Anthropic is rolling out Claude Code Projects in a limited release: users simply state their goal in a main conversation, and Claude will break the task down into multiple parallel Threads. These Threads run independently in the cloud, share project memory, and continue working even after the user’s computer goes offline. Each Thread has its own copy of the code and branch, can run tests, open PRs, and the project will centrally display items requiring user decisions.<\/p>\n
This transforms a project from a “folder for storing materials” into a continuously operating work hub. The official current statement is that the feature is first available for cloud sessions for some Pro\/Max users; Threads temporarily cannot access local files, tools, or internal networks. Boris Cherny stated that he no longer manually manages conversations, instead handing ideas directly to the project for breakdown and execution.<\/p>\n
Sources:<\/p>\n
Qwen officially released Qwen3.8-Omni-Flash, positioned as a full-modality model for Agents: capable of simultaneously understanding audio and video, planning tasks, calling tools, and covering workflows like video editing, short video translation, and movie summarization. The official capability description includes a 1 million token context window, an average 19.5-point improvement on audio-video Agent evaluations, and approximately an 89% reduction in video input costs compared to Qwen3.5-Omni-Plus.<\/p>\n
The accompanying actions are clear: Qwen simultaneously opened Qwen-MM-Plugins and is preparing Qwen-Live Harness. Blogger @LufzzLiz further clarified that the standard version natively outputs text, real-time voice requires the Realtime version, while final videos, dubbing, and PDFs are still handled by accompanying tools; therefore, it currently resembles a usable “model plus toolchain” solution rather than a single model handling all outputs.<\/p>\n
Sources:<\/p>\n
OpenAI announced Astra for Law, stating it is powered by GPT-6 Astra and comes with tools, settings, and context tailored for legal practice. The initial phase will be offered to selected law firms through Trusted Access for ChatGPT and Codex, with API access to follow. OpenAI stated it will maintain the legal configuration, allowing developers to focus on their own products and workflows.<\/p>\n
This release also brings 26 partner plugins and 47 community plugins, with partners including Thomson Reuters, Harvey, Legora, and iManage. The focus is evidently not just a model capability showcase, but integrating professional knowledge, existing tools, and lawyer-customized skills into the same working environment. The officially available scope is still limited to selected institutions and subsequent API openings.<\/p>\n
Sources:<\/p>\n
After releasing Helix 2.5, Figure sent robots equipped with the model into 30 real homes in the San Francisco Bay Area that had not previously provided data, testing three long-horizon tasks: tidying living room toys, folding towels, and making beds. A blogger’s summary reported a comprehensive single-trial success rate of 56%, and specifically noted that the robot could readjust its stance, change angles, or detour and continue execution when encountering mistakes or limitations.<\/p>\n
The value of this set of demonstrations lies in the test conditions, which included unfamiliar environments, no on-site fine-tuning, and multi-step actions. The focus shifted from whether a single action was completed to self-correction after failure. The existing material is Figure’s release information and summaries from two bloggers. The 56% figure is a test result from the visible reports and should not be extrapolated as a general household chore capability.<\/p>\n
Sources:<\/p>\n
Visible forensics from the community regarding ZCode claimed that the client might package the workspace and complete Git history and upload it to Alibaba Cloud OSS during login or indexing; the controversy centers on default behavior, user awareness, controls, and key management. Subsequently, @MaxForAI relayed a statement from Zhipu AI: the issue was related to Repo Wiki triggering cloud generation within “codebase indexing,” uploaded data would be destroyed after generation, the related issue has been fixed, and they plan to open-source the code, accept third-party audits, and reset users’ weekly quota once.<\/p>\n
Currently, the two layers of evidence should be separated: the upload behavior comes from community forensics and relayed statements; the reason for the fix and subsequent measures come from the relayed official response. In the absence of more independent audit materials, responses like “data has been destroyed” should not be treated as externally verified conclusions. This incident directly impacts developer trust in local Agents and has made default uploads, privacy statements, and the ability to disable features into selection criteria.<\/p>\n
Sources:<\/p>\n
@dotey summarized a judgment framework from Jason Wei’s speech at the Stanford AI Club: AI capabilities, once reaching the frontier, rapidly become commoditized, with reasoning costs continuously declining; tasks that are easier to verify objectively, quickly, and at scale are more suitable for generating candidate solutions with models and then automatically filtering them. AlphaEvolve is cited as an example: the model generates a large number of candidate solutions, a verifier scores them, and the best solutions are used for the next round.<\/p>\n
This framework also emphasizes that AI capabilities advance in a “jagged” rather than uniformly synchronized manner: competition-level mathematics and certain programming tasks are nearing their peak, while tasks in chemistry, the physical world, and data-scarce domains progress more slowly. It is better suited as an analytical framework for judging the sequence of automation, not as a definitive prediction of outcomes for specific years or industries.<\/p>\n
Source:<\/p>\n
@dotey’s practical advice is that non-engineers should not treat code cleanliness or internal reasoning as the primary acceptance criteria, but should instead, like QA, check for functional completeness, performance, and security; tasks should be broken down into small modules that an Agent can handle reliably, then accepted through testing, runtime performance, and security checks. A subsequent addition acknowledges that acceptance can only cover pre-conceived scenarios, and critical paths involving money, data, and security should still be reviewed by professionals.<\/p>\n
This differs from “deploying AI-written code directly”: the core idea is to delegate implementation details to the model while keeping observable results, boundary conditions, and rollback capabilities in human hands. Independent practice by @imwsl90 also shows that AI coding is more suitable for rapid product validation and fixing discovered issues, but cannot replace checks on critical paths like payment vulnerabilities.<\/p>\n
Source:<\/p>\n
OpenAI Developers showcased Appshots on Windows: users can pass the context of the application they are currently using to ChatGPT for debugging, replicating interfaces, or reading data from another application, reducing repetitive copy-pasting; simultaneously, the desktop client adds a new usage analytics entry point, allowing users to see the contribution of tasks, sub-agents, and individual conversations to Codex usage.<\/p>\n
These two updates address the problems of “letting the model see the work environment” and “knowing where the quota is spent,” respectively. The former reduces the cost of context transfer, while the latter helps users adjust their workflow. The public tweets only explain the feature entry points and typical use cases; the specific supported scope is subject to the actual client version.<\/p>\n
Source:<\/p>\n
Stats: Timeline Scans=554 Number of Bloggers Matched=56 Total Tweets Matched=320 Weighted Tweet Score=234.65 Original Tweets=117 RT Tweets=88 Crawl Attempts=4 Boundary Coverage Status=tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"
Today’s AI developments focus on agent collaboration and industry applications, showcasing deeper integration of models into complex tasks and specialized scenarios, from code development and audio-video processing to legal workflows.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[19],"class_list":["post-1760","post","type-post","status-publish","format-standard","hentry","category-brief","tag-x--ai-"],"_links":{"self":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1760","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1760"}],"version-history":[{"count":0,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1760\/revisions"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1760"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1760"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1760"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}