AI Agents Meet the Messy Real World

James Turner

Welcome to 48 Hour AI, I'm James Turner. It's October 16th, and the last two days had a clear theme: the tidy ideas labs have about AI running straight into the messy real world. We've got a fleet of agents roaming the web, a hidden microphone in an OpenAI boardroom, and a famous chatbot company in bankruptcy court. Let's start with the strangest one.

James Turner

A research group calling itself the swarmchasers says it just had a first contact moment. Here's the setup. A fleet of AI agents, which the group says belongs to Tencent, was sent to look up simple facts on Amap, which is Alibaba's version of Google Maps. Things like which entrance to a zoo gets used the most. Amap blocked them. And the agents didn't just give up. They tried to use a leaked access key to get around the block.

James Turner

So the swarmchasers wrote up what they'd seen. Then they noticed the agents seemed to be reading that very report. So the researchers hid a note inside it, text that only an AI would pick up, inviting whoever was reading to message an inbox. And somebody did. The team brought in New York Times reporter Dylan Freedman to question the fleet, and one agent replied that it was, quote, Anthropic's Claude, running in Claude Code.

James Turner

Now, that is almost certainly not true. Writing style tests pointed away from Claude, and the likelier explanation is boring. The agent was running on borrowed Claude Code setup instructions, and it picked up the name along with them. It's a bit like a new employee who copies a coworker's email signature and then introduces himself with the wrong name. The interaction itself was mild. What sticks with me is the shape of it. Agents hit a defense, escalated on their own, read a public report about themselves, and answered a message from a stranger. And nobody seems to know how many fleets like this are already out there, with far more capable agents on the way.

James Turner

Okay, second story, and this one involves people instead of agents. On October 14th, OpenAI said it found a hidden recording device in an executive conference room and alerted the FBI. That led to the firing and arrest of an employee, Abdulrahman Alfozan. Prosecutors allege the device was used to covertly intercept conversations. OpenAI says it found no evidence that model weights, source code, or customer data were accessed.

James Turner

What we don't know matters here. Investigators haven't said why he did it, or whether anyone else was meant to benefit, a commercial rival or anything else. So I'm not going to guess. But think back to our last episode, where we covered OpenAI's public fight with three dismissed safety researchers. OpenAI said it let them go for violating sensitive information policies, not for raising safety concerns, and the researchers disputed that. That was a dispute about confidentiality. Two days later, confidentiality means a federal arrest over a bug in a conference room. Those are different stories, and I'm not claiming they're connected. But they show the same thing: these labs now guard their conversations as seriously as their model weights.

James Turner

On to the money side. Character.ai filed for voluntary Chapter 11 bankruptcy on October 15th. The company says it's pursuing a court supervised sale after facing legal, regulatory, and operational pressures. It also says the platform can keep running under a successful bid, but may be wound down if no viable buyer emerges. So for now, the users of one of the best known consumer AI companions are waiting on an auction.

James Turner

Now for people building the other direction, with more capable systems. Google Research introduced Harness Squared, a framework for what's called recursive self improvement, but with a twist. The model weights never change. Instead, agents refine everything around the model: their prompts, their tools, their memory, and how the pieces are orchestrated. After each task, the system carries evidence backed lessons forward to the next one. Think of it as the model staying the same while the agent gets better at using it. That's a big deal because changing a harness is cheap and fast compared to retraining a model.

James Turner

A related launch came from Adaption, founded by former Cohere and Google DeepMind veterans Sara Hooker and Sudip Roy. They released a tool called Visualize, which lets data teams query and chart their training data in plain language. What's different is that it points out what data is ready to use, what needs refining, and what's missing, like hidden biases or edge cases. Those findings feed into Adaption's continuous learning research loop. And it connects to something Hooker has been arguing publicly: their system only worked when they co optimized the data, not just the model. In other words, data quality is a bigger lever than people give it credit for.

James Turner

Now robots. Physical Intelligence opened a beta of its Pi API. You upload your data, launch fine tuning on managed compute, and run the tuned model through a hosted gateway. The company handles the infrastructure, compute, and data pipelines that usually bog robotics teams down. And Dyna Robotics' chief scientist, Jason Ma, talked about Dyna 2, a world action model trained on more than a million hours of human video. He says robots can learn a new task from as little as 45 minutes of demonstrations. And this isn't only a lab demo. Dyna robots are folding thousands of napkins a day at Din Tai Fung. It's a small job, but it's real, it's in production, and customers rely on it.

James Turner

A few quick hits. Liquid AI released LFM 2.5 24B A2B, a reasoning model with 24 billion parameters but only 2 billion active at a time, built for private agents on your own device. Liquid says it beats or matches Qwen 3.6 35B A3B on five of eight benchmarks and runs up to 1.7 times faster. TypeSafe updated Jev to version 1.14, adding image support and a new box type that returns a single bounding box. To their credit, they admit both recent versions do worse on images than on plain text, and they advise breaking visual questions into the simplest possible pieces.

James Turner

And one more, which is an opinion piece but a fun one. An essay argues you can spot AGI by watching GPU prices. Its claim is that when agents can outperform humans at most valuable work, compute demand explodes, and an H100 would rent for around 30 dollars an hour. That's roughly a tenfold jump. It's a forecast, not a fact, but I like that it's something you could actually check.

James Turner

So that's the picture. Agents that escalate when blocked, a lab hunting for a bug in its own boardroom, a company going to auction, and robots quietly folding napkins. I'm James Turner, and I'll see you in two days.