6/22: Google Stumbles
Plus: Google x A24 partnership, Anthropic x Micron partnership, quantum executive order
Welcome back, monitors. I hope you enjoyed the three-day weekend, because we’re back in the Situation Room all week. Be sure to catch us on X and YouTube, and join our Discord to chat with our hosts live.
Today’s Experts
Christian Keil (Andreessen Horowitz)
Olivia H. Scharfman (Institute For Progress)
Miles Matthias (Stripe)
Elena Burger (Andreessen Horowitz)
Carmen Li (Compute Exchange, Silicon Data)
John D’Agostino (Coinbase)
Peter Wildeford (AI Policy Network)
Samo Burja (Bismarck Analysis)
Paul Bakaus (Impeccable AI)
Joshua Levine (Foundation for American Innovation)
Steven Glinert (Sphere Semi)
Making Sense of the World
Google Stumbles
Last week, Noam Shazeer — co-inventor of the Transformer, mixture of experts, and LaMDA; founder of Character AI; and co-lead of Gemini — left Google for OpenAI. On Friday, John Jumper — one of the creators of AlphaFold, for which he won a Nobel Prize — left Google for Anthropic. Both of these are heavy blows to the company. As one frontier lab researcher told me this weekend, there are maybe 50 people in the world who truly understand how to train a frontier model end-to-end. If AI is the most important thing in the world, it stands to reason that these people are among the most economically valuable workers on earth, and it’s a top priority for companies to attract and retain them1. The markets know this, and Google stock fell 5% today, wiping out a couple hundred billion dollars of market value.
More broadly, Google has been struggling for some time. Gemini 3.5 Flash, released a month ago, is roughly on the level of GPT-5.2, released in December, and substantially behind Fable, Opus 4.8, and GPT-5.5. Gemini 3.5 Pro is nowhere to be found. If you zoom out even further, Google’s fumble becomes even more apparent. In the mid-2010s, Google had two frontier labs (DeepMind and Google Brain), tremendous compute/data/cash flow/talent advantages, and a near-monopoly on the frontier tech. Google was so dominant that Elon Musk was personally inspired to co-found OpenAI because he feared Google would dominate the entire future if he didn’t found a competitor. The Transformer was invented not inside of OpenAI, but Google — but OpenAI, not Google, was the first to successfully commercialize it, and Google struggled to bring their own LLM to market for a while. It’s entirely possible that the phenomenon of Google as a frontier AI lab was only really true from ~2010 to 2018 and through most of 2025.
Then again, leads in AI change constantly. The “Mandate of Heaven” meme on X rarely reflects reality, and labs that seem totally invincible, like OpenAI in 2023, can lose their lead just as easily. Maybe Google will figure out their issues, leverage their compute and data advantages, and catch up to OpenAI and Anthropic again. Maybe they’ll shift their niche from building frontier models themselves to supplying TPUs and Google Cloud capacity for others to do so. Or maybe, like Meta, their small errors will compound, and they’ll find themselves far behind companies that once had a fraction of their resources. Whatever happens, we’ll be monitoring.
More Stories
President Trump signs two executive orders on quantum computing. One establishes a national effort to build the first quantum computer capable of accelerating scientific discovery and commercial applications, with the explicit goal of building a commercially relevant quantum computer by 2028. The other directs the government to implement post-quantum cryptography by 2031, in order to prevent against quantum-enabled cyberattacks. Reading these executive orders, it struck me that they seem far more comprehensive than executive orders about AI, an issue that is both more urgent and more important. It’s good to see that state capacity can still materialize where needed.
Google announces a partnership with A24. A24 is perhaps the most successful and innovative new film and TV studio of the 21st century. Their catalogue includes Backrooms (2026), Marty Supreme (2025), Everything Everywhere All At Once (2022), Uncut Gems (2019), Lady Bird (2017), Moonlight (2016), and many more. Google will invest $75 million in the studio and send forward-deployed Google DeepMind engineers to build new tools and workflows for filmmaking. With the cost of generating surprisingly decent-looking film crashing to zero while big movie budgets continue to climb into the hundreds of millions, disruption is inevitable.
Micron and Anthropic announce a partnership. Micron will design memory and storage products with Claude in mind, adopt Claude internally, and invest in exchange for an equity stake. Anthropic has made strategic compute partnerships before2, but this one is likely one of its largest yet. Micron will now be Anthropic’s main memory partner, as Samsung and SK hynix are for OpenAI. Memory supply is critical, as memory is currently the main bottleneck to further compute scaling.
Alan Greenspan dies at 100. He served as Chairman of the Federal Reserve for eighteen years, under four presidents. He was born in a very different world, to a Jewish family in upper Manhattan in 1926, an eternity ago. As a child, he played music, and even studied clarinet at Juilliard before entering economics. He studied at NYU and Columbia, got a job on Wall Street as an analyst, and worked his way up to being the chairman of a major economic consulting firm. As Fed Chairman from 1987 to 2006, he embodied laissez-faire economics: optimistic about free markets and economic growth, and skeptical of government control and regulation. He was close personal friends with libertarian writer-philosopher Ayn Rand. Elena Burger wrote a touching personal tribute to Greenspan that she read live on MTS.
Banger Review
At least, until the full automation of AI research, which, depending on who you trust, could happen in as little as two years.
While researching this, I stumbled across Anthropic co-founder and chief compute officer Tom Brown, basically unknown outside of the company. He has a very interesting background: he studied computer science and cognitive science at MIT, worked as a software engineer, and founded YC-backed friend-finding startup Grouper.
In 2015, he attended a workshop from the Center for Applied Rationality, a spinoff organization of Eliezer Yudkowsky’s MIRI and LessWrong, dedicated to teaching people the art and science of rationality and cognitive biases. Rationalists are known for AGI-pilling people very very early, and while we can’t conclude that Brown decided to change his career from going to CFAR, what we do know is that he left Grouper and joined OpenAI in early 2016 as one of the first 20 employees, at a time when AGI was a niche curiosity of various sci-fi cranks. Every time you think you’ve uncovered the last person connected to the rationality movement, there’s always another.
At OpenAI, he co-invented RLHF in 2017, and after a year at Google Brain from 2017 to 2018, he returned to OpenAI to discover neural network scaling laws and lead the engineering team behind GPT-3. In 2021, he joined Dario Amodei to co-found Anthropic and the rest is history.