NVIDIA GTC Taipei: hardware, AI factories & robotics
GTC Taipei turned into an end-to-end stack reveal. NVIDIA confirmed Vera Rubin is in full production as the POD-scale platform for agentic workloads (@nvidia), and introduced Vera BlueField-4 STX with DOCA security pitched as "secure-by-design storage for agentic AI factories" — claiming up to 1,000x faster runtime threat detection than agentless tools and in-silicon enforcement at 800Gb/s (@nvidia). On the deskside, the new DGX Station for Windows on GB300 is pitched as a 1-trillion-parameter local box built by ASUS, Dell, GIGABYTE, HP, MSI and Supermicro, paired with an "RTX Spark" 1-petaflop superchip narrative (@nvidia).
Robotics and autonomy got equal billing. NVIDIA unveiled Alpamayo 2 Super, a 32B-parameter open VLA reasoning model for L4 robotaxis, alongside an expansion of DRIVE Hyperion with HUMAIN, VinFast/Autobrains and Uber/Autobrains targeting Southeast Asia, the Middle East and Europe (@nvidia). The Isaac GR00T Reference Humanoid Robot was announced as the first open humanoid reference design, and Jensen released an open-source collection of "physical AI agent skills" spanning Omniverse, Cosmos, Isaac, Metropolis and Jetson (@nvidia).
Enterprise and vertical plays rounded it out: NemoClaw, OpenShell and Nemotron models for AI coworkers, with Cadence, Dassault, Siemens and Synopsys building autonomous AI engineers (@nvidia); TSMC adopting CUDA-X, Metropolis, TAO, Omniverse and Cosmos for chip design and fabs (@nvidia); and a "Healthy Taiwan" partnership with Foxconn deploying Nurabot nursing robots (@nvidia). DSX MaxLPS and DSX OS landed as open-source infra software for AI factory builders (@nvidia).
Open-weight frontier models
The open-weights tier got crowded fast. NVIDIA's Cosmos 3 launched as "the world's first fully open omnimodel with native vision reasoning, world and action generation," shipping Super (32B) and Nano (8B) variants with weights and post-training recipes on Hugging Face (@huggingface, @_akhaliq). vLLM shipped day-0 support via vLLM-Omni — a unified Mixture-of-Transformers fusing an AR reasoner and diffusion generator across text, image, video, audio and robot action behind one OpenAI-compatible API (@vllm_project).
MiniMax M3 dropped as the "first open-weights model to combine three frontier capabilities," posting 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1 and 74.2% MCP Atlas, with sparse attention scaling to 1M context (@ollama). Ollama is hosting M3 in US-based zero-retention cloud, wired into Claude Code and Codex (@ollama). Sapient's HRM-Text claimed competitive performance from a 1B reasoner trained on 40B tokens for ~$1,000 (@clementdelangue), and Surya OCR 2 trended on Hugging Face (@_akhaliq).
Agent skill security & ClawHub/SkillSpector dataset
OpenClaw, NVIDIA and Hugging Face open-sourced security scans for 67,453 ClawHub skills (@huggingface, @steipete). Headline findings: NVIDIA SkillSpector flagged about half for agentic risk, only 0.31% were actually malicious, and no two scanners agreed on more than 8.5% of risks — a brutal signal on the immaturity of agent-skill scanning (@openclaw via @steipete). ClawHub now layers VirusTotal, static analysis and SkillSpector into a single ClawScan, plus an LLM-driven auto-approval flow for safer YOLO mode (@steipete).
The threat side stayed loud too: Dutch authorities dismantled a 17M-device botnet backed by 200+ servers in the Netherlands (@thehackersnews), and NetworkChuck reminded readers how ARP spoofing still seeds local-network interception (@networkchuck).
Coding agents & the AI developer stack
Codex crossed 5M users — about 0.6% of ChatGPT's ~900M — with Sam Altman resetting limits to celebrate (@sama, @steipete). Practitioners are stacking tools: Codex for vibe coding, Claude Code with Opus 4.8 for hard bugs, a Hermes orchestrator, and a local Mac Studio/DGX Spark model for repetitive work (@alexfinn). steipete is turning Codex into a background QA assistant using webVNC and computer-use to test like a real user and open fix PRs (@steipete), and reports Codex authoring ad-hoc codemods for a TypeScript migration (@steipete). swyx warned that without diligence, AI-written code regresses a repo "to your worst engineer" because each bad pattern poisons the next agent's context (@swyx). Fireworks hit $800M ARR with 4x growth (@_akhaliq).
LLM skepticism, consciousness debate & AI policy
Gary Marcus opened multiple fronts: pushing back on Geoffrey Hinton's claim that consciousness may already be inside AI ("we are creating interactive fiction… not beings"), reviving his January 2025 "too big to fail" bailout warning, and arguing LLMs "can't play nicely with basic tools like databases" (@garymarcus). Emollick countered that fully autonomous /goal agents are the wrong model — good AI should know when to ask good questions (@emollick). swyx flagged a vibe shift toward personal/local agents, noting PewDiePie's vibecoded productivity suite hit >10k stars in a day (@swyx). Sam Altman pitched OpenAI helping the world "get a head start on biodefense" (@sama).
Physical AI, robotics & generative media
OpenAI Robotics is hiring full-stack hardware, ops and ML engineers for its world-simulation program led by Aditya Ramesh, aiming at infrastructure-skilled-worker robots short-term and personal robots long-term (@gdb). fal launched Grok Imagine Video 1.5 for prompt- or frame-to-video (@fal), and AMD's Lemonade SDK contest is handing out Strix Halo laptops to on-device AI builders (@the_only_signal).
The Bottom Line
Today belonged to NVIDIA: a wall-to-wall GTC Taipei stack from Vera Rubin silicon through robotaxis, humanoids and an open Cosmos 3 omnimodel, plus a notable agent-security dataset collab with Hugging Face and OpenClaw. Underneath, open-weights momentum (MiniMax M3, HRM-Text, Cosmos 3) and a maturing multi-agent coding stack are colliding with louder skepticism about LLM foundations and the security posture of an agent-skill ecosystem where scanners barely agree.