← Toutes les offres
I

Making sense of Recursive Self Improvement

invertedpassion

Global
Publié le 17 novembre 2026

Description du poste

This is a linkpost from my newsletter: https://invertedpassion.substack.com/p/making-sense-of-recursive-self-improvement There’s a whiff of an imminent RSI in the air. The theoretical possibility of AIs recursively improving themselves autonomously until we hit singularity has always existed but the recent METR report seems to have spooked a lot of rationalists and AI insiders about how soon it can arrive. The report that triggered the fear and paranoia is this one, and the chart to pay attention to is the following: It effectively shows how the rate of progress in AI R&D has roughly doubled at Anthropic since October 2025, and METR attributes most of the acceleration to Anthropic’s own use of AI in R&D. Before moving on to making sense of what’s driving this and its potential implications, keep two caveats in mind: • This report redacted many important datapoints as Anthropic didn’t let METR share them• Anthropic is soon going public, so it is in their best interest to project an acceleration (Keep this in mind for the other Anthropic charts below too: they’re from their IPO roadshows)Despite these caveats, I will make the case that a crude form of recursive improvement in AI models is already happening and it’s simply going to become more sophisticated and significantly scaled up. And whatever we call ASI at the end of it is likely to be jagged. Current self-improvement: AI models accelerating AI models releasesWell, models are releasing faster lately. There are several factors driving it and perhaps the biggest one is the sprawling data-generation industry that’s $10bn now, growing rapidly and perhaps could be $100bn in a few years. These data generation companies make RL environments to sell to labs on as many economically useful tasks as they can codify. What does a typical RL environment look like? Imagine slicing the entire economy into sectors and then sub-sectors and sub-sub-sub sectors, and then hiring an expert to design context, task and success rubrics