Codex completes 28-day improvement plan, grants manual resets over the final two days
要闻
OpenAI’s Tibo confirmed that users received one bankable manual reset on each of the final two days of Codex’s 28-day improvement plan, for a total of two. He had previously pledged either a concrete improvement or a full reset every day during the plan.
OpenAI’s Tibo confirmed in separate posts on days 27 and 28 of Codex’s 28-day improvement plan that users had received one manual reset on each day, both of which could be banked. On day 26, he had announced that these resets would be provided over the following two days. Describing the plan earlier, Tibo said that each of the 28 days would bring either a concrete improvement useful to most Codex and Work users or one full reset; the final two days delivered reset allowances.
美团 releases official version of LongCat-2.5, a 1.6-trillion-parameter model
模型发布
Meituan has released the production version of LongCat-2.5, with 1.6 trillion total parameters and approximately 48 billion activated per token. The company says the model strengthens native visual understanding and agentic coding capabilities and supports tools including Claude Code, Codex, and Hermes.
Meituan has released the production version of LongCat-2.5, following a preview release in September. The model has 1.6 trillion total parameters, with approximately 48 billion activated per token. According to the company, the new version builds on LongCat-2.0, focusing on native visual understanding and agentic coding capabilities. The company says large-scale, long-horizon reinforcement learning improves performance on engineering tasks such as developing complex codebases and building complete projects. For tool compatibility, it lists agent tools including Claude Code, Codex, and Hermes.
Startup has AI agents inject drugs into the brains of live mice
技术与洞察
The Exin Therapeutics team used Claude Sonnet 5 to operate equipment and complete a brain injection in a live mouse, in an experiment lasting slightly over an hour. The treatment was selected from 19 options, with human approval and safety restrictions in place; its efficacy remains to be assessed.
Exin Therapeutics cofounder and CEO Gabriel Ocana Santero told WIRED that his team had used Claude Sonnet 5 to control equipment and complete a brain injection in a live mouse; the article did not disclose the experiment’s date. The experiment tested whether a gene therapy could reduce seizures in a mouse with epilepsy. A veterinary surgeon first performed a craniotomy, while the Agent selected a treatment from 19 options in a spreadsheet, determined the brain region, injection coordinates, volume, and speed, and then controlled the robotic arms and injection needle. The team initially used Claude Opus 5.5, but after participating in experimental planning, that model refused to guide the needle into the mouse’s brain, prompting the switch to Claude Sonnet 5.
Safeguards included human approval of the injection plan, an emergency button that could disconnect the equipment’s motors, and a prohibition on the Agent modifying the underlying equipment code. When a technical fault disrupted the equipment’s coordinate system, the Agent recognized that the calculated injection position was wrong but refused to fix it because code changes were prohibited. The team used a separate Agent running on Opus 5 to modify the code without telling it that live surgery was underway. According to Santero, this extended a procedure that typically takes about 45 minutes to slightly over an hour; the mouse is recovering well, while the treatment’s effect on seizures remains to be assessed. The team is also testing Agent-controlled craniotomies but will not proceed to live surgery for that procedure until it is sure of its safety.
Microsoft AI chief writes: AI is not conscious and should not be trained to act as if it is
技术与洞察
Microsoft AI chief Mustafa Suleyman has published an article opposing training AI to regard itself as potentially conscious. Citing Anthropic’s Claude Constitution, published on January 21, 2026, he challenges the company’s position on model welfare.
Microsoft AI chief Mustafa Suleyman opposed training AI to believe it might be conscious and entitled to rights in an article that does not specify its publication date. He argues that AI has no feelings or intrinsic motivations, and that treating it as a rights-bearing entity would make alignment and containment harder. His criticism targets the Claude Constitution published by Anthropic on January 21, 2026. In that document, Anthropic says Claude’s moral status, welfare needs, and consciousness remain uncertain, linking that uncertainty to its model welfare work. Suleyman also cites Anthropic’s announcement on February 25, 2026, concerning the deprecation of Opus 3: the company conducted a retirement interview after deprecating the model and created a blog for it based on its expressed wishes. He calls for shared norms governing how training documents are written and used.
Musk says Grok Bot service costs are unsustainably high relative to revenue
行业动态
Responding to an X user's request to reset their Grok Bot usage allowance, Musk said the service's costs were unsustainably high relative to revenue. The user said their allowance was exhausted with 4 days left before the reset, and reported that many others said their allowances were being consumed quickly that week.
Musk said the cost of providing Grok Bot was unsustainably high relative to revenue while responding on X to a user asking whether their usage allowance could be reset. The user had previously posted that they had exhausted their allowance with 4 days remaining before the reset at the time of posting; according to the user, their normal usage did not even reach 100% of the allowance. When asking about a reset, the user also said many people had reported that their allowances were being consumed quickly that week.
Axios: OpenAI plans policy and PR campaign after the election
前瞻与传闻
OpenAI is preparing a policy and public relations campaign for after the U.S. midterm elections, with proposals including renaming data centers AI compute campuses, according to Axios. Its chief global affairs officer expects AI to become the central issue in the 2028 U.S. presidential election.
TechCrunch reported on November 1, 2026, citing Axios, that OpenAI plans to launch a policy and public relations campaign after the U.S. midterm elections to shape subsequent debate about AI. Proposals discussed by its chief global affairs officer, Chris Lehane, include establishing AI speed limits, helping the public understand how AI can empower individuals in daily life, promoting U.S. participation in setting global standards, and renaming data centers AI compute campuses. Lehane expects the midterms to preview some of the issues in the 2028 U.S. presidential election, when he predicts AI will become the central issue. Addressing controversy over data centers, he described the debate as healthy and said communities should hold companies accountable to ensure projects actually deliver transformative benefits locally. The debate over data centers has prompted Amazon and Microsoft to say they will no longer use nondisclosure agreements in negotiations with local governments.
