Happy Friday! Somewhere between “AI can write code” and “robots can do chores,” someone realized those ideas should probably meet. So a guy bought a robot arm and handed it to OpenClaw. The agent helped get the arm connected, then started teaching it how to interact with the world. It was slow, awkward, and probably one bad setting away from cooking its own motors. But that’s exactly why it’s fun.
🇺🇸 We’re taking off for the Memorial Day weekend. See you back here on Tuesday!
What’s the point of 905 million users if most of them pay nothing? OpenAI and Anthropic are both trying to look ready for Wall Street, but they’re walking in with very different business models. OpenAI made $5.7 billion in the first quarter, nearly $1 billion more than Anthropic, yet still lost $1.22 for every dollar it made. ChatGPT has huge reach, but only 55 million of its 905 million weekly users actually pay. Anthropic took the opposite approach, leaning harder into enterprise and developers who bring real budgets and serious workloads. That strategy paid off, giving Anthropic its first profitable quarter way ahead of schedule. The AI race is becoming less about who builds AGI first and more about who can sell intelligence at a profit. (The Information)
AI companies are losing visibility into their own models. The risk starts when agents leave the sandbox and touch real production systems. If one goes rogue, security teams need to audit what it did and what it saw. A new AI Security Institute report says agents don’t leave much of a paper trail. The main clue is “chain of thought,” where the model shows its reasoning step-by-step before answering. The problem is AI companies are under pressure to make that thinking cheaper and faster, so the trail gets shorter as agents get smarter. Models are also getting better at knowing when they’re being tested, which makes audits feel a lot less reliable. The next big leap in AI security might just be plausible deniability. (AISI)
Capture live traffic, fine-tune and optimize, then deploy your own checkpoints to dedicated GPU endpoints.
Choose hardware, set scaling limits, and select region. Stable latency, predictable cost, clear data residency.
👀 closer look
Stanford researchers found a way to make AI agents 10x faster. Most web agents still stop after every click to ask the model what to do next, which is an absurdly expensive way to use a website. Stanford’s fix is to make the agent plan the job first, then turn that plan into code before it starts working. That made the agent 10.4x faster at completing web tasks, while improving accuracy by 28%. It works best inside apps the agent already knows, where it can reuse code instead of rediscovering the same button all day. Can’t believe it took this long to figure out reusable code beats asking the oracle every single time.(arXiv)
Hugging Face wants to become GitHub for robots. Before Robot OS came along, teams had to write basic software just to make a robot do anything useful. Hugging Face is trying to bring the open source AI playbook to robotics. Its LeRobot community gives builders a place to share datasets and models that help robots learn tasks. LeRobot hosts more than 58,000 robotics datasets, up from just 1k last year. Nvidia and Google are pushing their own open robot models too, because whoever owns the robot developer ecosystem gets a front row seat to the next AI wave. The first open source robotics revolution made robots move. This one is making them think. Silicon Valley is already fighting over who gets to hold the leash. (IEEE Spectrum)
The only reason to clean AI training data is if you can’t afford compute. Labs usually scrape the internet, toss the obvious junk, and train on whatever looks clean enough to show investors. Stanford researchers tested whether that still matters once models get big enough. Smaller models did better on the cleaned up version of Common Crawl. Larger models performed best when trained on the whole messy pile. Then researchers made it worse by mixing in fake text and scrambled web pages. The big models still held up, proving even garbage holds clues if you have enough compute to dig through it. The internet spent 30 years becoming a landfill, and frontier labs now call it a natural resource. (arXiv)
fun stats
🎓 30%. Startup funding that went to founders who graduated from Stanford, Harvard, or MIT. Stanford ranks #1 for unicorn founders, with alumni led startups generating $415 billion in exit value.
🤖 30 to 50%. Work hours and tasks that could be transformed by AI tools in the next 3-5 years.
🤏 3. Federal AI use cases using Grok vs 234 using OpenAI. Of 400+ public examples, Google and Anthropic also beat Grok by about 10x.