Good morning. SpaceX filed for one of the biggest IPOs ever, and Grok’s “spicy” mode somehow made the risk section. Musk has set aside $530 million for potential legal losses, partly tied to complaints about explicit AI images. The filing also warns Grok NSFW modes could create regulatory headaches and reputational damage. Wall Street wanted the future of space travel and got a chatbot with a legal tab.
Also, rumor has it OpenAI could file as soon as Friday.
Your company’s code may no longer be safe on GitHub. The mess started when a GitHub employee accidentally installed a poisoned VS Code extension. That gave attackers access to GitHub’s internal systems, and they walked off with about 3,800 private repositories. GitHub says customer code wasn’t touched, but losing its own repos is still a big problem. That code reveals exactly how GitHub works, which systems talk to each other, and where attackers should strike next. TeamPCP, the crew behind the breach, is the same group tied to this week’s npm worm hitting OpenAI, Mistral, UiPath, SAP and OpenSearch. Now GitHub’s source code is listed for $50k on a cybercrime forum, which sounds cheap for a map to where everyone parks their trade secrets. (Hacker News)
Just to be clear, Microsoft Github was compromised when a Microsoft developer using Microsoft VSCode installed a rogue extension from Microsoft’s extension library, which is moderated and hosted by Microsoft, and now your company’s code is at risk. Any questions? 🙃
Anthropic expects its first profitable quarter. Last summer, Dario told investors Anthropic didn’t expect to turn a profit until at least 2028. Now it expects $559 million in operating profit this quarter. Anthropic is the first major frontier AI lab to make the math work. Revenue is expected to more than double to $10.9 billion as companies pour money into Claude’s coding tools. It’s pulling this off while paying Elon’s xAI about $1.25 billion a month for compute. The AI boom was supposed to be a cash bonfire, and Anthropic found a way to roast marshmallows over it. (WJS)
Fast APIs can still waste your users’ time. A quick wrong answer or retry loop won’t show up in a pretty benchmark table. You.com explains why raw latency misleads teams and how to measure what matters in production.
Stop optimizing for the demo. Download the guide.
👀 closer look
Startups are trading equity for AI tokens. Sam Altman told YC’s current batch that OpenAI will give each company $2 million in API credits in exchange for equity. The deal covers about 400 startups, with OpenAI offering $800 million in compute for roughly 2% of each company. Those tokens are fuel for startups trying to build AI products before revenue shows up. OpenAI gets a piece of hundreds of tokenmaxxing startups while becoming the platform they build on. Some investors are warning founders to be careful as OpenAI could see what works, then copy the best ideas into its own products. Nice little startup you’ve got there. Shame if your API provider became your investor. (TechCrunch)
Frontier models don’t seem to be getting smarter. A new study tested 25 top AI models across 510 questions in business, health, law, pets, and tech. Licensed professionals graded the answers, and no model scored above 73%. GPT-5.5 barely beat GPT-5.1, Claude Opus 4.7 barely beat 4.6, and Gemini 3 Pro stayed under 70%. More reasoning power improved results by only 1% to 2.6%, and sometimes made answers worse. The company behind the study sells expert-in-the-loop AI, so this is not exactly a neutral referee. Still, the results are awkward for labs selling every new model like a giant leap. All that compute and the smartest kid in class still came home with a C+. (Pearl)
AI notetakers are the new workplace liability. Nobody wants to manually type meeting notes anymore when AI can record the call, summarize the important parts, and send everyone the next steps. The tools are too useful to ignore, so they get invited to everything, including confidential meetings. That creates a problem when the recap does not stay in the room. Some companies have already lost legal privilege after AI notetakers sent confidential notes to the wrong people. It gets worse when a notetaker stays on a call after the person who invited it leaves. The office finally solved meeting notes and accidentally gave every private call a share button. (Fortune)
fun stats
🔎 $250 million. Series C raised by a16z backed AI search startup Exa at a $2.2 billion valuation. It’s building “the search engine for AIs,” as search shifts from people typing queries to agents hunting for them.
🌊 77,621. Scientific papers now submitted to arXiv each quarter, up from 44,000 when ChatGPT launched in 2022. The AI paper flood is overwhelming reviewers and raising serious integrity alarms.
📉 68%. Projected drop in humanoid robot prices by 2030, according to independent market analysis. The average selling prices are expected to fall from about $115k to $37k.
🦾 6 months. Current payback period for a humanoid robot in industrial conditions.