[AINews] Jensen Huang resigns after 33 years leading NVIDIA
Source
latent.space
Author
swyx .ai
Date
Key takeaways · AI-distilled
Artificial Analysis measured Gemini 4 Hydrogen at an 8% hallucinationWhen a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.Full definition → rate on AA-Omniscience, against 43% for Claude Sonnet 5.5 and 52% for GPT-6.1 Sol, but at 54% on Terminal-Bench 4.0, below Sonnet's 64%.
Simon Willison notes Hydrogen's $1.50/$7.50 per million token price is only a 25% discount on Carbon and Argon, though those models reportedly double in price on Dec 1. One tester saw xhigh spend 131K thinking tokens without answering.
Will Depue claims initialization patterns in released weights reveal the GPU type that made them, placing six Chinese models on Blackwell. Critics say his 2-3K Blackwell estimate is about 6MW, small next to China's compute, so controls may be working.
Thinking Machines released Inkling 2 Small under Apache 2.0: a 594B-total, 17B-active mixture-of-expertsA model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.Full definition → with up to 1M context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →. vLLM reports it fits on one GB200/GB300 node and reaches up to 630 tok/s per user at TP4.
Anthropic opened a private beta of self-hosted Claude Managed Agents, letting customers run the agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition →, sandboxes and session history in their own cloud or on-prem, with agent identity, OTel, kill switches and an evals product.
Terms in this piece · Glossary
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
