Codex lead Tibo said the global reset for all paid ChatGPT accounts is now in effect. He also said the slowdown caused by a load surge during the first two days after GPT-6.1 Sol launched has ended and the model is back to its expected speed.
Codex lead Tibo said the global reset for all paid ChatGPT accounts has now fully taken effect. He had previously said the reset would be implemented at 10:00 a.m. PST on October 2, though the source text does not specify the year. According to his explanation, GPT-6.1 Sol ran slowly during the first two days after its release because of a surge in load. He apologized for the issue and said the model has now returned to its expected speed.
Starting on October 9, 2026, Google will change Gemini Apps for personal accounts so that users without an AI plan can access only Flash-Lite. AI Plus users will be notified by email of the date when the changes apply to them.
Google announced that it will change model access in Gemini Apps for personal accounts on October 9, 2026, after which users without an AI plan will be limited to Flash-Lite. According to the company, AI Plus users will receive an email specifying when the changes take effect for them. Gemini Apps had previously changed usage limits and model availability for users aged 18 or older on May 17, 2026, and extended those changes to users under 18 on July 24, 2026. Google states that Gemini usage limits are calculated based on compute and refresh every 5 hours until the weekly limit is reached. Usage is determined by the complexity of the prompt, the features used, and the length of the chat. Paid users have higher limits than users without a Google AI subscription, while Premium models and features consume more usage and may cause users to reach their limits sooner.
Andrej Karpathy shares tips for understanding language model outputs
要闻
Andrej Karpathy said that as language models improve, human work will shift further toward supervising and understanding their output. He proposed using ASD-STE100 controlled language for explanations, requiring only 80% adherence when needed, and generating charts, interactive web pages, and custom explainer videos.
In a post on X, Andrej Karpathy argued that as language models become more capable, human work will focus increasingly on supervising and understanding model output, while models can also generate custom content to explain their results. He proposed having models explain topics in ASD-STE100 controlled language. The specification was originally designed for aerospace maintenance documentation and imposes strict requirements, so he sometimes asks for only 80% adherence. He also suggested generating charts and interactive web pages, and said he is most optimistic about explainer videos customized for any topic. According to Karpathy, this approach has already begun to work. He believes that as intelligence and code become more abundant, models can generate larger one-off custom products such as web applications and explainer videos, which previously were often not worth producing specifically for a single use.
Anthropic developer account ClaudeDevs announced that You should Know has been added to Claude Code. The plugin uses a sideagent to inspect Claude’s output, flag important information users may miss, and can be enabled with a specified command.
Claude Code has added the You should Know plugin, according to an announcement from Anthropic developer account ClaudeDevs. The official account said the plugin scans output generated by Claude and identifies important information that users may overlook but that could help them track current progress. It is implemented as a mod and spawns a sideagent at runtime to observe Claude’s output. Users can enable the plugin in Claude Code by running /plugin enable cc-plugin-you-should-know@builtin.
CodeBuddy exclusively integrates Space Bunny in China
开发生态
An article says CodeBuddy and WorkBuddy have exclusively integrated Space Bunny in China and are offering it as a built-in model. It says a limited-time 0.03x multiplier discount will be available from October 2, 2026, through October 7, 2026.
According to a WeChat public account article, CodeBuddy and WorkBuddy have exclusively integrated Space Bunny in China and are providing it as a built-in model, although the article does not give a specific integration date. It describes Space Bunny as an anonymous Flash large model and says it natively supports text, image, and video inputs, offers adjustable thinking depth, and has a maximum context length of 1M. The article also claims fast inference and strong coding capabilities. For the promotion, it says a limited-time 0.03x multiplier discount will be available from October 2, 2026, through October 7, 2026.
OpenCode announced that inclusionAI’s Ling-3.1-flash is now available free of charge on its platform. According to the official statement, the model has 560B total parameters, 25B active parameters, and a 262K context window during the free-access period.
OpenCode announced that inclusionAI’s Ling-3.1-flash is now available free of charge on the OpenCode platform. OpenCode described Ling-3.1-flash as inclusionAI’s latest model. According to the disclosed configuration, Ling-3.1-flash has 560B total parameters and 25B active parameters; the context window for this free release is 262K, with that specification applying during the free-access period.
Cua Spaces launches on macOS as a sandbox app for AI agents
开发生态
Cua launched Cua Spaces, a desktop app for AI agents, on macOS on the day of its announcement. Available as a free download with source code, it creates sandboxed virtual computers that agents can operate directly.
Cua announced the launch of its Cua Spaces desktop app on X, and the product became available for macOS on the same day. It can be downloaded free from Cua’s website, and its source code is available. The app creates and manages Spaces, each a sandboxed virtual computer running cua-spacesd that an AI agent can operate directly, with required permissions granted and tools installed in advance. macOS images are built locally, while Linux images can also run. Teleport moves an existing local app session into a Space so the agent does not need to log in again. User confirmation is required before the system reads any content, Touch ID is required for sensitive items, and transferred data is locked in the local Keyvault by default. Users can specify which sites and files are provided to each Space and delete everything else. Users and agents share the same desktop but have separate cursors, allowing users to take over or return control at any time. A Space can run on a local Mac, a privately owned Mac mini or Linux machine, or connect to the user’s own cloud environment. All Spaces can be listed together and viewed through live desktop streaming.
Meta open-sources the Muse Gadgets hardware project
开发生态
Meta has released Muse Gadgets, an open-source hardware project offering SDKs and firmware for ESP32 and Linux devices so developers can connect the personal AI assistant Muse to various components. A separate batch of 5,000 Home Link units is scheduled to ship in several weeks.
Meta launched Muse Gadgets and released the project’s code under the Apache 2.0 license. Developers can use its SDKs and accompanying firmware for ESP32 and Linux devices to build hardware connected to the personal AI assistant Muse. The project supports connecting Muse to displays, buttons, sensors, and actuators, while the code is provided as is and without warranty. Each custom-built device requires an SDK token obtained from the official website, after which the developer must enable developer mode in the iOS or Android Muse app and pair it with a device whose name begins with MuseGadget. Nat Friedman, product lead at Meta Superintelligence Labs, also announced the USB-C-powered Home Link. According to his description, the device connects Muse to a home network and allows it to communicate with televisions, speakers, and other devices. An initial batch of 5,000 units is expected to ship in several weeks, and Muse subscribers can claim one for free through the official website while supplies last.
llama.cpp adds support for decision models, retaining the System One API format
开发生态
llama.cpp server has added decision-model support. After receiving a text, JSON, or screenshot state and typed questions, it returns the probability of each option in a single forward pass through an interface compatible with the System One format.
llama.cpp server now supports decision models through the /v1/systemone interface. The API uses the System One format introduced with TypeSafe’s Jev model, so existing clients only need to change their base URL. A request contains one state and one or more questions, with 3 typed question categories. The state can be text, JSON, a screenshot, or a list of chat messages, and a data URL in an image_url part is read as an image.
A decision model does not generate text; it scores the supplied options and returns a probability for each one in a single forward pass. Uses include request routing, content moderation, checking whether an Agent step succeeded, and selecting the next action. According to the official description, some models can also read images such as documents or screenshots; OpenJev was supported at the time of publication, and the vision projector downloads automatically. In router mode, models load on demand and can be selected per request. /v1/models lists the ids, and the model field is ignored when only one model is loaded. Implementation details are in PR #29818, while the full reference is in the server docs.
ChatGPT Finances rolls out to Free and Go users in the US
产品应用
ChatGPT is rolling out Finances in ChatGPT to Free and Go users in the United States. According to the official announcement, users can connect accounts through Plaid and Experian to understand spending and credit, plan budgets, and review portfolios.
ChatGPT announced that Finances in ChatGPT is rolling out to Free and Go users in the United States. According to its description, users can securely connect accounts through Plaid and Experian, and the feature will use their own financial information to show their cash and credit positions. It can identify forgotten subscriptions that are still being paid, unfamiliar charges, possible duplicate payments, and recurring bills that have increased, while providing weekly financial updates and monthly spending breakdowns. Users can also create budgets based on actual spending, track credit scores and the factors affecting them, discuss debt repayment plans and emergency-fund contributions, and use voice to explore how changing jobs could affect their finances. The feature also supports viewing portfolio composition across accounts and discussing whether a large purchase fits within a budget.
Cambridge CASP publishes study assessing evidence for an intelligence explosion and policy responses
技术与洞察
The University of Cambridge’s AI science and policy project, CASP, published research saying automated AI R&D could compress years of progress into months or less, and called on policymakers to assess developments and devise methods to guide and constrain them.
The University of Cambridge’s AI science and policy project, CASP, published a research article assessing the evidence that automated AI R&D could trigger an intelligence explosion, its potential effects, and possible policy responses. The authors include Geoffrey Hinton, Yoshua Bengio, and Dawn Song. The article says that, compared with one year ago, AI systems now write most of the internal code at the companies developing them and are moving toward automating most or even all AI R&D within a few years. According to its account, further automation of the R&D process could compress years of progress into months or less, and preliminary evidence indicates that this could occur.
The authors state that substantial uncertainty remains. If an intelligence explosion occurs, the benefits of AI could arrive earlier, but it could also create three categories of extreme risk: capability growth outpacing society’s ability to respond, humans losing control of superhuman AI systems, and severe erosion of checks and balances within and among states, companies, and government departments. The research proposes that policymakers urgently gather more information about the automation of AI R&D, develop methods to guide and constrain an intelligence explosion, and prepare society to adapt to its effects. In a separate post, Geoffrey Hinton said that the idea of recursive self-improvement causing an intelligence explosion had not previously seemed imminent, but that many leading researchers now believe it could happen soon.
