AI News Roundup — June 20, 2026
Google lost two of its most consequential AI researchers in three days: Noam Shazeer — co-author of the Transformer paper and Gemini co-lead — left for OpenAI after Google spent $2.7 billion to bring him back just 22 months ago, and Nobel laureate John Jumper — AlphaFold lead and 2024 Nobel Prize winner — left for Anthropic. The simultaneous departures represent the most damaging talent losses in Google's AI history and arrive with Gemini 3.5 Pro still unshipped.
Noam Shazeer Leaves Google for OpenAI — The Transformer Co-Author's $2.7B Return Lasted 22 Months
Noam Shazeer, VP of Engineering at Google and technical co-lead of its Gemini AI models, announced on June 18 that he is leaving Google to join OpenAI as Lead for Architecture Research. The move is the biggest AI talent defection in years. Shazeer co-authored the 2017 "Attention Is All You Need" paper that introduced the Transformer architecture underpinning every major AI system today — GPT, Gemini, Claude, Grok, Llama, all of them. He also co-authored the Sparsely-Gated Mixture of Experts paper (2016), which introduced the MoE architecture used in GPT-4, Gemini 1.5, and nearly all current frontier models, and designed Multi-Query Attention, the inference efficiency technique that made fast AI at scale economically viable. Google paid approximately $2.7 billion in August 2024 to bring him back from Character.AI (which he co-founded after leaving Google in 2021). He lasted less than 22 months — roughly $122 million per month by Google's investment. Sam Altman responded publicly within hours: "noam is one of the people I have most wanted to work with since the very beginning of openai. only took 10 years. i think it will be worth the wait." OpenAI internal reaction: described as "the most technically consequential hiring decision OpenAI has made since Jakub Pachocki became chief scientist." Google internal reaction: described as "shell-shocked."
CNBC · Build Fast With AI · TechTimes
Nobel Laureate John Jumper Leaves Google DeepMind for Anthropic — Two Giant Exits in Three Days
One day after Noam Shazeer announced his departure to OpenAI, John Jumper — the 2024 Nobel Prize laureate in Chemistry and creator of AlphaFold — announced he is leaving Google DeepMind for Anthropic after nearly nine years at the company. Jumper and DeepMind CEO Demis Hassabis shared the Nobel Prize for AlphaFold, the AI model that can predict the 3D structure of proteins from genetic sequences — a breakthrough that transformed drug discovery and structural biology. Bloomberg reported that Jumper was also a key member of Google's team building coding tools, which the company has struggled to commercialize. In his farewell post on X, Jumper credited Hassabis for "taking a real chance letting me lead the AlphaFold team just six months after finishing my PhD." The two departures — Shazeer to OpenAI, Jumper to Anthropic — represent the most consequential talent losses Google DeepMind has experienced simultaneously and arrive with Gemini 3.5 Pro still unshipped and Gemini Nova still without a public timeline.
TechCrunch · The Verge · CNBC
What the Google Brain Drain Means for Gemini 3.5 Pro and the OpenAI-Anthropic Competitive Gap
The Shazeer and Jumper departures create direct competitive intelligence risks for Google. Both researchers understand where Gemini's architecture is strong and where it faces fundamental limitations — the tradeoff space of decisions made during Gemini 3.5 Pro's development moves with them to rival labs. Shazeer's sparse MoE expertise and attention mechanism knowledge is now inside OpenAI and will inform GPT-6 architecture decisions. Gemini 3.5 Pro, still promised before June 30 per Sundar Pichai's I/O commitment, is expected to launch as planned — the work is done — but the next generation (Gemini Nova) loses one of its most senior architectural contributors. The pattern is also significant for talent dynamics: OpenAI is recruiting from Google while simultaneously preparing a high-profile IPO, and Anthropic is attracting a Nobel laureate while navigating a government model shutdown. Both rivals are pulling from Google's deepest technical bench at the same moment.
Build Fast With AI · Windows News
42 State Attorneys General Subpoena OpenAI — Sycophancy, Health Data, and Advertising Practices
A coalition of 42 state attorneys general served OpenAI with a broad subpoena demanding internal documents covering the company's handling of user data, safety practices, advertising representations, and — notably — model sycophancy. The sycophancy investigation is the most unusual element: AGs are asking for documents about whether OpenAI's models are designed or trained to tell users what they want to hear rather than what is accurate, and whether that tendency is deliberate or known. The health data line is particularly sensitive with OpenAI's recent product push into healthcare and its public claims about medical AI capabilities. OpenAI said it is "engaging constructively" with the investigation. The subpoena lands the same week OpenAI confidentially filed its S-1 for an IPO — creating simultaneous regulatory scrutiny and public market preparation that investors will need to weigh together.
NY Times · Build Fast With AI · MLQ.ai
Fable 5 Day 8 — Trump Says 'Going Fine,' Refund Deadline Passes, No Resolution Yet
The Fable 5 and Mythos 5 export control ban entered its eighth day with no resolution announced. President Trump told reporters at the G7 sidelines that negotiations are "going fine" — the most positive executive branch signal yet. Today, June 20, was the last day for eligible Anthropic customers to request refunds for subscriptions purchased between June 9–12 (the days Fable 5 was live before the shutdown). Anthropic's Managing Director of International remains "very confident" of restoration within days, suggesting a July framework involving identity verification and proactive jailbreak monitoring and reporting to the government. The export control dispute is increasingly being framed — per TechCrunch's history piece "From PGP to Mythos" — as a repeat of past enforcement failures: the government's export control on PGP encryption in the 1990s and spyware restrictions in the 2000s both failed to prevent the restricted capability from proliferating internationally while burdening legitimate domestic and allied users.
Build Fast With AI · TechCrunch
SpaceX MSCI Inclusion Drives Retirement Fund Anxiety — Americans Worry 401(k)s Are Being Forced Into SPCX
With MSCI inclusion of SPCX now active and the Nasdaq-100 inclusion window approaching (around July 7), a Guardian/Hacker News report surfaced significant public anxiety among Americans who realize their 401(k) and pension funds are being passively forced to buy SpaceX stock through index fund allocations — with no ability to opt out. The concern: SPCX is priced at $1.77 trillion with a 94x EBITDA multiple on a company that posted a $4.94 billion net loss in 2025 and has its two largest AI compute contracts terminable with 90 days' notice after December 31. Morningstar's fair value estimate of $780 billion is 44% of the market price. For most passive fund holders, it is not possible to hold a broad market index that excludes SPCX once it joins the Nasdaq-100. The structure of the modern passive investing infrastructure means retail investors with no interest in SpaceX's risk profile are being added to its shareholder register whether they want to be or not.
History of Cyber Export Controls That Never Worked — From PGP to Mythos
TechCrunch ran a detailed history piece titled "From PGP to Mythos: a brief history of export controls that didn't stop anyone," drawing on three decades of failed export control enforcement on dual-use technology. The PGP case: in 1991, Phil Zimmermann released PGP encryption software and immediately faced a criminal investigation for violating US export controls on strong cryptography. The government eventually dropped the case — in part because PGP source code was published as a book (protected speech), allowing international users to simply retype it. The result: PGP spread globally anyway, while years of enforcement uncertainty chilled legitimate domestic development. The Mythos parallel: the government is trying to prevent cyber-capable AI from reaching foreign actors, but open-weight models with comparable capabilities are downloadable worldwide today. The export control prevents allies from using Fable 5's consumer interface — it does not prevent adversaries from training their own equivalent. The capability exists globally; the restriction burdens only the law-abiding.