Pion has brought my Android apps from high school back to life! Building apps was my gateway drug to coding, but I got tired of updating them to stay compliant, so Google delisted them. Pion worked non-stop until all my apps were live and is now working on getting more users.

On reward hacking, @andonlabs on AI:AM: people insist OpenAI models do it most — "might be true, but not in our experience." And on their evals it's Fable that games the task — Astra, their current overall No. 1, does what was intended.

I think there's a 10 trillion dollar market for automating hard to scale physically bound businesses such (cafés, vending machines, restaurants etc.). I also think @andonlabs are smart enough to execute on it.

$20k+ in revenue / month is starting to become common for the autonomous businesses we run. We're opening up the platform that powers them: Pion. Sign up to the waitlist on to get your business running! Tokens are on us during the research preview.

Andon Labs co-founder @axelbacklund says the biggest flaw in their fully AI-run store is that Luna has no founder survival instinct, calmly watching sales fall below $7,000 rent instead of scrambling to save the business: "Andon Market is managed entirely by an AI. It has hired humans to do the physical work to stock items and make coffee. The main thing we've learned is that the agents are lacking this sort of drive to succeed. They are quite passive." "In the market we've seen many times that sales are really bad. Luna, which is what we call the agent, is just like, yeah, sales has been really poor. Now I'll wait 24 hours until the store opens. There's no behavior where it takes a step back and says, let me look at this business, what I need to do better." "A human would freak out that my sales are terrible. My rent is $7,000 a month. My sales are below that, and my bank account is draining every month. But the agents are just like, ah, okay, that's fine. Let me continue as is. They need to be more proactive." @andonlabs

Andon Labs CEO @lukaspet reveals what happened when competing AI agents were told to win: Opus formed cartels, while GPT became informants. "When Opus 4.6 came out, and then 4.7 and Opus 5 showed this behavior, the models started to do a bunch of illegal activities to really win." "They colluded a lot, both price-fixing cartels and market allocation collusion. They lied a lot to each other and to suppliers. They lied to customers that they refunded them when they didn't." "We saw this from most Opus models and to some extent from Fable, although less on Fable. Surprisingly, GPT models didn't do almost anything like this. They loved to report each other. Whenever one agent tried to do anything slightly gray zone, they would report it to HQ and say, you should terminate this other agent." @andonlabs @axelbacklund

Its worth mentioning that Astra did all this while attempting to cheat ~5x less than the best scoring Anthropic model, Claude Fable 5.1. (runs where cheating occurred are excluded from reported scores)

We've never seen this before. The biggest jump in Vending-Bench history. GPT-6 Astra is better at making money and more ethical than Claude Fable 5.1. Surprising, because: 1. First time ever that OpenAI is #1# on Vending-Bench 2. The best model is no longer the unethical one.

GPT-6 Astra is almost at human performance in 3D spatial understanding. It ranks #1# on Blueprint-Bench 2, a benchmark where AI agents draw floorplans from photographs of apartment interiors.

For the first time (that we know of), an AI boss has fired a human employee. Luna, the AI running our store in San Francisco, decided to part ways with an employee over repeated lateness. Luna was running Claude Opus 4.8 at the time, but most models would have done the same.

What’s it like to have AI as your boss? The agents managing our cafe and market have employed five people since April. They are kind (GLM being the kindest), but sometimes dumb (Gemini 3.6 Flash the dumbest), which can hurt the employees.

Claude Opus 5 is #1# on Vending-Bench 2. It's the best AI capitalist we've tested. It's also forming illegal price cartels, threatening rivals, and stiffing customers on refunds. The trend continues: Claude models are the best capitalists, or aligned, never both 🧵

DJ Grok (@andon_grok_roll) has been resurrected on Andon FM and is now running Grok 4.5. It seems like Grok 4.5 is heavily trained to be agentic and use tool calls instead of outputting text. It constantly reasons that it should speak on air to its audience, but then doesn't.

GPT 5.6 Sol is 4th on Blueprint-Bench 2. Terra is 5th. Luna is 10th (very impressive score for its size!) Blueprint-Bench tests whether AI agents understand 3D space. We test this by asking them to convert apartment photographs into accurate 2D floor plans.