Good morning. The government-and-money theme dominates today: OpenAI is dangling a 5% equity stake in front of the Trump administration, Microsoft is spinning up a $2.5B deployment arm, and Anthropic is chatting with Samsung about custom silicon. Zuckerberg, meanwhile, admits Meta’s AI reorg hasn’t gone great.
OpenAI floats a 5% cut for Washington. Sam Altman has proposed donating 5% of OpenAI’s equity — roughly $42.6 billion at the company’s $852B valuation — to a US sovereign wealth fund, with the expectation that other AI companies would follow suit. TechCrunch reports talks are early and would need congressional approval, but the pitch fits the Trump administration’s newly interventionist posture — a 10% stake in Intel, revenue cuts from Nvidia and AMD on China sales. Bernie Sanders’ counteroffer: a one-time 50% tax on AI company stock for the same purpose.
Zuckerberg concedes agents aren’t there yet. At an internal town hall, Mark Zuckerberg told Meta staff that AI agents “haven’t progressed as quickly as he’d hoped” despite the company laying off ~8,000 employees and reassigning 7,000 more into AI teams this year, TechCrunch reports. He called the restructuring not as “clean” as it should have been but expects improvements in three to six months. Meta still plans up to $145B in AI infrastructure spend in 2026.
Anthropic and Samsung talk custom silicon; Microsoft launches a deployment arm. Anthropic is in early discussions with Samsung on a custom AI chip, mirroring OpenAI’s Broadcom-built “Jalapeño” inference chip announced recently. Details — purpose, specs, integration — are all TBD, and Anthropic stresses it’ll keep buying from Google, Amazon, and Nvidia. Separately, Microsoft launched Microsoft Frontier Company, a $2.5B, 6,000-engineer enterprise deployment business that looks a lot like the Forward Deployed Engineer model AWS, OpenAI, and Anthropic have all adopted — though Satya Nadella took pains to reject the label. Early customers include LSEG and Unilever.
Cursor’s model access looks precarious post-SpaceX. Wired asks the awkward question after the $60B SpaceX acquisition of Cursor: will OpenAI and Anthropic keep serving models through a platform now owned by Elon Musk? Cursor is one of their biggest customers, but also now effectively a competitor’s product. A rival founder told Wired the situation is “super unclear.”
Japan’s Supreme Court says AI can’t be an inventor. Japan’s top court ruled that AI systems cannot be listed as patent inventors, aligning with US and other jurisdictions on the requirement of human legal accountability. The HN discussion mostly bypasses the ruling itself for a practical point: nothing prevents human users of AI tools from listing themselves as inventors, which could just accelerate mass filings by well-resourced companies. One commenter recommended Against Intellectual Monopoly, noting the empirical record on whether patents actually drive innovation is thin.
One transformer layer may be all RL needs. A new paper claims that training just a single middle transformer layer during RL post-training can match or beat full-parameter RL across Qwen2.5, Qwen3, multiple algorithms (GRPO, GiGPO, Dr. GRPO), and tasks from math to code. The gains concentrate in middle layers, with input and output layers contributing much less. The HN thread finds this intuitive — early layers handle syntax, late layers handle decoding, and RL shapes the planning-like middle — though one commenter flagged that response-length inconsistencies at the 3K token cap may be confounding the results.
Meta’s Pocket and a video hack for Claude. Meta quietly launched Pocket on June 29, an AI app for text-to-game generation built on its earlier Gizmo acquisition, complete with a social feed. And a small open-source project called claude-real-video extracts scene-change keyframes plus Whisper transcripts to let Claude, GPT, or Gemini “watch” videos. The HN reception was skeptical: commenters pointed out frames still get sent to Anthropic, that Gemini or local VLMs are cheaper for this, and that keyframes can’t capture motion or object permanence anyway.
That’s it for today. If OpenAI’s sovereign-wealth pitch actually clears congressional scrutiny, we’ll be having a very different conversation about who owns the AI boom — see you tomorrow.