Desk
Broad artificial intelligence coverage.
TechCrunch's events desk highlights five safety sessions at Disrupt 2026 that frame trust as the gate between a demo and a paying customer, covering enterprise Claude deployments, agent permissions, cloud security, autonomous systems and the data gap holding back robotics. It is a promotional agenda piece, not independent safety research.
NVIDIA’s DLSS 5 bets that neural rendering stays grounded in the engine frame — lifelike lighting without rewriting the art bible, shipping now in NBA 2K27.
WIRED reports that The Biological Computing Company is offering a video-generation model derived from studying rat neurons in a limited AWS preview, claiming up to five times faster generation than an unnamed open-source model. The startup has raised more than $50 million, and AWS says a broader enterprise rollout is expected.
TechCrunch reports that British neocloud Nscale's IPO filing shows more than $103 billion in contracts, roughly 85% tied to Microsoft and Anthropic, with Anthropic's deal contingent on Nscale securing financing. The company posted $140.6 million in first-half revenue and a $1.02 billion net loss.
Nvidia released Isaac ROS 5.0 at ROSCon in Toronto, adding ROS Lyrical and Ubuntu 24.04 support, agent-ready skills for setup, perception fine-tuning and pick-and-place, and Jetson support from Orin Nano to Thor. Partner examples and performance figures are Nvidia's.
NVIDIA’s tutorial shows an AI agent migrating a ROS 2 node to CUDA-backed buffers — zero-copy where it counts, CPU fallback where it must.
In an OpenAI customer story, Parallel Web Systems says GPT-6 Astra cut research task time and cost by about 50% versus prior models on a labor-market research job, attributing the gain to more targeted queries and fewer research calls. The figures are from OpenAI and Parallel; full methodology is not published.
The Verge reports that Meta patched its Muse macOS app after researcher Patrick Wardle found a local zero-day that could let attackers take control of the AI agent through an undocumented setting. Meta's David Singleton called it local privilege escalation with 'quite low' practical risk before confirming the hotfix.
In MIT Technology Review, Timnit Gebru and Emily M. Bender argue that recent AI headlines, from Claude Mythos vulnerability claims to OpenAI's math announcements and a viral resignation warning, were marketing framed as civilization-scale breakthroughs. They say talk of rogue models shifts accountability away from companies.
NVIDIA’s SoL-Pi trims coding-agent token traffic nearly in half by evolving the harness — same agent brain, fewer wasteful round trips.
Grok 4.7 ships a larger base and longer RL run at Grok 4.6’s $2/$6 pricing — a price-performance shove, not a clean sweep of the leaderboard.
Hugging Face says its transformers library can now load GGUF quantized models with from_pretrained and run them using llama.cpp's ggml kernels, starting with Qwen3.5 architectures on Apple Silicon. Benchmarks on an M2 Max show speeds close to llama.cpp, though Hugging Face still recommends llama.cpp for pure local efficiency.