2026-07-11

Surgeons remotely controlled a Unitree G1 humanoid robot to perform two laparoscopic gallbladder removal surgeries on live pigs, marking the first use of a general-purpose humanoid robot for live surgery. The results were published in Nature. This demonstrates a low-cost alternative to specialized surgical robots like the da Vinci system, potentially expanding access to remote and resource-limited settings such as rural areas, battlefields, or space. It could democratize advanced surgical care. The Unitree G1 base model costs $13,500, and with dexterous hands it costs about $67,000, far less than the $500,000+ for dedicated surgical robots. The robot is 1.5m tall and weighs 27kg.

Binarly disclosed six vulnerabilities in U-Boot's FIT signature verification code, two enabling arbitrary code execution and four causing denial of service. The flaws affect over 50 stable versions dating back to 2013.07 and numerous downstream branches. Attackers can execute malicious code before the operating system boots, bypassing all security measures and potentially installing persistent firmware malware. For systems with remote firmware update capabilities like BMCs, exploitation requires no physical access, posing a severe threat to embedded and server ecosystems. The vulnerabilities reside in how untrusted FIT images are processed before signature validation completes. Remote exploitation is possible via malicious firmware updates, and while patches have been accepted into U-Boot, each hardware vendor must integrate and distribute them—legacy devices may never receive fixes.

vLLM v0.25.0 makes Model Runner V2 the default for all dense models, removes the legacy PagedAttention implementation, and achieves performance parity between the Transformers modeling backend and native vLLM. This release is a major architectural overhaul that simplifies the codebase, eliminates technical debt, and broadens model support, benefiting the LLM inference community with improved performance and compatibility. The release includes 558 commits from 232 contributors, adds new models like LLaVA-OneVision-2 and GLM-5, and introduces a new Streaming Parser Engine for tool-call/reasoning parsing.

SGLang v0.5.15 introduces Speculative Decoding V2 with zero-overhead scheduling, achieving +11% end-to-end tokens per second, and IndexShare Multi-Token Prediction (MTP) that reduces draft-step cost by up to 1.9x on long contexts. Additionally, GLM-5.2 with NVFP4 quantization on Blackwell GPUs now serves over 500 tokens per second per user on 8x B300. These optimizations significantly reduce inference latency and cost for production LLM serving, especially on NVIDIA's latest Blackwell architecture. The performance gains in speculative decoding and quantization make high-throughput, low-latency serving more accessible for large models like GLM-5.2, DeepSeek-V4, and others. Key improvements include Spec V2's CUDA-graphable draft-extend with fused metadata ops, IndexShare MTP reusing indexer top-k across steps, TopK V2 fusing top-k with page-table transform, and GEMM shape-specialized JIT routers for Blackwell. The release also enables breakable CUDA Graph by default and adds linear-attention kernels (KDA/GDN) for alternative attention mechanisms.

ClickHouse published a blog post describing how they scaled PgBouncer, a popular PostgreSQL connection pooler, to 4x its original throughput by enabling SOREUSEPORT and implementing peering between processes. PgBouncer is widely used in PostgreSQL deployments, and this performance improvement benefits many users who rely on it for connection pooling. The techniques used can inspire similar optimizations in other network services. SOREUSEPORT allows multiple processes to bind to the same TCP port, distributing incoming connections evenly. Peering forwards cancellation requests to the correct process, eliminating bottlenecks and improving scalability.

An article recommends using STRICT tables in SQLite, introduced in version 3.37.0 (2021-11-27), to enforce data type constraints and improve database reliability, especially when multiple applications share the same database. This tip addresses a common criticism of SQLite's dynamic typing, making it more robust for multi-application scenarios and encouraging safer database practices. It could drive adoption of STRICT tables and potentially influence future defaults. STRICT tables enforce exact data types and disallow type coercion, but they do not support some types like DATE. Each table must be explicitly declared as STRICT.

The VultronRetriever family of models has been released on HuggingFace, achieving 1 on the MTEB leaderboard with up to 16x smaller index storage and 12x higher throughput compared to previous 9B-class leaders. This breakthrough enables high-quality retrieval and embedding on edge devices like iPhones entirely offline, expanding AI applications to privacy-sensitive and resource-constrained environments while reducing infrastructure costs. The family includes three models: Prime-8B (global 1), Core-4.5B (outperforming models twice its size), and Flash-0.8B (outperforming models up to 5x its size). All models are trained on contamination-free datasets and leverage the Hydra architecture for late interaction retrieval with up to half the memory of comparable models.

SK Hynix CEO Kwak Noh-Jung warned that the global memory industry will face the worst supply shortage in history by 2027, with demand outstripping supply even after capacity expansion beyond 2030. The warning came on the company's Nasdaq debut day, where shares closed up 13.3% at $168.85. This forecast from a major memory maker signals that AI and computing demand will drive long-term memory tightness, potentially raising hardware costs and impacting the entire tech industry. As a leading player, SK Hynix's outlook carries significant weight for supply chain planning. SK Hynix reported a record operating profit of 47 trillion won ($31 billion) in 2025, with Q2 2026 expected to rise further to 65.5 trillion won. The company is considering overseas fab locations in the US, Japan, or Southeast Asia, prioritizing regions with cost advantages in land, power, and labor.

Apple filed a lawsuit on July 10, 2026, in the U.S. District Court for the Northern District of California against OpenAI, two former employees, and io Products, accusing them of systematically stealing trade secrets related to product design, manufacturing processes, and supply chain to accelerate OpenAI's consumer hardware business. This lawsuit could set a significant precedent for how AI companies compete with established tech giants in hardware, potentially impacting OpenAI's hardware ambitions and the broader industry's approach to intellectual property protection. Apple alleges former employee Chang Liu downloaded dozens of hardware files after leaving the company, and hardware head Tang Yew Tan sent supplier information to his personal email and encouraged job candidates to bring Apple parts to interviews. Apple also claims over 400 former employees now work at OpenAI.

Apple has agreed to use Intel's manufacturing facilities for some of its chips, following pressure from the Trump administration. The U.S. government converted $9 billion in federal grants into a 10% stake in Intel, becoming its largest shareholder, and facilitated deals with Nvidia and SpaceX. This marks a significant government intervention in the semiconductor industry, potentially reshaping global chip manufacturing dynamics and strengthening U.S. supply chain resilience. It also highlights the strategic importance of reviving domestic chip production amid geopolitical tensions with China. The $9 billion conversion came from CHIPS Act funds, and Intel's new CEO Lip-Bu Tan meets monthly with Commerce officials. Intel's stock has tripled since March 2025, and the company's foundry services have secured major customers like Nvidia and SpaceX.