AI that builds AI could come sooner than expected
Jack Clark, a co-founder of Anthropic, now thinks AI that can build its own successor is more likely to arrive in 2027 than 2028.
My take: I’ve followed Jack for years, and I think he’s right. If anything, he might be understating it. I think we’ll likely have it within the first few months of 2027. You think AI is improving quickly now? Just wait... this is going to accelerate things even further.
Gemini 4 Argon is out for paying users
Google made Gemini 4 Argon available to its paying Google AI Pro and Ultra subscribers.
My take: Same as I said Monday: I’ll believe it when I see it. Gemini models tend to look great on paper and then fall apart in real use.
Meta’s new model isn’t worth using
Meta released Muse Wonder, its first frontier model, and by Meta’s own tests it’s comparable to the best models from OpenAI and Anthropic.
My take: I’ve tried it, and so have people I trust. We all kind of hate it. It looked good on benchmarks, and in real use it’s not worth your time.
Linus Torvalds is coding with Claude Code
Linus Torvalds, who created Linux (the operating system that underlies many of the things we use every single day) and is one of the most respected programmers alive, posted experimental changes to Linux that he made with Claude Code.
My take: No one serious is writing code by hand anymore. The same thing is coming for other fields. Even the last holdouts are starting to say “oh, this is better,” and the people who still don’t believe it are so far behind it’s crazy.
More people are paying for AI
4.5% of US consumers paid for ChatGPT, Gemini or Claude in August, according to a16z, up from 2.1% a year earlier.
My take: That’s going to keep climbing. With AI, you really get what you pay for. My own AI spending is obscene, and I’m an outlier, but I think most people will spend more than they expect pretty soon, and it’ll be worth it.
— Matt
What should I write about next? Vote or suggest a topic →