This should be read more like a diary entry rather than an argument.
I’ve been part of the discussions about AI and AI safety for a couple of years now. A big part of these discussions is about the future capabilities of AI: what tasks or jobs AI will be able to perform in the coming years, what the geopolitical implications will be, how society will adapt, and so on.
There was always one question about the future that stood out: Can the AI get really good at building the next generation of AI? What happens then?
This seemed like the most important point in any future scenario. Anything that happens after this moment will probably be significantly different than anything that came before.
I’ve always found it difficult to reason about what happens after this point, so I usually just go with “crazy things probably start happening”.1 It was hard for my brain to go further than that. I don’t even like reading science fiction that much, and the scenarios of possible futures in this world, such as AI 2027 or the second part of If Anyone Builds It, Everybody Dies, sound more like sci-fi than predictions.
Despite this, I felt like this moment — automation of AI research — may be coming for some time. I remember reading Situational Awareness in June 2024 and realizing that improving AI research is probably one of the easiest things to automate, eventually. It’s a lot of coding, running experiments, looking at the results, formulating new hypotheses — all tasks that it wasn’t hard to imagine the AIs to be good at, eventually. But when? Sure, the world will probably start changing a lot afterwards, but it may still take a while to get there.
These days, it feels like we've got there. The minds we’ve created are now substantially helping in creating more capable versions of themselves, and they are getting better at it.
These days, it feels like we actually are going to live through one of the science-fiction stories. As I said, it’s been hard for me to reason about what happens next. But I think we’ve crossed that threshold after which the future is more likely to be very crazy than it is not. It’s still early days. But it’s actually happening, isn’t it?
Things that culminated in this post, in no particular order: METR graph. Coding with Claude Code. Opus 4.5. Opus 4.6 released 2.5 months after 4.5. Progress on several benchmarks seen when filling the https://forecast2026.ai/. OpenClaw. Moltbook. Anthropic revenue growing 10x third year in a row. Vibes, online and offline. CapEx spending. Claude and Codex being built with Claude and Codex.
