Compiled by Bao Yu based on an interview with AMD CEO Lisa Su, World Labs founder Fei-Fei Li, and Li’s open letter: AMD plans to acquire World Labs for approximately $8.2 billion in an all-stock deal. The transaction is still awaiting regulatory approval and is expected to close by the end of the year; Fei-Fei Li will join AMD as Executive Vice President and Chief Scientist. Notably, this deal brings spatial intelligence model R&D and chip and system capabilities under the same company to advance development together.<\/p>\n
World Labs focuses on enabling models to understand and generate 3D worlds. Fei-Fei Li introduced that its new model Atlas predicts subsequent perspectives from a small number of 2D images, emphasizing spatial and temporal consistency; the team has also acquired robotics simulation company SceniX. During the interview, Lisa Su stated that AMD will remain focused on a platform and solutions business model, and noted that the synergy between hardware, software, and models will help capture next-generation computing power demand.<\/p>\n
Source:<\/p>\n
Anthropic has launched Sonnet 5.5, with officials stating it is over 30% faster than Sonnet 5, and cuts costs by up to 30% for most tasks. Test and pricing information compiled by Bao Yu shows that API pricing remains $2 per million input tokens and $10 per million output tokens, with savings primarily coming from a reduction in the number of tokens required to complete tasks.<\/p>\n
In benchmark data compiled by Bao Yu, it scored 70.6% on Terminal-Bench 4.0, compared to 10.3% for the previous generation; Anthropic also noted that Sonnet is designed for well-defined tasks such as bug fixing, documentation, and office work, while Opus remains better suited for complex work that requires continuous judgment. Officials have released demos of the old and new models using the same prompts, but benchmark scores and internal cases should still be considered evaluation evidence provided by the vendor.<\/p>\n
Source:<\/p>\n
Bao Yu paraphrases Manus’ official announcement: Manus 2.0 uses the Cascade AI agent framework, and adds Manus Studio for video editing, a game development environment, and Cue, an independent personal AI agent application. Cue is currently available via invite-only access; Manus 2.0 is already live on web, desktop and mobile platforms.<\/p>\n
Studio allows users to directly replace timeline assets such as clips and music after the initial version is generated, then lets Manus continue adjusting based on the changes; the game development environment supports editing code and assets while running, and projects can be published as web pages. Officials report that under specific test configurations, token consumption is 23.2% lower, completion time is 28.2% shorter, and operating costs are 32% lower compared to the previous generation; these are official test metrics and do not represent that all tasks will achieve the same level of improvement.<\/p>\n
Source:<\/p>\n
OpenAI employee Tibo stated that the $200 Pro subscription will reopen to new users, with usage converted to approximately half of the old version’s allowance when calculated against API usage; the new relative multipliers are 1x for Plus, 5x for the $100 Pro tier, and 10x for the $200 Pro tier. Existing users will temporarily retain their 20x multiplier and receive extra usage allowance; he also promised not to reinstate the 5-hour usage limit and teased that subscriptions will add new benefits that do not consume usage allowance.<\/p>\n
This adjustment affects usage allowances for high-frequency users, and determining whether the new or old plan is more cost-effective depends on the new benefits and specific model pricing; public announcements have already confirmed the multiplier changes, but have not listed all details of the new benefits in this announcement.<\/p>\n
Source:<\/p>\n
Xiao Hu’s introduction to Google’s experimental product CC shows it is built for family groups of up to six people, and aims to let a shared assistant handle family matters scattered across emails, calendars, photos and files. Members can individually select which emails, folders or calendars to authorize, no need to share a single account, and unauthorized content will not be automatically shared.<\/p>\n
In demos, CC extracts time, location and required items from swimming class notifications to create family calendar events; when processing summer camp registration PDFs, it pre-fills existing information, asks family members for missing contact details, and requests member confirmation after filling in all gaps. This is a product demo paraphrased by Xiao Hu, which illustrates the collaboration process but cannot be used to judge long-term performance.<\/p>\n
Source:<\/p>\n
New product information compiled by Xiao Hu states that ElevenLabs has launched Eleven v4 and v4 Turbo: scripts can use tags or natural language to specify whispers, laughter, pauses and ambient sound, and it also supports multi-person dialogue, pronunciation control, and over 90 languages. It also emphasizes seamless transitions for long content and consistent character voices, targeting narration and continuous scene production.<\/p>\n
v4 Turbo focuses on real-time conversation, with official introduction citing around 100ms inference time, a median first audio output time of around 150ms, and supports bidirectional streaming generation; Instant Voice Clone is said to create a voice clone from approximately 10 seconds of audio. The above performance and function details are compiled from a single account’s coverage of the product launch, and are suitable as a list of capabilities pending actual testing.<\/p>\n
Source:<\/p>\n
Nan Yuan states she saw Keling’s official post announcing that 4.0 is scheduled to launch in October, with 4.0 Flash already available for small-scale testing to premium annual card members. The specs she lists include up to 30 seconds of native generation per pass, support for up to 10 keyframes, and mixed image and video reference; 4K, 10-bit HDR, 21:9 aspect ratio and stereo audio are also included in the announcement.<\/p>\n
The post also mentions that extended generation can be stretched up to 2 minutes, but this is marked as “coming soon” and should not be considered fully available as of this date. If these specs launch as planned, long takes and continuous narrative may reduce the need for repeated splicing of short clips; we can only confirm the release information paraphrased from social media at this time.<\/p>\n
Source:<\/p>\n
OpenAI Developers has announced the winners of the WebMCP Challenge, stating that these 10 projects demonstrate how after websites provide structured tools, users and AI agents can collaborate to complete tasks. Official examples include letting agents edit and run JupyterLite notebooks directly in the browser, browse 3D scans, and plan stargazing times based on visible celestial objects.<\/p>\n
Another winning project lets AI agents handle wedding seating arrangements while retaining constraints such as guest relationships and assigned seats. Compared to only demonstrating conversational ability, these examples more concretely show how websites can expose callable operations and the role of humans for review in the process; this is currently a challenge project showcase and does not equate to widespread deployment.<\/p>\n
Source:<\/p>\n
Statistics: Number of scanned timelines: 703 | Number of matched bloggers: 66 | Total matched tweets: 386 | Weighted tweet score: 296.3 | Number of original tweets: 141 | Number of retweeted tweets: 84 | Number of crawl attempts: 6 | Boundary coverage status: tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"
AMD is acquiring World Labs to integrate spatial intelligence and chip R&D, while the updated Claude model delivers faster processing speeds and lower operating costs.<\/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-1811","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\/1811","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=1811"}],"version-history":[{"count":0,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1811\/revisions"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1811"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1811"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1811"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}