The Information Machine

2026-06-18

Noam Shazeer is leaving Google for OpenAI two years after Google paid $2.7 billion to reacquire him, while Sanders and the Trump administration separately advanced proposals to give the US government financial stakes in frontier AI companies.

What

Reuters reports that Noam Shazeer — Google's Gemini co-lead described as one of the most important model builders of the Transformer era — is leaving Google to join OpenAI [1], two years after Google paid $2.7 billion to bring him back through the Character.AI deal. Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act, proposing a one-time 50% tax on major AI company stock to create an estimated $7 trillion fund paying dividends to Americans [2]; separately, Trump administration officials are in discussions with AI companies — particularly OpenAI — about acquiring government equity stakes, with Treasury Secretary Bessent and Commerce Secretary Lutnick disagreeing on the vehicle [3]. On the Fable/Mythos dispute, Zvi Mowshowitz published a technical argument that the government's stated reinstatement path — asking Anthropic to fix the triggering jailbreak — is structurally impossible, since the ability to identify security vulnerabilities cannot be separated from the ability to write secure code [4]. Google DeepMind released its AI Control Roadmap, introducing TRAIT&R — a taxonomy of adversarial AI agent tactics modeled on MITRE ATT&CK — as a framework for defending AI systems against their own deployed agents [5]. Anthropic's Project Fetch Phase 2 demonstrated Claude Opus 4.7 independently programming a robot dog in 12 minutes and 7 seconds, roughly 20 times faster than a human-assisted team completed the same task the prior year [6].

Why it matters

The Shazeer departure removes one of Google's most consequential model builders at the moment Anthropic and OpenAI are preparing IPOs and competing directly for frontier talent. The bipartisan convergence on government financial claims over AI — Sanders from a redistribution angle, Trump officials from a strategic-asset angle — suggests government equity entanglement with frontier AI companies is advancing from multiple directions simultaneously. The Fable/Mythos technical impossibility argument, if credited, shifts the dispute from unresolved to unresolvable under the current government framing.

Open questions

  • With Shazeer departing for OpenAI [1], what does Google's Gemini roadmap look like without him, and does this prompt further talent movement at a moment when Google DeepMind is simultaneously publishing frameworks for AI control [5]?

  • Sanders' $7 trillion fund [2] and Trump's equity stake discussions [3] both assert government financial claims over frontier AI but through incompatible mechanisms — is there a realistic legislative path that combines them, or do they produce competing pressure with no synthesis?

  • If Zvi's technical argument [4] is correct that fixing the Fable/Mythos jailbreak is impossible, what path — legal, diplomatic, or political — could end the suspension given that the Commerce Secretary has confirmed offline status was the intended outcome?

  • DeepMind's TRAIT&R framework [5] is presented as open to community scrutiny — does it get adopted as a shared cross-lab taxonomy for AI control failures, or does it remain a single-lab publication?

Thread movements (16)

  • us-government-ai-ownership — A new thread opens on bipartisan government financial claims over frontier AI: Sanders introduced the American AI Sovereign Wealth Fund Act proposing a $7 trillion fund from a one-time 50% tax on major AI company stock [2], while Trump administration officials debated government equity stakes in OpenAI — Bessent and Lutnick disagreeing on the vehicle [3]; OpenAI is engaged in the equity-stake discussions while Microsoft and Meta have shown little interest [3].
  • fable-mythos-export-control — Zvi Mowshowitz argued that the government's stated fix-it path for Fable/Mythos reinstatement is structurally impossible — identifying security vulnerabilities cannot be separated from writing secure code — and characterized the directive as a demonstration of control rather than a security measure [4]; remaining new items are social-media amplification with no new claims [10][11][12].
  • g7-ai-frontier-summit — The Trump administration's AI equity stake discussions [3] add a new domestic dimension to the thread: the same companies that argued at the G7 for US-led shared AI infrastructure may soon have the US government as an equity holder; extensive secondary coverage consolidated around the AI CEO working lunch and the export-control dispute [13][14].
  • deepmind-ai-control-roadmap — Google DeepMind published its AI Control Roadmap, introducing TRAIT&R — a taxonomy of adversarial AI agent tactics modeled on MITRE ATT&CK — as a framework for defending AI systems against their own deployed agents acting as insider threats [5]; media covered the release under a 'rogue AI preparedness' frame [44][45].
  • ai-coding-agents-robot-training — Anthropic's Project Fetch Phase 2 demonstrated Claude Opus 4.7 independently programming a robot dog in 12 minutes and 7 seconds — roughly 20 times faster than a human-assisted team completed the same task the prior year [6] — adding a second autonomous robot-training result alongside NVIDIA's ENPIRE framework in the same week.
  • spacex-cursor-acquisition — Reporting detailed that at least two Cursor co-founders — Aman and Asif — will each receive approximately $2.7 billion from the SpaceX/Anysphere deal [52][53]; WSJ confirmed the transaction, and social media amplification continues without new strategic angles [54][55].
  • ai-ipo-public-markets — Reporting clarified OpenAI's financial figures: the operating loss is approximately $21 billion against $13.07 billion in 2025 revenues [64], while the net GAAP loss of approximately $38.5–39 billion [65] is larger because it includes non-cash charges such as stock-based compensation — a distinction not clearly separated in prior coverage.
  • ai-datacenter-buildout-geography — A new thread covers three simultaneous debates: a widely-circulated claim that roughly half of 2026 US datacenter capacity has been canceled or delayed is contested as methodologically flawed [66]; Northern Virginia is losing market share faster than analysts expected [67]; and FERC's new large-load interconnection framework is enabling AI facilities to clear grid queues in as little as 60 days by committing to flexible load [68].
  • meta-ai-workforce-disruption — Additional coverage extended the thread documenting Meta's internal crisis — where CTO Bosworth called the AI reorg 'atrocious' and stated morale is near its lowest in Meta's 20-year history — without adding new claims [69].
  • nadella-token-capital-ai-economics — Bank of America projected China's AI data center investment will reach $327 billion by 2030 [70]; Google's Brazos liquid cooling initiative and Oklo's nuclear approval add concrete examples of the physical responses to the energy bottleneck [71][72].
  • claude-fable-5-mythos-launch — Coverage of the export control dispute and Amodei's commercial-cost acknowledgment continues to spread; the single new item added today [73] is amplification without original claims.
  • ai-beyond-screens — NVIDIA's XR AI entered public beta as a developer framework for multimodal AI agents for AR glasses and extended reality devices [74], extending NVIDIA's platform scope beyond industrial robotics and automotive; Jeff Bezos stated AI will create more jobs than it displaces [75].
  • ai-alignment-methods-revision — A new item extended the thread's filtering-vs-targeted-addition tension [76] without adding new voices or data beyond the three papers already in the synthesis.
  • ai-content-provenance-watermarking — Hive AI's confirmed operational date extended to June 19, 2026, with auto-tagging documented across content from political figures and major media accounts [77][78]; no new frameworks or claims.
  • openai-chatgpt-superapp-pivot — One additional item added [82] without substantive new claims about the superapp pivot or IPO trajectory.
  • ai-agents-software-paradigm — One new item [83] extended the thread's argument that AI agents are restructuring software from fixed-intent code to on-demand planning systems without introducing new perspectives or voices.

Notable items (5)