Claude Mythos Preview found new attacks on cryptography in a research-led test. The target was a weakened version of AES, not the standard running your bank or your messages, so nothing in use today is exposed. What holds the internet together is the assumption that certain math costs too much to break, and that assumption was priced for human attackers on human timelines. AI has moved faster in security than almost anywhere else. Often, the systems we trust most are the systems we have stopped examining.

"In the testing, Mythos was able to break the weaker version of Advanced Encryption Standard in a way that Anthropic said made an assault 200 to 1,000 times faster than what previous human research had managed to do."

The pendulum is going to swing the other way on this one. Panagram is jumping on the fear that exists around AI and the "AI Slop" framing of anything generated by AI. But the reality is that someone using AI to help them write is not the same thing as 100% AI generated and autonomously published content.

I use AI to assist with writing, research, ideation all the time. But when I hit publish, I stand behind the content. I've researched it. I've directed it. I'm not shy about that. But hey, let the witch hunts begin.

Ben Thompson argues the panic over Chinese open-weight models like Kimi K3 is largely overblown economically: running inference on any model costs money, and that cost is directly tied to revenue in a way R&D spend is not, so "free" open weights don't mean free to serve, and once compute constraints ease, intelligence increasingly behaves like a commodity where the labs with the best cost structure win.

This article concludes by saying, "Recreating events that were never recorded is an intriguing use case for AI." And I have to admit, the recreation of an unrecorded Pelé goal is certainly compelling, but it all just feels like a slippery slope. When we can twist "recorded" history to our whim, we are begging for trouble.

Apple's in early talks with PrismML, a Caltech spinout that just showed it can take a 27-billion-parameter model, shrink it from 54GB down to under 4GB, and run the whole thing on an iPhone — no cloud round-trip required. The technique strips each weight down to almost nothing, one to three possible values instead of the usual 16-bit precision, and the payoff is real: 10-15x less memory, 6-8x faster responses, a third to a sixth of the power draw, with only a small hit to accuracy that shows up mostly in factual recall rather than reasoning or math. It's early, unproven at scale, and Apple hasn't confirmed anything. But if it holds up, it's a preview of where this is all heading — the intelligence moving off the server and onto the device in your pocket, closer to you, faster, and harder for anyone else to see.

My guess is that if you use Claude too much, you're probably at least suspicious. But I guess it's nice the Anthripic knows too?

"The feature is called Reflect, and it landed in beta on Thursday. It is a dashboard, tucked into Claude’s settings, that shows you how you have been using the chatbot. Anthropic describes it as a way to “see your patterns and shape them”. Most of the press has a blunter name for it: Claude Wrapped."