The Information Machine

2026-06-19

Google lost two landmark AI researchers to direct rivals within 24 hours — Nobel laureate John Jumper to Anthropic and transformer co-author Noam Shazeer to OpenAI — while Trump publicly stated the US government may take equity stakes in frontier AI companies.

What

Google lost two of its most prominent AI researchers to competitors in rapid succession. Noam Shazeer — co-author of the 2017 'Attention Is All You Need' transformer paper and Gemini co-lead — is joining OpenAI after Google paid $2.7 billion to reacquire him less than two years ago [1]; the following day, John Jumper — who shared the 2024 Nobel Prize in Chemistry with Demis Hassabis for AlphaFold after nearly nine years at Google DeepMind — announced he is joining Anthropic [2], with both departures drawing broad media coverage [3][4]. On the government side, Trump publicly stated the US may take equity stakes in AI companies [5], and tech stocks gained when the statement became public [6]; separately, an unconfirmed single-source report suggests the White House and Anthropic may have identified a reinstatement path for Fable 5 and Mythos based on benchmark failure-rate tests rather than jailbreak elimination [7], while a research finding deepened expert consensus that perfect jailbreak immunity is technically impossible for any LLM [8]. OpenAI's collaboration with Boston Children's Hospital confirmed 18 clinician-validated new diagnoses from 376 previously unsolved pediatric rare-disease cases using an o3-based model [9], and OpenAI simultaneously moved GPT-5.5 Instant — tuned for health evaluations including urgent-care recognition and uncertainty acknowledgment — to the free ChatGPT tier [10].

Why it matters

Google losing Shazeer and Jumper to the two labs it competes most directly with — at the moment both Anthropic and OpenAI are in SEC review for IPOs — concentrates frontier model-building talent outside Google while its rivals gain it simultaneously. Trump's personal statement on government equity stakes moves the question from internal administration debate to a declared public position, materially raising the pressure on AI companies to engage with the proposal.

Open questions

  • With Jumper joining Anthropic [2] and Shazeer departing for OpenAI [1], what does Google's Gemini and AlphaFold successor roadmap look like without them, and does this accelerate further departures from DeepMind?

  • If the reported benchmark-based reinstatement path for Fable 5 and Mythos [7] is confirmed, does it create a workable resolution — or does research establishing that perfect jailbreak immunity is theoretically impossible for any LLM [8] mean the benchmark standard will itself be contested?

  • Trump's stated interest in government equity stakes [5] and Sanders' $7 trillion tax-fund proposal are incompatible mechanisms targeting the same stated goal — is there a legislative path for either, or do they produce competing pressure without synthesis?

  • SemiAnalysis argues NVIDIA's CUDA software ecosystem is the actual competitive moat and that 99% of chip startups fail by underestimating software depth relative to hardware specs [11]; as custom ASIC startups continue making unvalidated simulation-based performance claims, does this thesis get independently confirmed?

Thread movements (13)

  • google-ai-talent-exodus — A new thread opened covering Google's loss of two landmark researchers to direct rivals within 24 hours: Noam Shazeer (transformer co-author, Gemini co-lead) is joining OpenAI [1], and John Jumper (2024 Nobel laureate for AlphaFold, nine years at DeepMind) announced he is joining Anthropic [2], with broad media coverage following both announcements [3][4].
  • us-government-ai-ownership — Trump publicly stated the US may take equity stakes in AI companies [5], upgrading earlier coverage of administration officials discussing the idea to a presidential statement; tech stocks gained when the news became public [6], and additional items amplified the story without adding substantive new angles [45][46][47].
  • fable-mythos-export-control — An unconfirmed single-source report suggests the White House and Anthropic may have identified a reinstatement path based on benchmark failure-rate tests rather than jailbreak elimination [7]; a separate research finding reinforced expert consensus that perfect jailbreak immunity is theoretically impossible for any LLM [8]; and Zvi Mowshowitz's capabilities review confirmed Fable 5 as state-of-the-art while noting false-positive classifier problems on benign queries including the word 'cancer' [48].
  • openai-health-ai-democratization — A new thread formed around two OpenAI health AI moves arriving in parallel: a collaboration with Boston Children's Hospital produced 18 clinician-confirmed new diagnoses from 376 previously unsolved pediatric rare-disease cases using an o3-based model [9], and OpenAI moved GPT-5.5 Instant — explicitly tuned for health evaluations — to the free ChatGPT tier [10], with both framed as bringing frontier health AI to a broader audience [51][52].
  • asic-gpu-market-dynamics — A new thread formed documenting NVIDIA's persistent dominance over custom ASICs; SemiAnalysis published the argument that CUDA software depth is NVIDIA's actual moat and that 99% of chip startups fail by underestimating the software investment required [11], while multiple items added coverage of NVIDIA's CEO projecting continued growth and the gap between startup simulation-based performance claims and independently validated results [53][54][55].
  • meta-ai-workforce-disruption — Meta's $26.8 billion quarterly profit figure was added to the thread [60][61], sharpening the tension between the company's financial performance and CTO Bosworth's admission that morale is near a 20-year low comparable to Cambridge Analytica — the specific figure replacing the prior synthesis's reference to 'record profits.'
  • ai-coding-agents-robot-training — Social amplification of Anthropic's Project Fetch Phase 2 continued [62][63][64]; one item made the speed comparison concrete by specifying that the human-assisted team took 'roughly 4 hours' in 2024 versus Claude's 12 minutes and 7 seconds [65].
  • deepmind-ai-control-roadmap — Victoria Krakovna, a named DeepMind control researcher, publicly characterized TRAIT&R as 'a second line of defense against misalignment risk, if alignment is insufficient' [71], adding institutional named-voice framing to the roadmap; coverage spread internationally including Russian- and French-language posts [72][73].
  • ai-macro-economic-disruption-signals — Four items extended coverage of AI's macro economic footprint [74][75][76][77], including Fed Chair Warsh's argument that AI productivity gains could support lower interest rates and Accenture's approximately 20% stock decline in June 2026 as investors priced in AI-driven disintermediation of large IT consulting firms.
  • openai-chatgpt-superapp-pivot — Secondary coverage of the SpaceX/Cursor acquisition and OpenAI's superapp pivot continued accumulating [78][79][80][81][82] without new substantive claims or angles beyond what prior synthesis captured.
  • us-ai-policy-regulation — One item was added [83] to the thread tracking US AI regulatory actions, without new primary-source claims beyond the confirmed BIS export-control action forcing Anthropic to disable its advanced models for foreign nationals.
  • nadella-token-capital-ai-economics — One item was added [84] extending coverage of investor skepticism about whether hyperscaler AI capex produces returns, against a backdrop of Microsoft's approximately $190 billion calendar-2026 capex commitment.
  • ai-agent-benchmark-reality-gap — One item was added [85] to the thread documenting AI agent failures at real-world tasks, without introducing substantive new perspectives beyond the solve/judge dissociation and visual reasoning gaps already in the synthesis.

Notable items (2)