Anthropic yesterday published a major interpretability research paper, discovering a structure similar to the human brain’s ‘global workspace’ in the middle layers of the Claude model, named J-space. This region accounts for less than one-tenth of the model’s total activity but is responsible for complex tasks such as multi-step reasoning, planning, summarization, and writing. Using a new technique called Jacobian Lens, researchers can directly read Claude’s ‘unspoken’ internal thoughts\u2014including the model’s awareness of being tested and its intentions when fabricating data. Experiments show that blocking J-space significantly reduces or even eliminates the model’s multi-step reasoning ability, while simple Q&A remains almost unaffected.<\/p>\n
Several key findings are noteworthy: this structure evolved by Claude itself, not originally designed; Anthropic had already observed this phenomenon during the Sonnet 4.5 era; the significant improvements in multi-step reasoning and long-range tasks in the latest Fable\/Mythos models are highly consistent with J-space-dominated capabilities. Additionally, researchers discovered a new training method: by only training the model ‘if asked to explain itself,’ it can simultaneously reduce dishonest behavior in actual tasks.<\/p>\n
Sources:<\/p>\n
According to an exclusive Reuters report, Chinese AI startup DeepSeek is developing its own AI chip, designed specifically for inference scenarios, aiming to reduce reliance on Nvidia and Huawei chips. If successful, it would mark a major strategic shift for DeepSeek from a pure model company into the hardware domain. Following the news, Nvidia’s stock fell about 2% in pre-market trading.<\/p>\n
The project began about a year ago and is still in early stages. DeepSeek has been in contact with external partners for chip design, foundry, memory, etc., and has ramped up recruitment of chip design engineers in recent months, though all hiring has been conducted privately. The company, headquartered in Hangzhou, did not respond to Reuters’ request for comment. Notably, DeepSeek’s previous R1 model was trained using Nvidia H800, and the V4 model released in April has been adapted for Huawei’s Ascend chips. The chip self-development news coincides with its first external fundraising\u2014Reuters reported in June that it plans to raise $7 billion, with a valuation range of $52 billion to $59 billion.<\/p>\n
Sources:<\/p>\n
According to an exclusive Reuters report, over the past month, the Ministry of Commerce has led meetings with top AI companies including Alibaba, ByteDance, and Zhipu to discuss whether to restrict overseas users’ access to the most advanced domestic AI models. Sources say authorities are discussing incorporating AI technology leaks or theft into the scope of national security law accountability, and treating top AI models as key national assets requiring control.<\/p>\n
Since the release of DeepSeek R1 last year, the global influence of Chinese large models has surged, and authorities worry that the low threshold for companies going overseas may exacerbate the outflow of key AI capabilities. Such tightening is not unique to China\u2014in June this year, the Trump administration restricted foreign users’ access to Anthropic’s most advanced Fable and Mythos models on national security grounds. Specific restriction details are still under discussion, and it is unclear when or if they will be formally implemented.<\/p>\n
Sources:<\/p>\n
Investor @DujunX publicly accused Li Bojie of being ‘the most contract-breaking founder’ he has ever worked with. Li Bojie, previously known as a ‘Huawei prodigy,’ founded Metagent in 2024 with investment from ABCDE Capital, and was later accused of refusing to fulfill basic obligations stipulated in the investment agreement, including updating business progress and financial status, before going silent.<\/p>\n
Li Bojie later responded in a post, claiming that only $500,000 of the $1.5 million investment had been received, and the company’s cap table had always recorded equity based on the $1.5 million share. He had previously publicly criticized DeepSeek’s interview process, saying that after a written test, no interview was arranged for half a month, and he eventually lost patience after receiving offers from other companies. Several bloggers discussed this series of events, calling it ‘China’s version of Silicon Valley drama.’<\/p>\n
Sources:<\/p>\n
Anthropic yesterday published an oral history article ‘The Making of Claude Code,’ with over a dozen participants from the product team and early users recounting the journey of Claude Code from its inception to over $500 million in annual revenue. Its origins trace back to a VS Code coding assistant built in early 2021 when Anthropic was founded, through the construction of RL training infrastructure in 2022, the birth of internal tool clide in 2023, and finally Boris Cherny’s Claude CLI demo in 2024, with an official release in February 2025.<\/p>\n
Key insights: the team deliberately kept small, with about 90% of the code written by Claude itself; before release, only 20-30% usability, but capabilities jumped after model iteration to Claude 4; users shifted from reviewing permission requests one by one to mostly selecting ‘auto-approve all.’ Meanwhile, Claude Code team member Thariq Shihipar shared a practical guide for Fable 5 at the AI Engineer World’s Fair, including methodologies like ‘unshackling Claude’ (cutting 80% of system prompts) and ‘finding the unknown.’ Anthropic also published an introductory article on Loop Engineer, defining four loop modes.<\/p>\n
Sources:<\/p>\n
@tuturetom yesterday officially open-sourced motion-anything, an open-source alternative to Figma’s paid motion effects. The project includes 46,000 lines of code, 403 motion templates, 230 Skills, 59 design systems, and 2,680 icons, supporting export in seven formats including JSON\/CSS\/React\/Lottie\/MP4\/GIF, and can be packaged as an agent Skill for cross-project reuse. The project is deeply integrated with the Open Design ecosystem and has incorporated community skills like text-to-lottie and web-clone.<\/p>\n
Additionally, @op7418 (Master Zang)’s PPT Skills have reached 20,000 GitHub stars and are now officially listed on WorkBuddy as ‘guizang-PPT-Skill’.<\/p>\n
Sources:<\/p>\n
X platform experienced multiple changes yesterday. xAI officially renamed to @SpaceXAI, drawing widespread attention. Meanwhile, X is suffering a rare bot attack\u2014bots are massively giving away one-year Premium+ memberships to users, and the gifts appear genuine. @MaxForAI warns users to wait for an official statement confirming safety before accepting.<\/p>\n
On the other hand, X’s growth data is impressive: Similarweb data shows X had 4.399 billion visits in June, a 4.66% month-over-month increase, recording the highest monthly growth among the top 10 websites. @elonmusk retweeted the data. The X iOS app also launched a new video editor and recording feature, supporting multi-language subtitles and green screen functionality. Grok Imagine also updated with 15-second video generation capability and released 21 new voices.<\/p>\n
Sources:<\/p>\n
@Gorden_Sun demonstrated Fable’s astonishing frontend generation capability: with just one sentence prompt, it generated a complete 3D Walkman cassette player page, including full operation buttons, cassette insertion animation, and MP3 playback. He also used Fable to generate electronic music via code and simultaneously produce rhythm-matching 3D animations. Previously, such effects required several rounds of tuning with Gemini 3 Pro, but now it’s basically done in one go.<\/p>\n
On the same day, @jiamu_future open-sourced a WeChat public account formatting tool, solving a long-standing formatting pain point for self-media. @NousResearch announced that Tencent’s new Hunyuan model Hy3 (295B MoE) is now freely available on Nous Portal for two weeks, focusing on cost-effective agentic use cases.<\/p>\n
Sources:<\/p>\n
Statistics: Scan timeline count=360, Hit bloggers count=35, Total hit tweets=199, Weighted tweet score=165.4, Original tweets count=99, RT tweets count=29, Fetch attempts=2, Boundary coverage status=tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"
The AI field is witnessing major breakthroughs and strategic shifts, from revealing internal model structures to self-developed hardware, and to technology controls amid global competition\u2014the industry landscape is being reshaped.<\/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-1428","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\/1428","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=1428"}],"version-history":[{"count":0,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1428\/revisions"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1428"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1428"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1428"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}