AI news for Friday, September 4, 2026

Sanders and Casar introduced a bill that would ban superintelligent AI and pause frontier research until federal safety rules exist. Violators could face up to 20 years in prison; companies could face a corporate death penalty. It is an opening bid, not law.

Nvidia agreed to buy Hugging Face for $11.9 billion in cash plus roughly $1 billion in retention equity. Hugging Face says its hub will remain open to competing models and chips. The deal now tests how independent open AI's central platform can remain inside Nvidia.

MBZUAI's Institute of Foundation Models released six Apache 2.0 models spanning 0.9B to 375B parameters, with weights, code, training data and methods published. The smallest tiers claim state-of-the-art results at their sizes, and vLLM, SGLang, Ollama and Unsloth support them from day one.

South Korea's semiconductor exports tripled year over year in August to a record $46.65 billion - 47.5% of all goods exports. The concentration shows how powerfully hyperscaler AI spending is lifting Samsung and SK Hynix, and how exposed the wider economy would be to a sharp pullback.

BAAI's DisCo framework converts repositories and papers into reusable operational skills at about $40 per repository. Its authors report that adding those skills raised a baseline Codex agent's MLE-bench score from 31.1% to 72.9%. Those are paper-reported results, but the 5,000-skill library is available to inspect.

Nvidia researchers report that Nemotron-3-Ultra-CC scored 535.4 out of 600 on the IOI 2026 problem set, ahead of the top human contestant's 498.27. The result comes from Nvidia's own new arXiv preprint and has not been independently replicated.

GPT-6 Astra is live for a limited group, with paid ChatGPT plans, the API, Azure and Bedrock rolling out over days. It is also OpenAI's first model at the Critical cyber threshold - making access rules as consequential as the performance jump.

DeepSeek plans to deploy at least 160,000 Huawei AI accelerators at a 1GW site in Inner Mongolia, Bloomberg reports. If completed as described, it would be among the largest known Huawei AI clusters—and a major test of China's domestic compute stack.

NHTSA opened an investigation hours after Tesla's first production Cybercabs entered Austin without steering wheels or brake pedals. Tesla says the vehicles comply through self-certification; federal manual-control rules remain in force while regulators examine that claim.

Gimlet raised $300M at $3B, six months after an $80M round valued it near $400M. Andreessen Horowitz led, with Arm and Microsoft's M12 joining. The bet: route inference across competing chips instead of locking workloads to one stack.

Vals AI estimates some long agentic tasks carry roughly 10,000x a one-shot query's environmental footprint. Building one web app on some models matched 2.5 hours of home electricity. It is a task comparison, not a universal agent multiplier.

FineBooks tested 14 open OCR models on 2,165 expert-transcribed pages from 18th- and 19th-century books. A 3B-parameter model reached 97.6% reading accuracy, while two sub-2B models topped 96%. Better OCR for old books could unlock cleaner public-domain data for training and research.

The Last Translation Benchmark collects examples that break leading systems and pairs each with a failure rule. LTBv1 spans text, image, audio and video and accepts new cases - turning embarrassing edge failures into reproducible tests.

Ray Kurzweil joined Subsense, a startup developing a nonsurgical BCI using nasal nanoparticles and a magnetic-coil cap. The memorable concept is still preclinical: the company has raised $27M, but current experiments are in mice.