One of the most famous researchers in AI just quietly switched sides, and the implications for the frontier are hard to overstate.

Yesterday in AI | Thursday, May 21, 2026

One of the most famous researchers in AI just quietly switched sides, and the implications for the frontier are hard to overstate.

Google finally entered the agent race with something that could actually win, OpenAI's IPO is reportedly days from filing, and there's a new kind of enterprise AI customer emerging — one where a bad output doesn't just cost money. Andrej Karpathy's move is the headline, but there's a second announcement buried inside it that tells you exactly where the real AI competition is happening right now. Plus: the one earnings call this week that's basically a financial report card on the entire AI industry.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, Bluesky, or Substack. If you like the show, please take a minute to rate and review it so others can find it!

Links:
Website: https://www.yesterdayinai.news/
LinkedIn: https://www.linkedin.com/in/mrmichaelrobinson/
X: https://x.com/Yesterday_In_AI
Bluesky: https://bsky.app/profile/yesterdayinai.bsky.social
Substack: https://yesterdayinaipodcast.substack.com/

Transcript

Copy Transcript Copied

SPEAKER_00 0:05

Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Thursday, May 21st, and the biggest story of the day is one of the most consequential career moves in the history of this industry. Let's get into it. Andre Carpathi is joining Anthropic. If that name doesn't immediately ring a bell, here's the short version. Carpathi co-founded OpenAI in 2015, went to Tesla to run autopilot, came back to OpenAI for a stint, left in 2024, and then started an AI education startup. His YouTube series, Neural Networks Zero to Hero, has probably produced more self-taught machine learning engineers than any formal degree program. The man is genuinely beloved in this field. On Tuesday, he posted on X that he's joining Anthropic's pre-training team under Nick Joseph. His post got 17 million views in about a day, and alongside that pre-training role, he's building a new unit inside Anthropic focused on using Claude to automate and accelerate Anthropic's own training research. The model is going to help build the next version of itself. We've been hearing about this self-improvement loop from every major lab. Anthropic just hired the person who'd probably know best how to make it work. Pre-training is the most expensive, most consequential phase of building a top-tier AI model. It's where the model learns from raw data at scale before it's refined into the product you actually use. Marginal improvements there are what separate the top labs from everyone else. It's where most of the real research war is being fought, away from the benchmarks and the press releases. Carpathy's post was classic chem, quote, the next few years of the frontier of LLMs will be especially formative, end quote. LLMs or large language models are the class of AI that powers ChatGPT, Claude, and most of what you'd recognize as modern AI. No hype in that post. Just someone who knows exactly where the action is and decided to go there. For anthropic, this is a big deal. They've been on a run, overtaking OpenAI and US business adoption for the first time and closing a $40 to $50 billion funding round. Adding Carpathy to the research team signals they're as serious about the frontier as they are about deployment. Moving over to Google I.O., I covered the big picture yesterday, but one announcement deserves more attention than it got. Gemini Spark. Spark is Google's 24-7 cloud-based AI agent. It runs in the background even when your phone is locked, handling tasks across Gmail, docs, calendar, and third-party apps via MCP connectors, standardized links that let AI plug into outside services. It rolls out to a private test group now, then to Google AI Ultra subscribers in the US and beta next week. It runs on Gemini 3.5 Flash, their new model, which Google says is four times faster than other Frontier models at less than half the cost. The benchmarks are solid, 76.2% on Terminal Bench 2.1 and 83.6% on MCP Atlas, both of which test how well an AI can complete complex real-world tasks. That's enough raw capability to handle the kind of multi-step background tasks Spark is built for. Google also shipped a daily brief feature alongside Google Spark, an opt-in morning summary that pulls from your inbox and calendar, ranks what actually needs attention, and filters the noise. It's a small feature on its own, but it's the same design philosophy as Spark. Gemini staying active on your behalf instead of waiting to be asked. Google has 900 million monthly Gemini users across 230 countries. Reach has never been their problem. Getting those users to do more than ask quick questions has been the gap. Spark is how they try to close it. This puts Google directly in competition with background AI agent tools from Anthropic and OpenAI, and unlike those tools, Spark is launching into an ecosystem where hundreds of millions of people are already in Gmail and Google Docs every day. That's a real head start. Also, out of I.O., Google's AI glasses are now fully specified. Two products. First, everyday audio frames in partnership with Warby Parker and Gentle Monster, two very different aesthetics. They connect to your phone via Gemini and handle navigation, notifications, real-time translation, and photos. You activate them with Hey Google or Tap on the Frame. Samsung built the hardware Qualcomm the Chips, shipping this fall. Second, Project Aura, a display model that floats Google Maps and multi-window app views in your field of view, powered by a separate compute puck. Meta's Ray Bands have owned this category for a couple of years and they're genuinely useful as hands-free cameras with voice access. But they don't have displays. Google's everyday frames do, and Project Aura goes much further. There's also the Android ecosystem angle Meta Simply Can't Match. Real-time translation in 70 plus languages, deep Google Maps integration, Gmail and Calendar access, Gemini Voice. If you're already living in Google's apps, and most people are, these glasses are an extension of the phone already in your pocket. Privacy is a conversation we're going to keep having about cameras on faces in public spaces. But as a product, this is the most complete AI glasses package anyone's announced. Over to OpenAI, where CNBC reported the company is set to confidentially file its IPO prospectus as early as Friday. An IPO prospectus is the formal document a company files before listing on a stock exchange. Filing it confidentially means the financial details stay private for now, but the process has started. Goldman Sachs and Morgan Stanley are handling the deal. Private valuation is over $850 billion. OpenAI has raised more than $180 billion from investors and is reportedly spending at a rate that has its CFO raising flags internally. The timing is notable. The Musk trial wrapped this week. He lost on statute of limitations and plans to appeal. OpenAI exits the courtroom with Mira Marathi's testimony, and Greg Brockman's personal diary entry is now part of the public record. Not ideal pre-IPO reading material, but apparently not enough to slow anything down. If OpenAI prices near $850 billion, it would be the largest tech IPO in history. Anthropic crossed $1 trillion in valuation on private secondary markets, where investors trade shares in companies before they go public last month and is reportedly weeks away from closing a $40 to $50 billion funding round. OpenAI is 10 years old, Anthropic is $5 billion. Both are about to collide with public markets at valuations that would have sounded like fiction in 2022. The confidential filing is just step one. But Goldman Sachs and Morgan Stanley in the room means this is no longer maybe later this year, it's execution mode. Also new from OpenAI today, an enterprise program called Guaranteed Capacity. Companies can commit to one to three year compute agreements at volume discounts. Altman posted on X, quote, customers are increasingly asking us for certainty on capacity, end quote. He added that OpenAI intends to build, quote, as much compute as fast as we can, end quote. The target customer is the kind of organization that won't embed AI in essential high-stakes workflows without uptime guarantees. Banks, healthcare systems, these businesses have been watching from the sidelines precisely because the uncertainty made integration too risky. OpenAI also announced multipath reliable connection, a networking protocol for making GPU clusters, the banks' specialized chips that run AI workloads, faster and more reliable. Taken together, OpenAI is clearly building the necessary plumbing. Same arc Amazon ran with AWS. Start with capabilities, end up owning the infrastructure stack. One more enterprise-related story. Itachi announced a partnership with Anthropic to deploy Claude across all 290,000 of its employees worldwide. They're setting up a Frontier AI deployment center, started with 100 experts and scaling to 300 across North America, Europe, and Asia. Target sectors, energy, transportation, manufacturing, and finance. This is the latest in a string of whole organization commitments for Anthropic. KPMG, 276,000 people, committed the entire firm days ago. The Gates Foundation signed a $200 million four-year deal. Now Hitachi. The pattern is consistent. Anthropic launches a focused industry product. The biggest name in that vertical commits the whole workforce. Hitachi adds industrial infrastructure to that list, meaning AI deployments where failures have physical consequences. That's a different customer profile than a startup accessing Claude via the API. One last thing before we wrap. Nvidia reports its first quarter earnings for fiscal year 2027 after market close today. Analysts expect revenue to nearly double year over year. The stock would move 10% either way by tomorrow morning. Nvidia is basically the financial heartbeat of everything we cover on this show. They make the chips that power AI training and inference across every major lab. Every model, every agent, every data center deployment runs on their hardware. When CEO Jensen Huang speaks tonight, he'll be telling us whether the demand driving all this investment is real and sustainable. It's worth watching. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterdayinai.news. Or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast, please be sure to rate and review it so others can find it. Thanks. That's all for this clone free edition of Yesterday in AI. Stay curious, and I'll see you tomorrow.