From 29 items, 11 important content pieces were selected
- Eli Lilly’s retatrutide phase 3 trial achieves 28.3% weight loss ⭐️ 9.0/10
- Project Glasswing Update: 90.6% True Positive Rate for Vulnerabilities ⭐️ 8.0/10
- Why Japanese Companies Diversify So Much ⭐️ 8.0/10
- Antigravity 2.0 Tops OpenSCAD LLM Benchmark for 3D Models ⭐️ 8.0/10
- yt-dlp deprecates Bun support over Rust rewrite and AI concerns ⭐️ 8.0/10
- U.S. Researchers Face Unclear Restrictions on Foreign Collaborations ⭐️ 8.0/10
- AI’s HBM Demand Squeezes Consumer Memory, Driving Price Hikes ⭐️ 8.0/10
- BPF-based custom page-cache eviction proposed at Linux Summit ⭐️ 8.0/10
- ByteDance Open-Sources Lance: 3B Unified Multimodal Model ⭐️ 8.0/10
- China Cracks Down on Illegal Cross-Border Securities Trading ⭐️ 8.0/10
- Cloudflare global outage 25 min, 28% HTTP traffic affected ⭐️ 8.0/10
Eli Lilly’s retatrutide phase 3 trial achieves 28.3% weight loss ⭐️ 9.0/10
Eli Lilly announced results from the TRIUMPH-1 phase 3 trial of retatrutide, showing average weight loss of 28.3% at the highest dose (12 mg) over 80 weeks in adults with obesity or overweight without diabetes. All primary and key secondary endpoints were met. This result represents one of the highest weight loss percentages reported in a clinical trial for an obesity drug, potentially surpassing existing treatments like semaglutide. If approved, retatrutide could offer a more effective option for weight management, impacting the obesity treatment landscape. The trial enrolled about 2,500 participants; 45.3% of those on 12 mg lost at least 30% of their body weight. Discontinuation due to adverse events was 4.1% for the 4 mg dose, lower than placebo (4.9%), with gastrointestinal side effects being common.
telegram · zaihuapd · May 22, 02:18
Background: Retatrutide (LY-3437943) is an investigational triple agonist targeting GLP-1, GIP, and glucagon receptors, developed by Eli Lilly. It is designed to reduce appetite and increase energy expenditure. Phase 3 trials are the final stage before regulatory approval.
References
Tags: #obesity, #clinical trial, #retatrutide, #pharmaceutical
Project Glasswing Update: 90.6% True Positive Rate for Vulnerabilities ⭐️ 8.0/10
Anthropic’s Project Glasswing update reports that its security analysis tool achieved a 90.6% true positive rate for high- or critical-severity vulnerabilities, as validated by six independent security research firms. This demonstrates that AI-assisted vulnerability detection can be highly effective, potentially transforming how open source software is secured at scale, which is critical as AI agents increasingly rely on such code. Out of 1,752 high- or critical-rated vulnerabilities assessed, 90.6% were valid true positives, and 62.4% were confirmed as high- or critical-severity. The tool is based on Anthropic’s Claude Mythos model.
hackernews · louiereederson · May 22, 19:31 · Discussion
Background: Project Glasswing is an initiative by Anthropic to secure critical open source software using advanced AI models. Traditional static analysis tools can catch common vulnerabilities but often miss complex or context-dependent issues. AI models like Claude Mythos aim to find such vulnerabilities proactively, offering a new layer of defense for software maintainers.
Discussion: Comments show mixed reactions: some users like mdeeks report high accuracy and adoption in their workflows, while others like mukmuk and demorro express skepticism about real-world efficacy compared to existing tools and question cost-effectiveness. The maintainer of curl, Daniel Stenberg, saw no significant improvement over existing tools.
Tags: #AI, #security, #vulnerability detection, #Anthropic
Why Japanese Companies Diversify So Much ⭐️ 8.0/10
An analysis argues that Japanese companies diversify across many fields due to lifetime employment and an insular corporate culture that prioritizes employee interests over shareholder value. This diversification strategy contrasts sharply with Western focus on core competencies, challenging assumptions about corporate efficiency and offering an alternative model for long-term stability. The system only works when companies are insulated from outside pressure, such as takeovers, and employees develop firm-specific skills that are not transferable to other employers.
hackernews · d0ks · May 22, 15:22 · Discussion
Background: Lifetime employment in Japan has historically fostered a workforce that stays with one company for decades, encouraging investment in company-specific knowledge. The keiretsu system, a network of interlocking business relationships and cross-shareholdings, further insulates companies from market pressures and enables long-term planning.
Discussion: Comments highlight that this analysis captures the core role of lifetime employment and insulation, but some note that Western companies were also diversified in the past. An East Asian commenter warns against romanticizing the system, pointing out subtle class issues.
Tags: #Japanese business, #corporate culture, #organizational behavior, #diversification, #lifetime employment
Antigravity 2.0 Tops OpenSCAD LLM Benchmark for 3D Models ⭐️ 8.0/10
Antigravity 2.0, an AI agent powered by Gemini 3.5 Flash, achieved the top score in the OpenSCAD LLM benchmark by generating complex 3D architectural models like the Pantheon, including interior details such as coffered ceilings. This benchmark demonstrates that LLMs are advancing in parametric 3D modeling, which could significantly accelerate architectural design and 3D printing workflows. It also highlights the potential of AI agents to autonomously handle intricate geometric tasks. The benchmark used the Pantheon as a single test case, which some commenters noted limits generalizability. Antigravity 2.0 was the only agent that implemented the Pantheon’s interior ceiling pattern with repeated square coffers visible through the oculus.
hackernews · jetter · May 22, 10:38 · Discussion
Background: OpenSCAD is a free, script-based 3D CAD modeller that uses its own description language for creating solid objects via constructive solid geometry. Antigravity 2.0 is Google’s AI agent platform that integrates with the Gemini API and provides CLI and IDE tools for orchestrating AI tasks.
References
Discussion: Community reactions were mixed. Some users praised Antigravity’s output quality, while others criticized rollout stability issues and noted that the benchmark’s single test case is insufficient to prove general superiority. A commenter also observed that Gemini models are the least ‘jagged’ across different 3D model types.
Tags: #LLM, #3D modeling, #OpenSCAD, #benchmark, #AI agents
yt-dlp deprecates Bun support over Rust rewrite and AI concerns ⭐️ 8.0/10
yt-dlp has deprecated support for the Bun JavaScript runtime, citing foreseeable compatibility and security issues stemming from Bun’s upcoming Rust rewrite and AI involvement. This decision affects many yt-dlp users who rely on Bun and highlights broader tensions in open-source communities regarding AI-driven development and major rewrites. The deprecation was announced before Bun’s Rust rewrite has even been released, leading to criticism that the move is based on politics rather than engineering evidence.
hackernews · tamnd · May 22, 17:24 · Discussion
Background: yt-dlp is a popular command-line video downloader for sites like YouTube. Bun is a fast JavaScript runtime initially written in Zig, but it is being rewritten in Rust. The rewrite and AI involvement (via Anthropic acquisition) raised concerns among maintainers about code reviewability and security of 1 million lines of new code.
Discussion: Community comments are divided. Some users agree with yt-dlp’s caution, citing the difficulty of reviewing 1 million lines of rewritten code. Others argue the decision is premature and based on political discomfort with AI rather than engineering facts.
Tags: #bun, #yt-dlp, #rust, #open-source, #deprecation
U.S. Researchers Face Unclear Restrictions on Foreign Collaborations ⭐️ 8.0/10
U.S. researchers are being informed individually by agencies like NIH and NASA that they cannot publish with foreign collaborators without prior approval, yet no official guidance has been publicly released. This lack of transparency creates confusion and could hinder international scientific collaboration, which is essential for research progress. The unclear policy may also affect U.S. competitiveness in science and technology. The restrictions apply to any publication involving a ‘foreign component’ and have been in place since at least 2003, but are now being enforced more strictly without formal public guidance. Researchers report being told about these rules on a case-by-case basis.
hackernews · ceejayoz · May 22, 16:23 · Discussion
Background: International collaboration is a cornerstone of modern research, with many papers involving co-authors from multiple countries. Previously, agencies like NIH and NASA permitted such collaborations more freely; the current shift toward tighter restrictions, without clear rationale, has caused concern in the academic community.
Discussion: Commenters express frustration at the lack of transparency and official guidance, calling the policy changes confusing and potentially harmful to science. Some highlight the asymmetry with countries like China, which do not extend similar openness to foreign researchers.
Tags: #policy, #research, #academia, #US, #international collaboration
AI’s HBM Demand Squeezes Consumer Memory, Driving Price Hikes ⭐️ 8.0/10
The growing demand for High Bandwidth Memory (HBM) in AI data centers is diverting wafer capacity from consumer DDR and LPDDR memory, leading to price increases in consumer electronics. Memory manufacturers are allocating an expected 20% of wafers to HBM by end of 2026, up from 2% previously. This shift means smartphones, laptops, and other consumer devices will likely become more expensive, especially impacting budget markets like sub-$100 smartphones in Africa and South Asia. It highlights how AI infrastructure investments can have ripple effects across the broader electronics industry. A single gigabyte of HBM consumes more than three times the wafer capacity of a gigabyte of DDR or LPDDR. Memory companies deliberately under-provision fabrication capacity to avoid overcapacity, constraining supply for years.
rss · Simon Willison · May 22, 22:01
Background: High Bandwidth Memory (HBM) is a 3D-stacked DRAM technology used in high-performance GPUs for AI and graphics, offering high data rates while saving space. Wafer capacity refers to the number of silicon wafers a fabrication plant can process, which is fixed in the short term. Only three major memory manufacturers remain (Samsung, SK Hynix, Micron), and they allocate wafer capacity among different memory types like DDR, LPDDR, and HBM.
References
Tags: #memory shortage, #HBM, #consumer electronics, #AI, #hardware pricing
BPF-based custom page-cache eviction proposed at Linux Summit ⭐️ 8.0/10
Tal Zussman proposed cache_ext, a BPF-based system that allows custom page-cache eviction policies to be loaded from user space without kernel changes, at the 2026 Linux Storage, Filesystem, Memory Management, and BPF Summit. This enables workloads like financial databases to avoid thrashing by tailoring page-cache eviction to their specific access patterns, potentially improving performance without application-level caching or direct I/O. cache_ext is implemented as a BPF struct_ops program with callbacks for folio addition, removal, access, and eviction requests, and is attached to control groups to manage per-workload policies.
rss · LWN.net · May 22, 14:37
Background: The Linux kernel’s page cache uses LRU (least-recently-used) or multi-generational LRU algorithms to manage folios, but these generic policies can be suboptimal for certain workloads. eBPF is a technology that allows safe and efficient extension of kernel functionality at runtime without modifying kernel source code or loading modules. The proposal is inspired by sched_ext, which enables custom CPU scheduling policies via BPF.
References
Discussion: During the session, Vlastimil Babka questioned why the existing access-twice heuristic fails for this scenario, and Zussman explained that concurrent scans fool it. Shakeel Butt asked whether the interface could manage all memory, not just the page cache; Zussman stated his focus is on file-backed memory for now.
Tags: #Linux kernel, #BPF, #page cache, #memory management, #storage
ByteDance Open-Sources Lance: 3B Unified Multimodal Model ⭐️ 8.0/10
ByteDance has open-sourced Lance, a 3 billion parameter multimodal model that natively unifies image and video understanding, generation, and cross-modal editing under the Apache 2.0 license. Lance’s lightweight yet unified architecture achieves competitive results on benchmarks like GenEval and VBench, potentially lowering the barrier for multimodal AI research and applications. Lance uses a shared-context dual-stream expert architecture, with Qwen2.5-VL and Wan2.2 encoders for understanding and generation respectively, and introduces modality-aware positional encoding to resolve sequence boundary confusion.
telegram · zaihuapd · May 22, 06:40
Background: Unified multimodal models aim to handle both understanding and generation tasks in a single framework, traditionally requiring separate models. Qwen2.5-VL is a vision-language model series by Alibaba Cloud, while Wan2.2 is a video diffusion model by the Wan-Video team using Mixture-of-Experts. Modality-aware positional encoding helps the model distinguish between different input types within a shared sequence.
References
Tags: #multimodal, #ByteDance, #open-source, #AI, #computer vision
China Cracks Down on Illegal Cross-Border Securities Trading ⭐️ 8.0/10
Eight Chinese government departments jointly issued a plan to crack down on illegal cross-border securities, futures, and fund operations, allowing existing investors only to sell holdings and withdraw funds. The CSRC has launched investigations into Tiger Brokers, Futu, and Changqiao, issuing administrative penalty notices. This marks a significant tightening of capital controls and enforcement against fintech platforms facilitating unapproved cross-border investments, affecting millions of Chinese investors and signaling a push towards legal channels like Stock Connect and QDII. The two-year rectification period limits existing investors to sell-only and fund withdrawals, after which all domestic websites, trading software, and servers must be shut down. The CSRC seeks to confiscate all illegal gains from Tiger, Futu, and Changqiao and impose severe penalties.
telegram · zaihuapd · May 22, 08:26
Background: Illegal cross-border securities operations involve foreign brokers soliciting mainland Chinese clients without approval. Legal channels for overseas investment include the Stock Connect programs (Shanghai-Hong Kong and Shenzhen-Hong Kong), QDII (Qualified Domestic Institutional Investor), and Cross-Border Wealth Management Connect (for Greater Bay Area). These regulated channels allow capital outflows under strict supervision.
References
Discussion: The Telegram post includes no community comments; it simply relays the regulatory announcement and separately mentions Lagou.com’s bankruptcy as additional context.
Tags: #regulation, #fintech, #cross-border investment, #China, #securities
Cloudflare global outage 25 min, 28% HTTP traffic affected ⭐️ 8.0/10
On December 5, 2025, at 08:47 UTC, Cloudflare’s global network experienced a 25-minute outage that impacted approximately 28% of HTTP traffic, caused by a fix for the Next.js security vulnerability CVE-2025-55182 (React2Shell). This outage impacted a significant portion of global HTTP traffic, highlighting the risk of emergency security patches causing collateral damage, and underscores the complexity of mitigating critical RCE vulnerabilities like CVE-2025-55182 in widely-used frameworks. The outage was triggered by Cloudflare’s WAF update to patch CVE-2025-55182, a critical pre-authentication remote code execution vulnerability in React Server Components and Next.js, and mainly affected customers using legacy FL1 proxy with Cloudflare managed rule sets.
telegram · zaihuapd · May 22, 16:15
Background: CVE-2025-55182, also known as React2Shell, is a critical vulnerability that allows unauthenticated remote code execution in React Server Components and Next.js applications. Cloudflare operates a global content delivery network and web security platform that includes a Web Application Firewall (WAF) to block such exploits. The outage occurred when a rule intended to mitigate this vulnerability caused unintended disruptions on the network.
References
Tags: #Cloudflare, #故障报告, #网络, #CVE, #WAF