Today's AI news: Reflection's 501B open model Beam, ChatGPT ads launch, EU watermarking, Opus 5.5 science discovery, and a continual-learning research breakthrough.
Here is what happened in AI today — Monday, 6 October 2026.
Reflection AI Launches Beam: A 501B Open-Weight Model
The biggest model news of the day comes from Reflection AI, which has introduced Beam — a 501-billion-parameter open-weight language model. The announcement attracted significant attention on Hacker News, where it ranked among the day's most discussed posts. Open-weight models at this scale are rare, and Beam's release is likely to draw immediate interest from researchers and developers who want to run or fine-tune a frontier-scale model without proprietary restrictions. Reflection AI's blog post formally introduces Beam and outlines its capabilities, though full technical details are still being digested by the community.
OpenAI Launches Ads in ChatGPT
OpenAI has announced a new visual advertising format inside ChatGPT, alongside expanded measurement tools, attribution partnerships, and brand-suitability controls for advertisers. The move marks a significant commercial shift — ChatGPT, which has operated as a subscription and API business, is now opening a third revenue stream through advertising. OpenAI says the format is designed to match the way people actually use AI assistants, rather than replicating traditional web-display advertising. Brand-suitability tools suggest OpenAI is also trying to reassure advertisers about content context.
OpenAI Sets Out Its Approach to EU Text Watermarking
Also from OpenAI today: the company has published details of how it plans to handle EU rules on text provenance. The post explains where watermarks will apply to AI-generated text, how detection will work, and why access to watermark-detection tools will initially be limited to researchers. The move reflects growing regulatory pressure in Europe for AI systems to label or authenticate machine-generated content, and gives OpenAI's first clear public statement on how it intends to comply.
Opus 5.5 Agents Identify Two Room-Temperature Magnetic Semiconductor Candidates
In a notable AI-for-science result, the team at Vals AI reports that Anthropic's Opus 5.5 agents have autonomously identified two candidate room-temperature magnetic semiconductors. This is a materials-science target of long-standing interest — such materials could underpin a new generation of electronic and spintronic devices. The agents apparently worked through the discovery process without direct human guidance for each step. With 237 Hacker News points at time of writing, it is one of the most-upvoted stories of the day, signalling strong community interest in AI-driven scientific discovery.
Research: Off-Policy Merging Beats On-Policy Self-Distillation for Continual Learning
A new paper on arXiv challenges a widely held assumption in AI training: that on-policy data is a prerequisite for continual learning in large language models. The researchers introduce a method they call grafting, which learns a weight update from new (off-policy) data on an earlier checkpoint and then applies — and scales — that update to the post-trained model, optionally masking the most sensitive directions. Tested across three continual-learning scenarios — distilling from expert traces, self-improvement via STaR and Pedagogical RL, and injecting knowledge after a pre-training cutoff — grafting Pareto-dominates both standard supervised fine-tuning and on-policy self-distillation on both new-task and old-task performance, while avoiding the cost of on-policy sampling. The result is a meaningful step forward for keeping models up to date without catastrophic forgetting.
Two-Year Study Finds Mixed Results for AI Tutoring with Khanmigo
A newly published working paper from the Education Working Papers series examines a two-year school experiment using Khanmigo, Khan Academy's AI tutoring tool. The study offers one of the longer-run empirical looks at AI tutoring in real classroom settings. While the full findings are nuanced, the paper's publication adds important evidence to the ongoing debate about where and how AI tutoring tools deliver genuine learning gains — a question of growing relevance as schools worldwide consider AI adoption.
Sources and links are listed below. All information is drawn exclusively from today's published reports and announcements.
Frequently Asked Questions
What is Beam and who made it?
Beam is a 501-billion-parameter open-weight language model released by Reflection AI on 5–6 October 2026. It is notable for being one of the largest openly available models at frontier scale, attracting significant attention from the AI research and developer community.
Is ChatGPT now showing ads?
Yes. OpenAI announced on 5 October 2026 that it is introducing a new visual advertising format inside ChatGPT, along with expanded measurement tools, attribution partnerships, and brand-suitability controls for advertisers.
What did Opus 5.5 agents discover in materials science?
Anthropic's Opus 5.5 agents autonomously identified two candidate room-temperature magnetic semiconductors, according to a report published by Vals AI. Room-temperature magnetic semiconductors are a long-sought target in materials science with potential applications in electronics and spintronics.
What is the grafting method for continual learning in AI?
Grafting is a technique proposed in a new arXiv paper (October 2026) that learns a weight update from off-policy data on an earlier model checkpoint and then applies and scales that update to the post-trained model. It optionally masks sensitive update directions when new data is far from the post-trained distribution. In tests, grafting outperformed both standard supervised fine-tuning and on-policy self-distillation for continual learning, while avoiding expensive on-policy sampling.