OpenAI Announces Internal Version of Astra Model Solves Ten Open Math Problems
要闻
OpenAI has published research on what it calls its next major model, "Astra." An internal version of the model has produced ten new results on long-standing open problems in mathematics and theoretical computer science, covering areas such as high-dimensional sphere packing, non-sofic groups, and arithmetic circuit complexity. The Lean formal proofs of the arguments have been made public on GitHub for machine verification.
OpenAI published a blog post introducing research results from an internal version of its next-generation flagship model, Astra. The model produced ten new results on a set of long-standing open problems in mathematics and theoretical computer science, covering high-dimensional sphere packing, binary and spherical codes, non-sofic groups, the Connes rigidity conjecture, arithmetic circuit complexity, quantum parallel repetition, the Closest Vector Problem, multicolor Ramsey numbers, and extremal graph theory.
OpenAI stated that the tokens consumed to find these results cost approximately $2,000 at the pricing of the GPT 5.6 Sol API. The mathematical arguments were generated by the model, then organized into a paper by humans with model assistance, and finally the proofs were formalized into Lean certificates by the model. OpenAI also made the relevant proofs and process descriptions public for the mathematics community to inspect, discuss, and further investigate.
Tibo Says Usage Limits for Codex and ChatGPT Work Have Been Reset
要闻
At noon Beijing time on August 1, Tibo, the head of Codex, said that usage limits for Codex and ChatGPT Work had been reset, and that this was to celebrate Efficiency Week, allowing users to run 100,000 Luna task sessions this weekend.
At noon Beijing time on August 1, Codex lead Tibo announced on social media that he had reset usage limits for Codex and ChatGPT Work. Tibo said this was to celebrate a week of focused efficiency work, allowing users to run 100,000 Luna task sessions over the weekend.
DeepSeek Harness Recruits Developers of Agent Harness Open Source Projects for Beta Testing
要闻
DeepSeek Harness is currently in its beta phase. The person in charge said that developers of open source projects related to Agent Harness can contact them to participate in the beta, with their GitHub ID and representative open source work.
DeepSeek Harness lead Tianyi Cui announced that DeepSeek Harness is currently in internal testing and is recruiting internal testers from developers of open-source projects related to Agent Harness. Developers can apply by replying or sending a private message to Tianyi Cui, and the application should include a GitHub ID and representative open-source works. No mention was made of public beta timing or official release plans.
SpaceXAI Adds Text-to-Video, Citation Features, and Native 1080p to Grok Imagine
模型发布
SpaceXAI announced an upgrade to the Imagine Video 1.5 model, adding text-to-video, image and voice citation, and native 1080p support. These features are now being rolled out across official platforms.
SpaceXAI's video model Imagine Video 1.5 has received feature upgrades, adding text-to-video, image reference, voice reference, multi-reference, and native 1080p support. The multi-reference feature allows binding up to 7 reference images per generation, locking characters, scenes, or products separately. Image and voice reference are currently available to SuperGrok Heavy and SuperGrok Plus subscribers in the U.S. on grok.com/imagine and iOS; text-to-video and native 1080p are fully available on grok.com/imagine, iOS, and Android. On the API side, image reference, text-to-video, and native 1080p are now live, while voice reference support requires contacting sales.
OpenAI Team Submits Multiple Performance and Correctness Patches to Git Upstream
开发生态
OpenAI is continuing to push Git improvements for large repositories, with performance and correctness patches partially merged into the upstream main branch. The company says these optimizations will also improve Git usage efficiency in the Codex app.
OpenAI is continuing its work on Git improvements for large repositories, led by Ted Nyman and Taylor Blau, developed in a public fork, with patches contributed upstream to Git when they have general value. Currently, 2 patch series have been merged into the upstream Git master branch, 5 series are under review in the upstream seen branch, and another 23 series are still in development on OpenAI's topic branches. Merged improvements include path-limited git stash sparse-index optimization, reducing index expansion time from 18.87 seconds to 0.06 seconds, and path-traversal bitmap support. Ted Nyman said that performance improvements are also being applied to the Codex app, with more efficient Git usage reflected in upcoming new versions.
Report: OpenAI and Anthropic Lobbying to Require Competitors to Accept Same Scrutiny
前瞻与传闻
According to reports, OpenAI and Anthropic are quietly urging the U.S. government to require competitors such as Meta and SpaceXAI to undergo the same frontier model reviews they face.
According to The Information, OpenAI and Anthropic are quietly urging the U.S. government to require rivals such as Meta and SpaceXAI to undergo the same frontier-model reviews that the two companies currently face. The report says that OpenAI and Anthropic have already accepted such frontier-model reviews, and the two companies hope competitors will also be brought under the same review framework. The report notes that this alliance could protect OpenAI and Anthropic from being slowed down if their competitors avoid such scrutiny. The news comes from a media report by The Information, and neither the companies involved nor Washington has issued an official response.
Note: Content is AI-assisted and may contain hallucinations and errors.
