AI News: June 8, 2026

1. Trump Administration in Talks to Take an Equity Stake in OpenAI

President Trump confirmed discussions about a government equity arrangement with OpenAI, framing it as a way for “the American people to benefit from the success of AI,” with some equity potentially seeding a public wealth fund that distributes returns to citizens. The concept echoes earlier proposals from Sam Altman and follows the equity-for-support pattern the administration used with Intel in 2025. No stake size or terms have been disclosed, but a direct government holding in a frontier lab would mark a significant shift in how the U.S. relates to the AI industry. Source

2. Great American AI Act Draft Would Freeze State AI Laws for Three Years

A bipartisan discussion draft of the Great American AI Act drew sharp opposition this week for proposing a three-year moratorium on state laws that specifically regulate AI model development. The 269-page bill would require semi-annual third-party audits for frontier developers with over $500 million in revenue, create a Center for AI Standards and Innovation inside the Commerce Department, and allow penalties up to $1 million per day for non-compliance. Consumer and civil rights groups argue it converts a state consumer-protection floor into a weaker federal ceiling. Source

3. xAI Reportedly Trained Grok Coding Models on Claude Outputs

A report from The Information says xAI distilled Anthropic’s Claude outputs to train its coding models, and that after Anthropic revoked official API access in January 2026, engineers continued via personal accounts and the intermediary service Blackbox AI. Elon Musk acknowledged in court that xAI “partially” used OpenAI models in a similar fashion, characterizing it as common industry practice. The report also says xAI’s pretraining team shrank to under five engineers and that four Grok coding leads departed within months. Source

4. Deepseek Tops Ramp’s Corporate Spending Charts as U.S. Firms Chase Cheaper AI

Spending data from Ramp, drawn from more than 50,000 companies, shows Deepseek topping its June 2026 trending-vendor list following the April launch of Deepseek V4 at pricing well below Western alternatives. Ramp’s chief economist flagged concerns about U.S. companies routing sensitive data through a Chinese-owned platform as adoption rebounds. The data point lands alongside reports that Chinese models now account for over 44 percent of downloads on Hugging Face, surpassing U.S.-origin models. Source

5. Sakana AI Launches RSI Lab to Pursue Recursive Self-Improvement

Sakana AI, founded by Transformer co-author Llion Jones and former Google Brain director David Ha, established a new RSI Lab to build systems that iteratively redesign their own code, benchmarks, and architectures as an alternative to brute-force compute scaling. The lab outlined a four-phase roadmap progressing from agent-native models to full recursive self-improvement, citing prior projects such as the Darwin Godel Machine and The AI Scientist as early evidence the approach is “no longer purely theoretical.” The bet is that efficiency, rather than raw scale, can democratize access to frontier capability. Source

6. Anthropic Hires OpenAI’s Second-Ever Chip Engineer

Anthropic recruited Clive Chan, formerly the second hardware hire in OpenAI’s custom chip program and previously a chip engineer on Tesla Autopilot, for an internally coined “perplexity per picojoule” role focused on maximizing model performance per unit of energy. The hire follows April reports that Anthropic was exploring proprietary AI silicon while currently running Claude on Google TPUs and Amazon chips. Custom inference hardware would help margins as the company is reported to be approaching an IPO. Source

7. Researchers Explain Why Emergent Abilities Appear Only in Large Models

A study covered this week argues that small models fail at rare tasks not because they lack capability in principle, but because frequent tasks repeatedly overwrite the learned weights for uncommon patterns during training. Larger models have enough capacity to anchor frequent tasks first, leaving room for rare skills to accumulate without interference. The practical takeaway is that increasing the frequency of rare tasks in training data can be a cheaper path to new capabilities than simply scaling model size. Source

8. Open-Source Audio-Interaction Model Does Continuous Real-Time Speech

Researchers from China, Hong Kong, and Singapore released Audio-Interaction, a 3-billion-parameter model that processes a live audio stream in 0.4-second chunks, emitting either a silent or a response token each cycle to enable continuous conversation without waiting for speaker pauses. The model handles transcription, translation, dialog, and sound recognition in a single system and scores 58.15 on the MMAU benchmark after training on 302,000 hours of synthetic audio. Code and weights are available under Apache 2.0, with the full training dataset to follow. Source

9. White House AI Advisor Sriram Krishnan to Step Down

Sriram Krishnan, who spent roughly 18 months as senior AI policy advisor in the Trump administration, will leave his post at the end of June 2026 to found an external institution focused on AI policy, energy, and data-center infrastructure. During his tenure he oversaw the AI Action Plan, which prioritized data-center development over regulation, and executive orders challenging state-level AI laws. No successor has been named. Source