Researchers say GPT-6 Astra directly rewrites model weights
要闻
Researcher Rikiya Takehi reports that GPT-6 Astra can directly edit model weights without gradients. Changing just 0.016% of Qwen3.5-0.8B’s weights reportedly raised its MATH-500 score from 7.9% to 48.1%.
In an article with no publication date given, researcher Rikiya Takehi reported Weight-Space Programming experiments, claiming that GPT-6 Astra can directly read and modify neural network weights without an optimizer or gradient descent. According to the account, with Astra prohibited from writing training loops, it changed 0.016% of Qwen3.5-0.8B’s weights to address a reasoning-loop issue mentioned in the model’s model card. Under the same sampling settings, the MATH-500 score reportedly rose from 7.9% to 48.1%, while IFEval increased to 3 times its original level. The article also claims that small game-playing networks Astra wrote from scratch without environment feedback outperformed networks trained with PPO for 100 million steps. In weight-reading experiments, Astra reportedly identified the games that four networks had each been trained to play using only their raw weights, without being given a list of possible games. For DDPMs separately trained to generate circles, triangles, and squares, it reportedly identified the shapes without tools or forward passes, but drew outlines even though the models actually generated filled shapes.
OpenAI puts its in-house Git backend Pando into production, cutting median push latency by nearly 80%
要闻
OpenAI has put its in-house Git backend, Pando, into production for monorepo development, reporting a nearly 80% reduction in median push latency. The project took 10 weeks from start to production, and engineers and agents can use it without changing their existing Git workflows.
OpenAI disclosed that its in-house Git backend, Pando, is in production and now supports all development work in its monorepo; the source does not specify a production launch date. According to the company, the switch reduced median push latency by nearly 80% and improved merge reliability. The system delivers nearly one PB of Git data daily and has delivered more than 57 TB in a single hour. Engineers and agents can continue using their existing Git workflows. Previously, growth in parallel development driven by Codex exceeded the capacity of the existing system, resulting in slower pushes and fetches, failed merges, and reliability that at one point fell below 90%. Written in Rust, Pando took 10 weeks from project start to production. One engineer handled the core design, while multiple agent teams carried out the implementation, and some Git client improvements have been contributed upstream.
Kimi unveils solutions for game development and creative work
产品应用
Kimi has announced a solution for game development and creative work that connects to professional software through plugins and operates on existing projects. According to the company, teams with limited programming experience can build playable prototypes in 1 to 2 days, while the solution also covers full development and post-launch operations.
Kimi has released a solution for game development and creative work, although the source does not specify a release date. Built on the Kimi K3.1 Pro model, Kimi Code, and the Kimi Work series of Agents, it connects through plugins to professional software including Unity, Unreal Engine, and Blender, allowing it to operate directly on existing projects. According to the company, planners or creative teams with limited programming experience can produce a playable prototype in 1 to 2 days before formally starting a project, then use the playtest experience to decide whether to invest. For full development, the company says Kimi can map the code structure, make changes in the engine editor, and validate them against runtime data after receiving project files and the necessary permissions. Post-launch uses include producing user-acquisition assets, in-game assistants, organizing community feedback, content localization, and content moderation.
阿里云 launches Qovest, an AI investment research workbench
产品应用
Alibaba Cloud has launched Qovest, an AI investment research workspace for professional investors that turns research requests into reports and tracks research subjects over time. Invite-only testing is open to individuals, with registration credits for a free trial and subsequent subscription plans priced at 99 yuan or 199 yuan per month.
According to the November 3, 2026 edition of Juya AI Daily, Alibaba Cloud has launched Qovest, an AI investment research workspace for professional investors, and opened invite-only testing to individual users. New users receive credits upon registration for a free trial, with subsequent subscription plans priced at 99 yuan or 199 yuan per month. According to the product description, users can submit a research request in a single sentence, after which Qovest can automatically retrieve data, gather materials, conduct in-depth analysis and produce a report. It can also track research subjects at a specified frequency. Research results are saved in an investment research space for continued work on the same project, rather than serving only as one-off answers.
商汤 releases SenseNova 6.8 Flash, now free for a limited time
模型发布
SenseTime has released the production version of SenseNova 6.8 Flash, currently free for a limited time. Through sales follow-up and staff handover scenarios, the company describes how the model connects product updates with earlier customer requirements and organizes customer information from authorized historical records.
SenseTime has released the production version of SenseNova 6.8 Flash and is offering it free for a limited time; the supplied material specifies neither the release date nor the duration of the free offer. The company's workplace examples cover customer-information handovers and sales follow-up. According to its description, when an employee changes roles, the model can use authorized historical work records to compile deal progress, unresolved requirements, customer feedback, and communication considerations for the incoming employee. In another example, a customer had declined to proceed because an enterprise AI product did not support private deployment. When a product update added that capability one month later, the model linked the feature change to the previously recorded requirement during preparations for a regular sales meeting report and proactively prompted the salesperson to follow up with that customer.
TRAE 宝 goes free for a limited time: Q&A and tasks use no credits
开发生态
TRAE announced that its personal assistant, TRAE 宝, is free for a limited period through November 8 at 23:59, with no credits consumed for questions or task assignments after users log in to the latest desktop app. The assistant can split tasks for parallel execution, but official documentation states that video generation is not currently supported.
TRAE announced a limited free period for its developer personal assistant, TRAE 宝, starting with the announcement and ending on November 8 at 23:59. Users who log in to the latest TRAE desktop app can ask questions or assign tasks without consuming credits; the source does not specify the year. According to the company, TRAE 宝 can understand users’ plans and decisions across projects and break complex requests into subtasks for parallel execution. Users can assign new tasks without waiting for an earlier one to finish. After progressing tasks in the background, the assistant returns a consolidated set of questions requiring user judgment, pending items, and deliverables. The company also states that TRAE 宝 does not currently support video generation.
ZCode launches three official plugins, enabling one-click deployment of web projects
开发生态
ZCode has introduced three official plugins in version 3.17.1: Computer Use, Browser Use, and ZCode Sites. They cover local desktop and browser operations, as well as one-click deployment of web projects to publicly accessible links.
With the release of version 3.17.1, ZCode launched three official plugins—Computer Use, Browser Use, and ZCode Sites—providing computer control, browser operation, and web deployment capabilities, respectively. Users must first install and enable the relevant plugins in the plugin marketplace, then enter /computer-use or /browser-use, or request computer or browser operations in natural language. According to the official description, the first two plugins can interpret screen content and operate local desktop software and browsers for tasks such as filling out OA forms and conducting image-and-text research. They can also open a browser to test web page interactions and fix problems. ZCode Sites deploys completed web projects in one click, generating publicly accessible online links.
Anthropic Message API launches message threads in public beta
开发生态
Anthropic has launched a public beta of message threads in the Messages API, caching conversations server-side so developers no longer need to resend the full history each turn. The company says the feature is enabled by default for most Claude Code users, with token usage and standard pricing unchanged.
Anthropic has released message threads in public beta in the Messages API. Once developers enable the feature, each turn requires only new messages and the previous message ID, while the server caches the conversation history. According to the company, the feature is integrated into Claude Code and enabled by default for most users. For requests containing 500,000 to 1,000,000 tokens, it reports an approximately 35% reduction in median time to first token, while the median size of an individual request in long conversations fell from approximately 800 KB to 1.8 KB.
Threads reduce the repeated transmission of historical messages without changing how Claude reads the entire conversation each turn. Token usage therefore remains unchanged, and standard token pricing and existing rate limits continue to apply. A thread expires 5 minutes after the last request, so developers must still retain a copy of the conversation in their applications. The feature works with all models but does not support Message Batches processing, token counting, or conversation compaction.
OpenAI releases LifeVal benchmark to measure AI's ability to complete everyday tasks
技术与洞察
OpenAI released LifeVal, a benchmark for everyday tasks spanning caregiving, health, education and other areas. According to the company, scores measured by unweighted average reward rose from 49.3% for GPT-5.2 to 80.6% for GPT-6 Astra over the past year.
OpenAI released LifeVal to measure AI’s ability to help people complete everyday tasks; the source does not specify a release date. Its tasks cover caregiving, health, education, household administration, personal finance, shopping and access to government services. The company said it has used LifeVal to evaluate and improve models over the past year, with scores measured by unweighted average reward reaching 49.3% for GPT-5.2 and 80.6% for GPT-6 Astra. According to its explanation, the benchmark incorporates feedback from thousands of ChatGPT users, while task design draws on the American Time Use Survey and cross-national time-use research, alongside an additional demographically representative survey of 3995 respondents. The company described its aim as moving beyond academic evaluations to test whether models can help ordinary people accomplish practical tasks in their daily lives.
OpenAI deploys activation classifiers to block policy-violating model outputs in real time
技术与洞察
Starting with GPT-5.6, OpenAI has added activation classifiers to all assistant completions to monitor policy violations using internal model signals. The company says these monitors can interrupt generation, but depend on labeled data and require retraining for each new pre-trained base model.
OpenAI began running activation classifiers on all assistant completions with GPT-5.6 and has deployed them on internal and external traffic for its most capable models, adding a monitoring layer alongside text classifiers. According to the company, these lightweight ML models, also called probes, read internal residual stream activations to score tasks such as whether a completion violates policy. During generation, a message is flagged as soon as its score crosses a threshold, allowing a violating output to be blocked midway. Training data combines synthetic and de-identified conversations, typically using up to 1 million messages, with the company’s most capable reasoning models supplying labels as LLM judges. For each hyperparameter configuration, the company typically trains at least 32 random initializations, selects a candidate using a validation dataset, and checks performance on a held-out evaluation set. OpenAI says probes can detect only known, measurable forms of undesirable activity and must be retrained for each new pre-trained base model. Even within the same model family, performance can decline as weights diverge from the checkpoint used for probe training. The experimental results in the article are not necessarily representative of deployed classifiers.
腾讯混元 unveils Hyra 2.0: from solving problems to building tools
技术与洞察
Tencent Hunyuan has released Hyra 2.0, an automated scientific research system that shifts the focus of Agent tasks toward developing and upgrading reusable research tools. The system uses three tiers of responsibilities and requires human approval for certain operations, while the company says results will be published on GitHub on an ongoing basis.
Tencent Hunyuan has released its automated scientific research system Hyra 2.0 and has already completed the development and upgrading of a set of tools. Rather than having Agents directly solve scientific problems, the system now focuses on having them develop and improve reusable scientific models, research workflows, and software tools. For execution, it organizes tasks into three tiers—Overall, Mission, and Worker—and connects to WeCom, computing platforms, and code repositories; operations involving substantial computing resource consumption or shared-knowledge synchronization require human approval. Tencent Hunyuan says the related results will continue to be made public in the GitHub repository Tencent-Hunyuan/hyra-results.
SemiAnalysis releases GDPx benchmark, with the best model scoring just 12%
技术与洞察
SemiAnalysis has released GDPx, a benchmark for general Agent capabilities, with the highest-scoring model reaching just 12%. It focuses on areas including software engineering and recursive self-improvement; the publisher says safety classifiers affect some models’ overall scores, while performance also varies across domains.
SemiAnalysis has released GDPx to evaluate general Agent capabilities, although the supplied article does not specify a publication date; according to the publisher, even the best-performing model scores only 12%. Domains are selected for the potential application value of frontier intelligence rather than GDP share or the SOC/O*NET occupational taxonomy. The benchmark therefore includes software engineering and recursive self-improvement (RSI), while excluding government, real estate, and insurance. The publisher believes the first trillion dollars in AI revenue will concentrate in industries that can deploy frontier intelligence at high ROI, rather than being distributed evenly across the economy.
According to the publisher, Opus 5.5 performs best overall and leads in quant trading, hardware design, and chip design; GPT-6 Astra leads in RSI and materials science, while Gemini 4 Carbon leads in biology and software engineering. Almost every model scores exactly 0 on the finance and legal tasks. The publisher says safety classifiers lower the overall scores of models including Opus 5.5 and GPT-6 Astra, so it also presents results excluding cybersecurity, biology, and alignment. To develop tasks for individual domains, SemiAnalysis partnered with specialized RL environment startups and says all partners underwent a QA process.
OpenAI launches Grid Protectors initiative, with compute campuses set to reduce peak demand on the power grid
行业动态
OpenAI announced its nationwide Grid Protectors program to reduce computing campuses’ electricity use when the grid is under strain. The first implementation site is in Georgia, where its agreement with Georgia Power includes up to 1000 megawatts of flexible demand response.
OpenAI announced Grid Protectors, a nationwide program intended to support electricity supplies for local communities by reducing computing campuses’ power consumption when the grid is under strain. The first implementation site is a planned campus still under development in Effingham County, Georgia, and the agreement with Georgia Power includes up to 1000 megawatts of flexible demand response. Operating arrangements will be determined for each location, with OpenAI working with utilities and other partners to establish how much demand can be reduced, how quickly reductions can be made, and how long they can last. According to the company, the program aims to support affordable, reliable electricity for communities, and OpenAI will cover all infrastructure and electricity service costs required to serve the first project.
