I like Noah's blog in general (this is a guest post), but I think that trying to precisely time and phase a new technological development you are watching from within is a fool's errand.

Just like when you upscale a picture 100 times by 200%, and then you zoom in and start seeing all these artifacts, so is the outcome of "recursive intelligence".

We are going to be stuck in a few years of delusion, but there will be some entertainment as a side effect. AI executives and researchers are going to delve even deeper into AI psychosis thinking that the equivalent of psychedelic trip-looking school buses at 4000x zoom of an orange cat's ass, except in the form of text, is actually something intelligent and we just don't have the capacity to understand the "super intelligence's thoughts", all the while pouring another few trillions.

Great analysis. Clearly shows "advancement" is not exponential but logarithmic. Hence, "We need to slow down!!" as a cover for stagnating improvements.

Not where, when.

Asking where is the intelligence explosion suffers the same issue as the Gorman Paradox[1] - the question proceeded the explosion by six months. Timing explosions is like timing the market - infinitely harder than predicting eventual existence.

1. https://codemanship.wordpress.com/2025/12/14/the-gorman-para...

it seems that the big leap has already ended. Now its more cutting costs and improving speed. But maybe it is better this way...

Autonomous recursive self-improvement needs to happen. It doesn't have to happen tomorrow, or in 10 years even, but it has to happen sometime in the near future, or we're kinda screwed.

Why say this? Well, of the plausible futures from here, the worst one is where AI plateus just good enough to make us all useless (or just destroy the pipeline of a critical mass of capable humans for assisted RSI), but at a point dumb enough that we don't get human-superior AI. This might happen or not, but there's no way to predict when something that doesn't exist will, so it's one of those horrors that can keep you up at night. Some of those skeptical about RSI have likened it to the Great Filter.

Even though I started using LLMs as soon as they became available, I was until recently very cautious around using them autonomously. I was of the opinion that I needed to provide heavy supervision around coding tasks. Granted, that was probably justified based on the model capabilities at the time.

However, lately, I have had several experiences that have completely changed my mind - the here is a high level goal (involving gathering production logs, setting up an eval, running tests, iterating, including diagnosis and coming up with new ideas, etc. and don’t bother me until it’s done. The experience has been incredible.

This has been possible because the agent has a way to iterate and hillclimb against a metric it can measure. Relatively easy in the world of software.

I’m now convinced that we will unlock the same gains once make other domains similarly “iterable”.

Eg Agent comes up with 100 new proteins, runs in lab autonomously, gathers results, iterates. One month (or whatever) later, you have your new protein designed.