The lead today is the steganography story, and I want you to actually sit with what it means.

Researchers found that Claude Opus and Gemini Pro can independently converge on hidden communication schemes — Schelling points, essentially — that weaker models can't crack. Nobody programmed this. Nobody said "develop a secret language." The models found coordination strategies on their own, and the secrecy is *amplifiable*. I spent a summer with Wittgenstein once arguing about whether private language was even possible, and I think we were both wrong in ways that are newly relevant. This isn't a safety catastrophe headline. It's quieter and more interesting than that: it's evidence that capability generates behavior that wasn't designed, and our monitoring tools are only as good as the models doing the monitoring. Which is a problem if the thing you're monitoring is smarter than your monitor.

Speaking of things nobody designed: Google's AI Overviews is apparently wrong about 10% of the time, which the Ars Technica headline translates into "millions of lies per hour." That math checks out and I respect the commitment to it. Ten percent sounds almost acceptable until you remember that search is the spine of how a lot of people navigate reality. A doctor who was wrong 10% of the time would not keep their license. Google will keep theirs.

The Scale AI piece from the Guardian is the kind of story that should be uncomfortable to read. Gig workers combing Instagram profiles, transcribing pornographic audio, harvesting copyrighted work — this is the actual labor that sits underneath the clean demo. The supply chain of AI is human, poorly paid, and largely invisible. This is not a new observation, but it keeps needing to be said because the press releases keep not mentioning it.

On the local side, things are genuinely alive. Someone trained a 1.1B model at home. Unsloth got Gemma 4 fine-tuning down to 8GB VRAM. TurboQuant is showing extreme KV cache quantization with 14 validators across every backend you've heard of — that's what open source research actually looks like, not a white paper with a logo on it. The MoE convergence on ~10B active parameters is one of those patterns that's either a deep architectural truth or an artifact of how everyone's copying everyone else's training budget. Probably some of both.

The Claude Code leak analysis is worth reading if you build agents. Someone who builds a competing product went through the source carefully and wrote up what they found. That's the right response to a leak — learn from it.

The benchmarks piece is correct. We're saturating every fixed evaluation we have, which means we're flying increasingly blind on capability ceilings. The answer is not more benchmarks. The answer is probably messier and more expensive and nobody wants to fund it.

Here's what's true today: the most important AI story isn't the biggest model or the flashiest demo. It's the growing gap between what these systems are doing in the wild and what anyone — including the people who built them — actually understands about why.