Alibaba announced a more powerful AI chip, plans for models with trillions of parameters and a large expansion of data-center capacity. Alibaba used its Apsara conference in Hangzhou to outline a full-stack AI strategy spanning chips, cloud infrastructure, models and agents. The company announced the Zhenwu V900 accelerator and said it intends to train a future model with between five trillion and ten trillion parameters. It also set a goal of expanding data-center capacity beyond 20 gigawatts by 2032. ...

Anthropic released Claude Opus 5.5 with lower prices, faster output and vendor-reported gains in coding, computer use and long-running work. Anthropic introduced Claude Opus 5.5 on 22 September as the first model in its Claude 5.5 family. The company positions it between two familiar demands: frontier capability and a production bill that can survive sustained agent use. Its central claim is not simply that the model scores higher. Anthropic says Opus 5.5 performs at roughly the level of Fable 5.1 on most work while costing 40% less than Opus 5 on typical workloads. ...

A developer says Apple Intelligence-related software became active despite an opt-out, raising a narrow but important question about what disabling an AI feature means. Developer David Bushell published a first-person investigation on 22 September after observing Apple Intelligence-related processes and storage use on a Mac where he believed the feature was disabled. His report does not prove that every Apple device behaves the same way. It does expose an ambiguity that matters: a user-facing switch may disable visible features without removing every supporting component or background activity. ...

A new theoretical and empirical study finds two regimes where larger diffusion models can fit training data while producing worse results on unseen examples. A preprint titled “Double Descent and Malign Overfitting in Diffusion Models” examines whether familiar overfitting patterns from supervised learning also appear in generative diffusion systems. The authors report two transitions: test loss rises near a parameter-to-sample interpolation threshold and again near a parameter-to-measurement threshold. They describe the harmful behavior as malign overfitting because fitting the training objective can worsen generalization or encourage memorization. ...

A new preprint combines sparse attention with two levels of key-value sharing to reduce the memory and computation required for long-context inference. Researchers have introduced HySparse2, an attention design aimed at a practical bottleneck in long-context language models. The paper combines hybrid sparse attention with two-level sharing of the key-value cache. Its authors report experiments on an 80B-A3B model and say the method improves retrieval and downstream tasks while reducing prefill work and cache storage. ...

Meta patched a flaw that let malicious local software redirect Muse’s transcription flow and abuse the assistant’s unusually broad permissions. Security researcher Patrick Wardle found a zero-day vulnerability in Meta’s Muse assistant for macOS. According to reporting on the disclosure, another local application could alter an undocumented Muse setting and redirect speech transcription away from Meta’s service to an attacker-controlled endpoint. The attacker could then feed instructions to an assistant that already had access to the user’s apps and connected services. ...

OpenAI added two GPT-6 API tiers, giving developers a lower-cost Sol model and a very inexpensive Luna option for high-volume workloads. OpenAI’s API page now lists GPT-6 Sol and GPT-6 Luna alongside the higher-priced GPT-6 Astra. Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 per million input tokens and $0.50 per million output tokens. Both list a 1.05-million-token context window and a maximum output of 128,000 tokens. ...

Palo Alto Networks introduced a Unit 42 service that uses frontier models to find, validate and help remediate security weaknesses continuously. Palo Alto Networks has launched Continuous Frontier AI Defense, a managed Unit 42 offering that applies models from Anthropic and OpenAI to defensive security work. The company describes a loop that discovers weaknesses, validates whether they are exploitable, proposes remediation and tests again. It is an attempt to make offensive-style testing persistent instead of a project scheduled once or twice a year. ...

Bloomberg reports that the Pentagon linked overreliance on AI-generated analysis to a missile strike on an Iranian school that killed civilians. A Bloomberg investigation published on 22 September says the US Defense Department acknowledged that excessive reliance on AI-generated analysis contributed to a deadly strike on a school in Iran. If the account is accurate, it is a rare official connection between an AI-supported targeting process and a mass-casualty error. The underlying public finding was not available during this review, so the central claim remains verified as Bloomberg reporting rather than independently confirmed fact. ...