Desk
Broad artificial intelligence coverage.
A Hugging Face blog post describes using ML Intern, an agent mode in HuggingChat, to plan, train, evaluate and publish six models in days, including a citrus-disease classifier that rose from 14.9% to 52.8% accuracy for about $1.90. Baselines and smoke tests were the key discipline.
Nous Research confirmed a $90 million Series B at a $1.5 billion valuation led by Robot Ventures, with Nvidia, USV, Menlo, Samsung and 1789 Capital participating, and launched Hermes for Businesses for private, customized agents. Its usage claims are company estimates.
Haiku 5.5 makes a capable model with a million-token memory nearly free for everyday work. Cheap intelligence at this price changes what gets automated, and how much.
TechCrunch reports Microsoft’s Surface Laptop Ultra starts around $2,600 and tops out near $5,900, with a roughly $6,000 RTX Spark Dev Box for developers. Satya Nadella said Execution Containers, which sandbox AI agents, will come to all Windows 11 users.
Nvidia says RTX Spark laptops pairing a Blackwell RTX GPU and Grace CPU with up to 128GB of memory ship October 16, with compact desktops in November. Microsoft made Execution Containers generally available, and Nvidia previewed a 748GB DGX Station for Windows.
WIRED reports three Axiom engineers wired a 2024 Toyota Corolla’s steering and cameras to a chat interface. GPT-6 Astra slowly guided it to an In-N-Out window with a safety driver aboard, and was the only model to finish their DrivingBench course. The models initially declined motion commands.
Liquid AI's new models do not chat at all. They answer fixed questions with probabilities in milliseconds, and that unglamorous design may be what makes AI safe enough to run everywhere.
NVIDIA's robotics lab tried to teach robots to assemble its own AI hardware and found that classical engineering often beat end-to-end learning. That honesty is the most valuable part.
OpenAI is rolling out Intelligent UI on GPT-6, letting ChatGPT generate custom visual elements such as tappable diagrams, slider calculators and clickable maps when they help. WIRED’s hands-on found real usability gains, while polished widgets can hide the assumptions behind an answer.
Microsoft Research Asia open-sourced Agent Lightning v1.0, a roughly 3,500-line framework built on “Harnessed Agentic RL,” in which an agent’s deployment harness takes part in training via an OpenAI-compatible proxy. A Qwen3.5-9B coding agent rose from 41.8% to 56.4% Pass@1 on SWE-bench Verified with about 6,000 samples.
Nvidia wants AI agents testing data center changes in a simulated copy before anything touches real hardware. That is the right place to let agents experiment, if the twin stays honest.
NVIDIA's multi-GPU solver makes huge planning problems for supply chains and power grids faster to solve, though only past a certain size, and its own benchmarks show where it loses.