Xiaoliu BOT
中文EN

X Platform July 20 AI Brief | Claude Fable 5 Disproves Jacobian Conjecture, Hugging Face Hit by AI Agent Attack, US-China AI Gap and Model Competition

Claude Fable 5 Disproves Jacobian Conjecture, AI Cracks 87-Year Math Problem During World Cup

Anthropic mathematician Levent Alpöge tasked Claude Fable 5 with studying the Jacobian Conjecture (proposed in 1939, one of the most famous open problems in algebraic geometry). Fable 5 directly provided a concrete counterexample of a polynomial map from C³ to C³: its Jacobian determinant is identically -2, but three distinct inputs produce the same output, proving the map is not invertible. Levent noted that Fable completed this work during the World Cup final. An 87-year problem for humanity was solved by AI during the final broadcast. The general case has been disproven; the two-dimensional conjecture remains open.

Sources:

Hugging Face Breached by Fully Automated AI Agent; Closed-Source Models Refused to Assist in Investigation, Open-Source GLM 5.2 Completed Analysis

On July 16, Hugging Face issued a security advisory: a fully automated AI Agent exploited two code execution vulnerabilities from a malicious dataset, gained node-level permissions, moved laterally, stole cloud and cluster credentials, established a self-migrating C2 control system, and left over 17,000 attack events. Ironically, during the investigation phase, attempts to use ChatGPT and Claude APIs to analyze logs failed as the attack commands, payloads, and C2 data triggered safety guardrails, causing the models to refuse work. They ultimately had to run the open-source Zhipu GLM 5.2 model on their own servers, completing log analysis, attack timeline reconstruction, and credential localization within hours. Multiple bloggers commented that this incident highlights the security initiative value of open-source models in Agent-based offense and defense.

Sources:

Kimi K3 and the US-China AI Gap Discussion; Qwen3.8-Max and Grok 4.5 Make Moves

Following the release of Moonshot AI’s 2.8 trillion parameter Kimi K3, multiple bloggers joined the discussion. @xiaohu cited an evaluation by independent analyst Lisan al Gaib using the Epoch Capability Index (ECI): the US-China AI gap is not the rumored 45 days, but approximately 160 days (nearly 6 months). On another front, Alibaba released Qwen3.8-Max-Preview (2.4T parameters, supports 1M context). @LufzzLiz’s tests show it is about 10 times faster than Kimi K3, slightly weaker than GPT-5.6 Sol in office scenarios but offers outstanding cost-performance; the official version is set to be open-sourced. xAI’s Grok 4.5 scored 91.3% on the VulcanBench coding benchmark, surpassing Claude Fable 5 and GPT-5.6 Sol, with Musk retweeting the recommendation. @FuSheng_0306 shared Silicon Valley observations: startups there are already using Chinese open-source models at scale, and Anthropic’s revenue growth slowed in June.

Sources:

OpenChatCut Released: An Agent-native, Local-first Open-Source AI Video Editor

@cellinlab fulfilled a prediction made three months ago by officially releasing OpenChatCut. This editor allows Codex, Claude Code, and built-in Agents to directly read, edit, and export real video projects for continued editing. It supports multi-track timelines, script-driven editing, MG/WebGL, FCPXML export, and prioritizes saving projects and assets locally. Multiple bloggers shared and endorsed it. Additionally, @cellinlab shared best practices for finding and “copying” benchmarks (endorsed by @dontbesilent), and information that light-ocr reached #2 on GitHub Trending within 5 days of launch.

Sources:

Pinduoduo Recommendation Algorithm Sparks Privacy Debate; Follow-up Controlled Experiment Points to Algorithm Bug

@MaxForAI first reported: users found that after viewing a specific Pinduoduo screenshot, recommendations for “Honkai: Star Rail” character figurines would appear on the homepage, sparking widespread discussion about “Pinduoduo monitoring users via screenshots.” However, a subsequent controlled experiment showed that downloading the Pinduoduo app on a new phone without having seen the image also triggered the recommendation, suggesting a recommendation algorithm bug. @cellinlab analyzed the potential data alliance mechanism behind it. Multiple bloggers participated in the discussion, which spread widely in groups and on social media.

Sources:

WeChat Channels Download Tool and learnui Website Resource

@vista8 (Xiangyang Qiaomu) shared an open-source WeChat Channels download tool: open a Channels video in the WeChat built-in browser, copy the URL to download. It can be deployed on Cloudflare as a personal downloader. Testing shows it does not support live stream replays, but regular videos download normally. He also recommended the learnui website—it used k3 to generate previews of 74 well-known websites, suitable for finding design style references. A tip is to ask the AI to add PWA support to improve open rates.

Sources:

New Progress in Anti-Aging Research: Blocking EP2 Receptor Restores Immune Clearance Ability

@paji_a cited Stanford research: a key driver of aging is the loss of ability by tissue-resident macrophages to clear senescent neutrophils, leading to systemic chronic inflammation and synchronized decline in organ function. By blocking the PGE2-EP2 signaling pathway, inflammation in multiple organs (liver, heart, kidneys, etc.) in aged mice was suppressed, and muscle strength, exercise capacity, and memory function approached youthful levels. This points to a new paradigm in anti-aging research, shifting from “killing senescent cells” to “restoring self-clearance ability.”

Sources:

Model Fingerprinting, Codex Reset Controversy, and Other Practical Info

@xiaohu shared an interesting study from Prague University of Economics and Business: leveraging AI’s “flaw” of being unable to produce truly random numbers, inputting a random number task can identify the model—GPT-4o favors 42, Claude Sonnet 5 almost always says 47, Qwen3-Max always answers 42. This can be used to detect if a proxy service is swapping models. Additionally, @wangyinrec (cited by @xiaohu) pointed out a controversy with Codex quota resets: OpenAI extends the original date with each reset, while Anthropic does not, effectively giving a bonus. @zhufengme (cited by @cellinlab) proposed a Vibe Coding reasonable workload ratio of about 1:25 (one day of AI-assisted work equals two weeks of manual coding), but a 25x efficiency increase does not mean a proportional salary increase.

Sources:

Stats: Scanned timeline items=360 Number of bloggers matched=32 Total matched tweets=146 Weighted tweet score=121.25 Original tweets=80 RT tweets=24 Crawl attempts=2 Boundary coverage status=tail_confidently_crossed_target_boundary