Everything that mattered in AI, one page a week.

Moonshot AI published Kimi Linear on Thursday: a 48B-parameter mixture-of-experts model that activates 3B parameters per token, carries a 1M-token context window, and was trained on 5.7 trillion tokens. The architecture interleaves three Kimi Delta Attention layers — a channel-wise-gated refinement of Gated DeltaNet — for every full multi-head latent attention layer. Moonshot says that 3:1 ratio cuts KV-cache demand by up to 75% and reports 1.84 ms versus 11.48 ms per output token against MLA at 1M tokens, a 6.3x decode advantage. Checkpoints and the KDA kernel are open-sourced.

The claim worth taking seriously is not the speed table but the quality comparison: the report argues this is the first hybrid linear attention design to beat full attention under fair comparison across short-context, long-context and RL-scaling regimes, with MMLU-Pro at 51.0 against MLA's 47.2 and RULER at 128K at 84.3 against 81.3. Linear attention's historic failure mode is exact recall — a fixed-size state cannot copy arbitrary earlier tokens — so the interesting question for builders is where that trade still bites, not whether the benchmark moved.

OpenAI released a research preview of gpt-oss-safeguard on Wednesday: open-weight reasoning models at 120B and 20B, fine-tuned from gpt-oss and published under Apache 2.0. Instead of a fixed classifier trained on labeled examples, the model takes two inputs — a developer-written policy and the content — and reasons its way to a label, which means policies can be changed without retraining. OpenAI is explicit about the trade-offs: a classifier trained on tens of thousands of labels can still outperform it, and it is compute-intensive enough that OpenAI runs smaller high-recall classifiers in front of its internal Safety Reasoner.

The next day, Chain-of-Thought Hijacking showed what happens when reasoning is the mechanism you are trusting. Wrapping a harmful request inside a long benign puzzle-solving trace — often five minutes of reasoning — defeated refusals 99% of the time on Gemini 2.5 Pro, 94% on ChatGPT o4 Mini and Claude 4 Sonnet, and 100% on Grok 3 Mini; activation probing suggests refusal depends on a low-dimensional safety signal that weakens as traces lengthen. Reasoning that reads a policy is also reasoning that is long enough to forget one.

OpenAI completed its recapitalization on Tuesday, converting the for-profit into a public benefit corporation called OpenAI Group PBC, controlled by the renamed OpenAI Foundation. The Foundation holds equity currently valued at about $130 billion, a 26% stake; current and former employees and investors hold 47%; Microsoft holds roughly 27%, valued at about $135 billion. The reworked Microsoft agreement extends Microsoft's IP rights through 2032 — now including post-AGI models — drops Microsoft's right of first refusal to be OpenAI's compute provider in exchange for an incremental $250 billion Azure purchase, and requires any AGI declaration to be verified by an independent expert panel.

The post from board chair Bret Taylor notes the deal closed after nearly a year of dialogue with the California and Delaware Attorneys General — the regulatory follow-on to the 2024 for-profit conversion that OpenAI abandoned in May 2025. Microsoft's own summary of the agreement carries a line developers should read twice: OpenAI is now able to release open weight models that meet requisite capability criteria. Read alongside the same week's model drops, Monday and Wednesday stop looking like generosity and start looking like contract terms.

MiniMax open-sourced M2 on Monday: a mixture-of-experts model with 230B total parameters and 10B activated per token, built for tool use rather than conversation. It carries roughly a 200K-token context window, and MiniMax lists API pricing at $0.30 per million input tokens and $1.20 per million output, which the company puts at about 8% of Claude Sonnet 4.5's price at roughly double the inference speed. Weights are on Hugging Face with day-0 serving paths in vLLM, SGLang and MLX.

Reported benchmarks put M2 at 69.4% on SWE-bench Verified and 46.3 on Terminal-Bench, with the top composite score among open-source models on Artificial Analysis' index. The number that decides whether you can use it is the 10B activation: 4-bit weights land near 130GB, so a 2x80GB node serves long agent loops at interactive speed. One operational catch — M2 is an interleaved-thinking model, and its thinking blocks must be preserved in message history or quality degrades.

Brave published its second agentic-browser disclosure on October 21, extending its Comet prompt-injection work into hidden-text and screenshot vectors and reporting the same class of flaw across the entire category of AI browsers, with one further vendor's issue withheld at that vendor's request. The mechanism is unchanged: the browser forwards page content to the model without separating trusted user instructions from untrusted site text, so summarizing a hostile page or Reddit comment executes attacker text with the user's authenticated session as the blast radius. Traditional same-origin policy and CORS do nothing here, because the request is made by the assistant, not the page.

OpenAI did not contest the class of attack. In an October 22 post, chief information security officer Dane Stuckey described red-teaming, reward shaping to ignore malicious instructions, overlapping guardrails and new detection systems, then said plainly that prompt injection remains a frontier, unsolved security problem — a launch-week admission that Atlas ships with a known unfixed vulnerability class. Researchers reproduced injections against Atlas the same day, and by October 24 a clipboard-hijack demo showed agent mode clicking a malicious link and silently overwriting the user's clipboard.

Meta told employees on October 22 that it is cutting roughly 600 positions across Superintelligence Labs, the umbrella covering its AI infrastructure teams, its product AI groups and Fundamental AI Research. The cuts, announced in a memo from chief AI officer Alexandr Wang, do not touch the small group building Meta's next-generation foundation models — many hired this summer at compensation in the hundreds of millions — leaving Superintelligence Labs at just under 3,000 people. Affected staff get 16 weeks of severance plus two weeks per completed year of service.

The timing is the story. Reuters noted Meta had struck a $27 billion financing deal with Blue Owl Capital one day earlier — its largest-ever private capital agreement — to fund its biggest data center project. So the same week saw record infrastructure borrowing alongside a roughly 20% haircut to the research organization, which Meta attributes to organizational bloat accumulated during three years of fast AI hiring and a lukewarm developer reception to Llama 4 in April. Meta is keeping the newly hired researchers and cutting the legacy layers around them.

Alibaba's Qwen team completed the Qwen3-VL rollout on October 21 with the 32B and 2B dense models, in both Instruct and Thinking editions. The 32B is a dense 32B-parameter model — no mixture-of-experts routing, so you pay for all 32B per token — with an enhanced ViT encoder, interleaved-MRoPE positional encoding for interleaved image, video and text sequences, DeepStack multi-level feature fusion, native 256K context expandable to 1M, OCR across 32 languages, and second-level video timestamps. Benchmarks put the Thinking edition at 78.1 on MMMU, and the full line-up — 2B, 4B, 8B, 32B dense plus 30B-A3B and 235B-A22B MoE — ships under Apache 2.0 on Hugging Face with no gating.

The deployment maths is the point: the 32B fits a single 24GB card at Q4_K_M quantization, about 19.8GB, and the 2B runs on edge hardware, so multimodal agentic work that previously required a metered API — GUI element recognition, tool invocation, screenshot-to-code, document and long-video QA — can now run on hardware you already own, with no per-token bill and no data leaving the machine. The same window brought DeepSeek-OCR, an MIT-licensed 3B MoE with roughly 570M active parameters that showed 97% OCR fidelity at roughly 10x text-to-vision-token compression.

OpenAI launched ChatGPT Atlas on October 21, a Chromium-based browser built around ChatGPT rather than a search box. It runs worldwide on macOS for Free, Plus, Pro and Go accounts, with beta access for Business and, if a plan administrator enables it, Enterprise and Edu; Windows, iOS and Android builds are promised later. Three pieces matter technically: an Ask ChatGPT sidebar that reads the current page, optional browser memories that persist context across sessions and can be toggled per-site from the address bar, and agent mode — a preview for paid tiers — that takes actions in the user's own logged-in browser session instead of a remote sandbox.

The architectural claim is that the browser is the only surface holding your tabs, sessions and context at once, and OpenAI extended that logic the same week by acquiring Software Applications Incorporated, makers of the unreleased Mac agent Sky, and folding all twelve staff into ChatGPT. For anyone building web apps the practical consequence lands in two places: published ARIA tags on your pages now change how agent mode behaves on your site, and page visibility, memory and training opt-in are user- and publisher-controlled settings rather than server-side defaults you can assume.