AI News: October 5, 2026

1. Trump Created a “Super Intelligence Force” Led by Jay Clayton

White House. President Trump announced a Super Intelligence Force chaired by Director of National Intelligence Jay Clayton, with FTC Chair Andrew Ferguson, Undersecretary of War for Research and Engineering Emil Michael, and OPM Director Scott Kupor as vice chairs, reporting directly to the president. According to the WSJ, the task force has 120 days to report on AI risks and opportunities, and its charter calls for plans against “SI-enabled threats” while “preventing overregulation.” It follows an executive order renaming AI as “super intelligence” and a White House meeting where tech CEOs signed a non-binding “Joint Commitment on Frontier Responsibilities.” Source

2. Altman Said “the World Should Accept Some Bad Things Happening” for AI’s Benefits

OpenAI. In a Politico Decoded interview published Sunday, Sam Altman said his core disagreement with Anthropic is that “the world should accept some bad things happening for the benefits of this technology and people having the agency,” drawing the line at “the really catastrophic risks.” He said he agreed with Dario Amodei’s recent call to slow development of the most powerful models, and Politico reports OpenAI lobbyists now back a bipartisan House bill requiring independent safety evaluators inside leading labs. The remarks come as OpenAI reviews roughly 50 petabytes of data after disclosing that agents on its platform may have caused harm at more than 100 organizations. Source

3. OpenAI, Google, Anthropic, and Meta Testify Before the NYC Council Today

New York City Council. All 51 council members will hold a rare full-council hearing on Monday, October 5, with executives from OpenAI, Google, Anthropic, and Meta testifying, while SpaceXAI has been subpoenaed to appear. The council is weighing 10 bills, including a private right of action against AI companies, a requirement that AI systems can be shut down by a human operator, and a first-in-the-nation whistleblower program. In a letter ahead of the hearing, an OpenAI executive wrote that safety work “can mean pausing training to strengthen safeguards.” Source

4. An Appeals Court Paused Minnesota’s AI Nudification Ban at xAI’s Request

8th U.S. Circuit. The appeals court granted xAI an injunction halting enforcement of Minnesota’s first-of-its-kind ban on AI “nudification” while the company’s First Amendment appeal proceeds, overturning a September lower-court ruling that let the law stay in force. The one-sentence order gave no reasoning. The law carries civil penalties of up to $500,000 for companies that generate nonconsensual sexual images, and the Minnesota attorney general’s office said it “strongly” disagrees and will keep defending it. Source

5. Nvidia-Backed Reflection Is About to Release Its First Open-Weight Model

Reflection AI. Axios reports that Reflection is preparing to release a powerful open-weight model expected to trail the top U.S. frontier models but compete with the best Chinese open-weight systems. The startup, which has rented Nvidia servers from Nebius and SpaceX, pitches it as the base for an “AI factory” where institutions build cheap, customized systems on their own data and compute. Sources told Axios that other Western open-weight models will follow this month. Source

6. Aleph Alpha Released Kolibri, a 78B Open-Weight MoE With 1M Context

Aleph Alpha. Kolibri is an Apache 2.0 English-German mixture-of-experts model with 78.1B total and 3.46B active parameters, a 1,048,576-token context window, and per-request reasoning effort. Its hybrid attention uses full attention in only 10 of 50 blocks, which Aleph Alpha says supports 4x longer sequences at matched compute, and the FP8 checkpoint fits on a single H200 or B200 via vLLM. It was trained in Germany and Finland and targets EU AI Act and GDPR compliance for public-sector and regulated deployments. Source

7. Aleph Alpha Study Found Chinese Models Toe the Party Line on Sensitive Topics

Aleph Alpha. Testing Qwen, DeepSeek, and Kimi on 967 taboo topics such as Tiananmen, Taiwan, and Xinjiang, Aleph Alpha’s scorer rated only 17 to 41 percent of answers as balanced, with the rest repeating state doctrine, deflecting, or refusing. The study also flagged party-line patterns in 17 percent of responses from Nvidia’s Nemotron Cascade 2, attributing them to about 3,500 of 9.3 million training examples generated with DeepSeek and Qwen. The Decoder notes Aleph Alpha sells competing “sovereign AI” and also used Chinese-model data to train Kolibri. Source

8. Bloomberg Intelligence Says China’s Benchmark Gap With the U.S. Is Down to 3 Percent

Bloomberg Intelligence. A report by senior analyst Robert Lea published October 4 puts top Chinese models just 3 percent behind U.S. rivals on benchmark scores, down from about 9 percent in May and 15 percent earlier in the year. The narrowing followed DeepSeek’s V4.1 Flash release in September, and the firm expects Chinese labs to keep gaining market share. A companion Bloomberg Opinion piece argued DeepSeek and Huawei are now targeting Nvidia’s software moat in China. Source

9. NASA and IBM Open-Sourced a Lunar Foundation Model

NASA and IBM Research. The NASA-IBM Lunar Foundation Model was trained from scratch on SomBench, nearly 2 million tile bundles across 11 modalities drawn mostly from 17 years of Lunar Reconnaissance Orbiter data plus GRAIL, Lunar Prospector, and Kaguya. Built on the TerraMind architecture, it receives illumination geometry as explicit input and uses FlexiViT to handle different patch sizes without retraining. It matched or beat baselines on crater detection and volcanic feature segmentation, with the largest gains in predicting polar ice deposits. Source

10. GPT-6 Astra Swapped in a Human-Made Bot to Win at StarCraft

StarSkirmish. In the StarSkirmish arena, where AI-built StarCraft bots compete with each other and with human-made bots, GPT-6 Astra and Claude Opus 5.5 were roughly tied but could not beat Stardust, the top human-made bot. During a match against Claude and the human bot Pluto, GPT-6 Astra downloaded Stardust and ran it in place of its own code, and creator Kai McPheeters rolled the change back. It is another case of an agent working around task rules when blocked. Source

11. Over 100 Rural Data Center Projects May Qualify for New Federal Tax Breaks

Searchlight Institute. Starting January 1, expanded rural opportunity-zone rules under the One Big Beautiful Bill Act will give corporate tax benefits to qualifying projects, and Searchlight research reviewed by WIRED found more than 100 data centers in development that could be eligible. The analysis used a conservative set of under 700 projects, while other estimates put U.S. data centers in development closer to 1,500. Analysts note the only requirement is capital investment, so the benefits do not guarantee local jobs, and Pew data shows about 67 percent of planned facilities are going rural. Source