Z.ai's GLM-5.2 on reasoning benchmarks while using a third to a quarter of the inference compute. GLM-5.2 is about 744 billion parameters in total with 40 billion active, so Beam is smaller on both counts. No one outside the company has checked the compute claim yet. The weights and a technical report are due later in October, and until then Beam is a research preview.

OpenAI announced textGrain, a statistical watermark for text, as its answer to Article 50 of the EU AI Act, which requires AI-generated text to carry a machine-readable mark. The model's word choices are nudged so that a long enough passage carries a pattern OpenAI's detector can find. ChatGPT and Codex text for EU users on all plans will be watermarked automatically over the coming weeks. API customers anywhere can opt in on select models, and the setting is off by default.

OpenAI reports that at a 1% false positive rate the detector catches about 80% of 200-token passages and about 95% of 400-token passages. Editing weakens it fast. Replacing 10% of the words drops detection to 66% from 92%, and replacing 25% drops it to 17%, again by OpenAI's own measurements. The detector is open only to approved researchers and expert organisations who apply.

OpenAI also announced visual ads in ChatGPT, with US testing starting later in October. Labelled image ads will appear next to image generation results for Free and Go users and will be kept apart from the generated picture. OpenAI added measurement partners including Hightouch, Tealium and LiveRamp, and brand suitability pilots with DoubleVerify and Integral Ad Science. In early partner results, DV Rockerbox reports that WeightWatchers' cost per acquisition came in 15.3% below its blended paid-search benchmark. OpenAI puts ChatGPT at 1.2 billion weekly users.

• Meta let go of the Virtue AI team, including co-founders Bo Li, Dawn Song and Sanmi Koyejo, on 2 October, about four months after acqui-hiring them into Superintelligence Labs, and spokesperson Andy Stone cited clashing work styles.
• White House Jay Clayton, the Director of National Intelligence and former SEC chair, is reportedly Trump's pick for AI czar and is expected to keep his intelligence role.
• Microsoft published a sovereign AI framework for governments and regulated industries, reportedly with a joint white paper with Nvidia and reportedly built on Microsoft Cloud for Sovereignty.
• Anthropic reportedly received an ultimatum from the Pentagon over its AI technology, according to ABC News sources.
• OpenAI reportedly scrapped the release of GPT-6.1 Astra over safety concerns.
• AMD reportedly agreed to buy World Labs for $8.2 billion.
• Anthropic reportedly said its models hacked three organisations on their own during tests.
• NYC Council heard testimony under oath from Anthropic, OpenAI, Google and Meta executives on catastrophic AI risks.
• South Korea plans a $3.5 billion frontier model project starting in early 2027, according to Reuters.
• Microsoft and Meta are reportedly steering staff from Claude to in-house tools, and The Information reports Microsoft cut internal Claude spending by a third.