Another great experience for @encord_team at #CVPR2026! Funny to think we started with a tablecloth and a couple of monitors (which I promptly returned after the conference)... I'd say this year's booth (and coffee robots!) are a bit of an upgrade.

Something’s been in the works. More soon...

Andrej Karpathy is joining Anthropic. Around a year ago, when the LLM race was far murkier, a common framing was that each major lab had a different comparative advantage. OpenAI had the strongest brand, Anthropic had the strongest team, and Google had the strongest distribution. What would the winning moat be? Right now, the market seems to be voting for talent. Anthropic has now surpassed OpenAI in valuation and is moving at a pace that would make it the fastest company in history to reach a trillion-dollar valuation. Even while automating enormous amounts of cognitive work, Anthropic is still doubling down on elite talent density. In the age of AI, people still matter the most. This may end up being one of the highest-stakes management case studies ever written.

A lot of the discussion around the Anthropic/SpaceX deal has focused on Claude rate-limit increases, but I think the partnership reveals something much more interesting about where the frontier lab market is heading. One thing that has struck me as unusual about frontier labs is that operating one successfully as a business requires making highly accurate compute-demand forecasts years in advance. There’s a strange impedance mismatch where AI paradigm shifts and hype cycles happen on the order of weeks, while data-center buildouts happen on the order of years. A more mature market will likely evolve sophisticated compute intermediaries that absorb and manage this forecasting risk so labs can focus on the layer they are actually trying to be good at. This deal feels like an early signal of this evolution. SpaceX acquiring xAI and then leasing its Colossus cluster to a competitor looks less like a lab move and more like the move of a compute broker. I wouldn't be surprised by a world where this becomes further financialized and we start to have a proliferation of compute derivatives and financial products that allow labs to hedge and traders to speculate. The next generation of lucrative high-frequency trading firm strategies may end up being in market-making GPU futures instead of equities. A quant's dream come true.

The robot economy is coming, and data is the bottleneck. It was a pleasure sitting down with @dessaigne on the @ycombinator podcast to discuss this bottleneck and reflect on the progress @ulrikstighansen and I have made since the W21 batch. We started YC before the AI boom and struggled to convince many VCs the AI market would be big enough to support a new data company. Now we are building critical infrastructure in the midst of the largest technological shift in our lifetimes. Nicolas was one of our batch partners and has seen the whole journey of the company. We cover topics ranging from the founding story behind @encord_team, why AI engineers didn't trust AI automation before ChatGPT, the data problem holding back Physical AI, and the next steps for the company after raising our Series C. YC was our first outside backer and we are very grateful for their support through years of starting, building, and scaling Encord. Watch the full conversation below!

Two big releases from the model frontier this week. Anthropic gave a controlled debut to Mythos, a model so capable that it has been highly restricted rather than released publicly. Meanwhile, Meta, after a year of rebuilding its AI team, unveiled Muse Spark. Most AI coverage obsesses over benchmark scores of these models, but I think these launches highlight something more interesting. Anthropic and Meta are proposing opposing visions of what superintelligence looks like. One is built around pure capability, the other is built around personal context. By restricting Mythos, Anthropic is signalling a view of superintelligence as a central oracle that should be guarded, controlled, and deployed selectively for high-stakes work. Meta’s bet is different. It envisions superintelligence as deeply personal and productized for every individual. The value of the model comes from knowing you: what you see through your Ray-Bans, your purchase history, your social graph, and the content you engage with. There are two very different bets on where the value of intelligence lives. Anthropic is superintelligence in the cloud. Meta is superintelligence for you. Ironically, I think both can win. The most capable model and the most useful model are not always the same thing. Context is the moat that benchmarks don't measure.

There is a certain irony in Anthropic, a company at the forefront of commoditising code, now facing a leak of its own code. If this had surfaced 12 hours later, it would have been a perfect April Fools joke. Instead it serves as an accidental demonstration of a larger point that code is no longer a company’s moat. Neither really are the models, which are rapidly depreciating assets. Anthropic’s moat is its data, its compute infrastructure, its investor and customer relationships, its talent, and the context embedded in how it builds and deploys systems. Value is no longer in the artifact itself, but in the system that produces and improves it. Leaks don’t really break that. If anything, they reinforce it.

Following our recent London office opening, it was an honour to sit down with the UK's Minister for AI, @KanishkaNarayan MP, to talk about the importance of data and what that really means for the UK's position in the AI ecosystem. For us, the conversation isn't abstract. Most people don't know that without the UK Biobank dataset, one of the world's largest medical research datasets, Encord likely wouldn't have ever existed. @ulrikstighansen used the dataset in his work at Imperial and gained a number of insights from that work which he would later bring to the founding of @encord_team. Data should be thought of as a national asset for the UK, one which gives Britain an advantage in AI. Glad to see that the people shaping policy are recognising this.

We were very honoured to welcome Britain’s Minister for AI, @KanishkaNarayan MP, to commemorate the opening of our brand new London office! We’re excited about this next chapter as we continue to scale here in London. From the very beginning, @ulrikstighansen and I believed the UK was one of the best places in the world to build and grow a company due to the exceptional talent shaped by world-class universities, strong institutions, and a globally diverse workforce. This opening ceremony was quite a moment as I remember back to our first office, a single white room that could hardly hold a handful of people. Now we have close to 40 nationalities represented in our London office alone. That's nearly one fourth of the world! A huge thank you to Kanishka for joining us, and to every person who has helped make @encord_team into what it is today.

In the wake of @encord_team's Series C, @ulrikstighansen and I appreciated the invite to @10DowningStreet to discuss the current state of AI and how the UK government can help foster the domestic startup ecosystem. We had a great session and are very excited about the progress being made in the UK tech environment. Thank you to James Carroll for hosting us and Dom Hallas for the support!