AI Mornings with Andreas Vig
Your daily AI news briefing in under 10 minutes. New models, product launches, research breakthroughs, and industry shifts, explained clearly, no hype.
AI Mornings with Andreas Vig
OpenAI's Math Breakthrough & Anthropic's First Profit
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Hey, welcome to AI Mornings with Andreas Vig. It's the 21st of May 2026, and today we have some genuinely historic news. An open AI model has just done something no AI has ever done before. It autonomously solved a famous open problem in mathematics. The ADIS unit distance problem has been around since 1946, and it asks a deceptively simple question. If you place n points on a plane, how many pairs can be exactly one unit apart? For 80 years, mathematicians believed that square grid constructions were essentially optimal. Turns out they weren't. An internal OpenAI model found an infinite family of examples that beat the conjecture, and here's what makes this remarkable. The model wasn't specifically trained for math or targeted at this problem. It was a general-purpose reasoning model that, when tested on a collection of Adder's problems, produced a valid proof. Fields medalist Tim Gowers reviewed it and said he would recommend acceptance to the annals of mathematics without hesitation. The proof also brought in unexpected ideas from algebraic number theory to solve what looked like an elementary geometry question. This is the first time AI has autonomously resolved a central open problem in an active mathematical field. It's not just solving equations, it's having what Gowers called original ingenious ideas. In business news, Anthropic is about to hit a milestone that's been elusive for AI Labs, its first profitable quarter. The company told investors it expects $10.9 billion in second quarter revenue, more than double the previous quarter, with an operating profit for the first time. That puts them in a strong position relative to OpenAI, though Anthropic warned they may not stay profitable throughout the year due to major compute costs coming down the pipe. Speaking of compute costs, here's where those billions are going. Anthropic just signed a staggering deal with XAI, $1.25 billion per month through May 2029 for 300 megawatts of compute capacity. That's the entire output of XAI's Colossus 1 data center in Memphis. The total deal could exceed $40 billion. The details emerged from SpaceX's IPO filing and they reveal an interesting dynamic. XAI has overbuilt its infrastructure while Grok usage has declined, so it's now selling spare capacity to one of its closest competitors. It's a neo-cloud model that lets AI companies offset infrastructure costs by acting as cloud providers when their own usage falls short. Those same SpaceX filings gave us our first real look at XAI's financials, and the numbers are stark. XAI lost $6.4 billion from operations on just $3.2 billion in revenue last year. That's a widening gap between what they spend and what they earn. For context, Anthropic is approaching profitability while XAI is burning billions. Grok has 117 million monthly active users, but that's only about one-fifth of the combined X and Grok ecosystem. The filing also reveals plans to scale Grok to multiple trillions of parameters and to begin deploying orbital AI compute satellites as early as 2028. Elon Musk has been pitching space-based data centers as cheaper than terrestrial ones. We'll see if that materializes. On the model front, Alibaba released Quen 3 7 Max, their latest model built specifically for what they're calling the Agent Era. It's designed for sustained autonomous execution across long-horizon tasks, and they demonstrated this with a 3-5-hour kernel optimization run where the model made over a thousand tool calls and achieved a 10x speed up over the reference implementation. What's interesting is that the model generalizes across different agent frameworks, performing consistently whether deployed through Clawed Code, OpenClaw, Quencode, or other scaffolds. On benchmarks, it scored 69.7 on Terminal Bench 2, outperforming Deep Seek V4 Pro Max, and hit top marks on GPQA Diamond at 92.4 and HMMT 2026 at 97.1. For Creative AI, Stability AI released Stable Audio 3, a family of models that can generate professional grade music over 6 minutes long. That's more than double the length of their previous generation. They're releasing four models ranging from 459 million to 2.7 billion parameters, with the smaller and medium models available as open weights. Importantly, these models are built on fully licensed data from partnerships with Warner Music Group and Universal Music Group, which could matter a lot given the ongoing legal battles facing Suno and Udio. Stability also hired Ethan Kaplan, formerly of Universal Audio and Fender, to lead their professional music offerings. Finally, some numbers from Nvidia. The company posted another record quarter with $81.6 billion in revenue, up 20% from the previous quarter, with $75.2 billion from data center alone. They're forecasting $91 billion for next quarter and authorizing $80 billion in share buybacks. But here's a number that caught my attention. Nvidia's holdings in privately held startups nearly doubled in one quarter, from $22 billion to $43 billion. They spent $18.5 billion buying stakes in private companies, and that doesn't even include their pending $30 billion investment in OpenAI. Jensen Huang also highlighted their Vera CPU, which he says opens up a brand new $200 billion market for Nvidia. The pitch is that agents run on CPUs, not GPUs, and Vera is purpose built for token processing. Nvidia has already sold $20 billion worth of standalone Vera CPUs this year. That's the news for today. I'll see you tomorrow.