When OpenAI's Nonprofit Promise Was Called "A Lie" In Court

Yesterday in AI | Wednesday, May 6, 2026

When OpenAI's Nonprofit Promise Was Called "A Lie" In Court

A 2017 journal entry just became evidence in one of the biggest trials in AI history. GPT-5.5 is the new default and the announcement came with a business model move most people glossed right over. An Anthropic co-founder ran the numbers on when AI starts building itself, and the timeline is closer than you'd expect. The company that told the government no paid a real price for it. And AI compute may soon be drifting through the Pacific Ocean without engines.

Remember to subscribe, rate, and share this podcast if you like it!

Transcript

SPEAKER_00 0:05

Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of artificial intelligence in 10 minutes or less. I'm Mike Robinson. It's Wednesday, May 6th, and the Musk vs. OpenAI trial got genuinely uncomfortable in court. GPT 5.5 just became ChatGPT's new default model, and an anthropic co-founder put 60% odds on AI training its own successors before 2029, with data to back it up. Let's get into it. Let's start in the courtroom, because Tuesday's Musk vs. OpenAI trial testimony produced the kind of material that would be hard to write as fiction. OpenAI president Greg Brockman took the stand. That alone would have been a headline, but Musk's lawyers went further. They read Brockman's personal journal aloud to the jury. Two entries stood out, the first from 2017. Quote, financially that will take me to$1 billion, question mark, end quote. The second also from 2017, where Brockman described OpenAI's public commitment to its nonprofit mission as, quote, a lie, end quote. Brockman's defense from the stand was that these were hundreds of pages of stream of conscious self-doubt, private mental noise that nobody's supposed to present as a master plan. That's a reasonable defense. But it's the kind of thing that sounds different when read aloud in a courtroom. Then, journalist John Gruber did some digging and found that Y Combinator quietly owns roughly 0.6% of OpenAI, a stake worth around$5 billion at OpenAI's current$852 billion valuation. YC co-founders Paul Graham and Jessica Livingston potentially have billions riding on Sam Altman keeping his job. Gruber noted that Graham had spent recent weeks publicly defending Altman while conspicuously stopping short of ever calling him trustworthy or a man of integrity. And there's one more detail. Musk had reached out to Brockman about a potential settlement two days before the trial started. Brockman declined. This trial has a long way to run, but the journal entries in the YC revelation will shape how people think about one core question. Was OpenAI's nonprofit mission ever genuinely meant? That question matters because a lot of the regulatory and governance frameworks being built around Frontier AI labs assume a certain level of mission alignment. Tuesday's testimony made that harder to assume. From the courtroom to the product launch, OpenAI has started rolling out GPT-5.5 instant as the primary model for most ChatGPT users. OpenAI is claiming over 50% fewer hallucinations on high-stakes medical and legal prompts in their own tests. That's specific and measurable, and if it holds in real-world use, it's a meaningful improvement for anyone using ChatGPT for anything health or legal adjacent. There's also a new memory panel that shows users exactly which past conversations shaped a given response. This addresses one of the more persistent frustrations with AI assistance. You can't tell why it said what it said or what it's working from. Now you can see it. And then, buried at the end of the announcement, OpenAI launched a self-serve advertising platform, currently in beta. That's the part we're slowing down on. OpenAI has been running on subscriptions, API revenue, and a spectacular amount of venture funding. Adding advertising signals they're building toward a consumer-at-scale business model, the kind of revenue diversification you pursue when you're preparing for an IPO or trying to hit specific benchmarks before one. The model improvements are real, but the ad platform launch tells you something about where the company thinks it needs to be financially and how fast. While OpenAI is updating its model and its business strategy, Anthropic is pushing Claude directly into finance operations. Yesterday, Anthropic released 10 ready-to-run agent templates covering financial workflows, pitchbook drafting, month-end close, financial analysis. They ship with connectors designed to plug into financial data feeds, so a bank or hedge fund can hook them up to sources like Moody's, SP, or Pitchbook rather than building that plumbing from scratch. This is a different story from the$1.5 billion Wall Street joint venture we covered yesterday. The joint venture is a go-to market strategy. These templates are the actual product landing inside finance teams. A bank or a hedge fund gets something usable on day one, backed by the data sources their people already work with. Finance is one of the highest-stakes environments you can deploy AI into. Accuracy matters, data freshness matters, audit trails matter. Connecting to Moody's and Pitchbook rather than just offering a general purpose assistance signals that Anthropic understands what finance professionals actually need. When AI lands in finance and actually works, it handles the task. An agent that drafts the pitchbook, runs the month end close, and pulls from live data. That's a different category than a chatbot you can ask questions about balance sheets. The combination of yesterday's joint venture and today's agent release is the full picture. Anthropic is building both the distribution infrastructure and the actual tooling to win an enterprise finance. Now to something that's either the most important AI story of the week or very close to it. Jack Clark, co-founder of Anthropic, published an analysis Tuesday putting 60 plus percent odds on AI systems training their own successors before 2029. He built this case on public benchmark data, not intuition. Meter's data shows that AI's independent work window grew from 30-second tasks in 2022 to 12-hour autonomous runs in 2026. 100-hour sessions are projected by the end of this year. Sweebench, which tests AI on real GitHub coding problems, moved from 2% accuracy with Claude 2 to 93.9% with Anthropics Mythos preview model. That shift happened in under three years. OpenAI is already targeting an automated AI research intern by September of this year. What Clark is pointing to is a trajectory that, if it continues, reaches a point where AI meaningfully accelerates its own development. The phrase self-improving AI has lived in theoretical discussions for decades. Clark's argument is that you can now chart its arrival using data that already exists. To put it directly, if these projections hold, the AI industry in 2029 looks very different from today. The bottleneck in AI development has always been human researchers. If that bottleneck shifts, development speeds up in ways that are genuinely hard to model and advance. The data supports the trajectory. What we don't know yet is the exact rate and whether the governance infrastructure humans are building right now is anywhere near adequate for that future. That gap between development speed and governance readiness is the thing to watch. Here's a story about where AI compute might actually live a few years from now. A startup called Panthalassa just raised$140 million led by Peter Thiel to build AI data centers on the ocean. Each unit is an 85 meter steel platform that bobs in open water, converts wave motion into electricity for onboard AI chips, and cools naturally using seawater. They navigate using hull shape alone, no engines. Connectivity runs through Starlink. The first Pacific deployment is targeted for 2027. Why does this exist? Because land-based data centers are hitting a wall of public resistance. Communities are pushing back on the energy use, the water consumption, the noise, and the sheer scale of what gets built. Grid access is also intensifying as a constraint, with data centers competing for power that everyone else needs too. The ocean sidesteps a lot of those problems. No neighbors, renewable energy from wave motion, natural seawater cooling, no grid dependencies. Thiel has argued that extraterrestrial or off-earth solutions to compute are no longer science fiction, and ocean-based platforms are the near-term version of that compared to true space data centers. Obvious questions about maintenance and reliability in open water still need answers, but$140 million raise and a near$1 billion valuation tell you that investors think those problems are cheaper to solve than the zoning battles on land. Our closure is a governance story and deserves your full attention. Several big labs have agreed to give the U.S. government early access to some frontier models for national security evaluation, including cases where safety settings are relaxed, though the details vary by company and agency. Anthropic has taken a harder line in negotiations, publicly refusing government requests for exceptions that would allow mass domestic surveillance and fully autonomous weapons. After those talks broke down, the Pentagon followed through on a threat to designate Anthropic a supply chain risk, which the company is now contesting. A few things to keep in mind. These government agreements are distinct from the White House's proposed pre-release review process we covered yesterday. That proposal is still being discussed. These agreements are already signed. The government already has early access to models from three major labs under terms that can include relaxed safety settings. The AI safety community has spent years arguing that guardrails on frontier models are non-negotiable. The labs have largely agreed publicly. What Tuesday revealed is the practical test of that commitment. When a major government customer says it needs the model without some of these guardrails, what do you do? Three of the four largest labs just said yes. Anthropic said no. Anthropic's position looks principled. It also costs them a Pentagon contract and a supply chain risk label that carries real business consequences. These decisions aren't abstract. They have dollar signs attached, and they're going to shape which companies end up at the center of how AI gets used in defense and national security for a long time. One more thing. Thanks. That's all for this edition of Yesterday in AI. Stay curious, and I'll see you tomorrow.