6/18: Medjourney
Also: is OpenAI back now?
It’s Thursday, monitors, and Gabriel is back in the Situation Room with us. GTA VI pre-orders will finally be starting next Thursday, so maybe we’ll actually get to play the game this year after all. As a Floridian, I look forward to seeing how accurate the map is. Be sure to catch us on X and YouTube, and join our Discord to chat with our hosts live.
Today’s Experts
Steve Hou (Silicon Data)
Ariel Cohen (Navan)
Tanishq Abraham (Sophont AI)
Jeanne Shih (PhD)
Benjamin Liu (Formation Bio)
Ashwin Sreenivas (Decagon)
Warren Pies (3F Research)
Matt Parlmer (General Fabrication)
Making Sense of the World
Is OpenAI back?
Yesterday, star researcher Noam Shazeer announced that he’s joining OpenAI. Given that OpenAI has been struggling to match Anthropic’s momentum for some time, this is a huge boon to the company. Some people don’t follow AI researchers like NBA players1, so we’ll explain who Shazeer is and why he’s such a big deal.
In 2000, right after grad school, Shazeer joined a small Series A startup called Google. In his early days, he worked on spelling correction in search and a key algorithm behind Google’s advertising. In 2012, he joined Google Brain, the company’s internal AI research organization, at the very beginning of the deep learning revolution in the tech industry. Shazeer’s research at Brain focused on sequence prediction, language modeling, and scaling. In 2016, he co-discovered an early version of what would later become neural network scaling laws, and helped make mixture of experts work with large neural networks.
In 2017, he was one of the lead authors on perhaps the most important paper in AI history: “Attention is all you need”, which introduced the transformer architecture that still powers every frontier model today. Though it wouldn’t release an LLM for some time, Google realized the potential of the transformer for things like Google Translate, so it sent Shazeer to work scaling them. He spent the next four years making multiple key advances that allowed transformers to be run bigger, faster, and cheaper. He was a major contributor to the team that built LaMDA, Google’s first conversational large language model that was so convincing (for the time) that it made one engineer think it was sentient.
In October 2021, Shazeer left Google to found Character.ai, a platform that allowed users to build custom chatbots in the image of various real and fictional characters. After ChatGPT came out and LLMs really started to take off, Character.ai became a breakout success. In 2024, Google (sort of) acquired Character.ai for $2.7 billion, and appointed Shazeer co-lead of Gemini along with Jeff Dean and Oriol Vinyals. Before Shazeer, Gemini was well behind OpenAI and Anthropic. Shazeer helped turn Gemini 2.5 Pro, 3 Pro, and 3.1 Pro into genuine frontier-class models, and made the Flash and Gemma series into highly competitive lower-cost options.
So why is Shazeer joining OpenAI as the new lead for architecture research so important? Because of the labs’ biggest goal for the next couple of years — recursive self-improvement. AI models can already meaningfully assist with AI research, and if trends continue, they may be able to fully automate AI research and trigger an intelligence explosion. If you believe what the labs believe, whoever wins the race to RSI wins everything. Aside from scaling compute and getting better data, the only way to improve models is through improving the algorithms, exactly what Shazeer excels at.
Also, Dean W. Ball, friend of the show and author of Hyperdimensional, is joining OpenAI to lead its new Strategic Futures team, advising the company’s leadership on frontier AI policy. He’s an incredible pick, OpenAI is lucky to have him, and we’ll be monitoring OpenAI’s policy work with great interest.
More Stories
Midjourney announces Midjourney Medical. Yesterday, Midjourney, maker of creative and tasteful image models, finally announced the first of their long-awaited hardware devices. Is it a pair of AI-enhanced earbuds? Augmented reality glasses? A smart speaker? Nope, it was extremely not any of those things. Midjourney is breaking into medical devices with the Midjourney Scanner, a machine that uses hundreds of thousands of tiny ultrasonic sensors to scan your body with no radiation, no magnets, and no contrast dye needed, all much cheaper and faster than a traditional MRI or CT scan. The hope is that you’ll be able to understand your body better, identify changes over time, and potentially catch issues before they become dangerous. They’ll be putting thousands of these machines in Midjourney spas across the country, along with saunas, cold plunges, and hot tubs, with the first location in downtown SF set to open in 2027.
Meta is buying compute capacity from Crusoe. Crusoe is a major partner of OpenAI and helped build Stargate. Meta is still well behind in the AI race despite major talent acquisition.
Amazon may sell Trainium chips, competing with Nvidia. OpenAI, Anthropic, and other labs use Trainium chips for training and inference, but access them through Amazon Web Services. Amazon has never sold the chips before. Google recently began selling its TPUs for the first time.
Europe will invest less in data centers. The current EU proposal includes $1.15 billion in subsidies for domestic data centers, a downsizing from prior proposals and a tiny fraction of the hundreds of billions China intends to spend on state subsidies. Someone needs to make the Eurocrats read Europe 2031 before it’s too late.
Banger Review
The median NBA player makes $4.9 million per year, much less than superstar AI researchers, who can make tens of millions. The highest-paid NBA player this season was Steph Curry, who made $59.6 million, but the highest-paid AI researchers make a lot more than that — researcher Andrew Tulloch was reportedly offered up to $250 million a year by Meta.