Recursive Self-Improvement might not be imminent despite the lab hype

The idea of Recursive Self-Improvement (RSI) and a software intelligence explosion is everywhere right now, but the actual evidence is still pretty thin. I've been looking into the discussion between Nathan Lambert and JS Denain from Epoch AI, and the main takeaway is that while labs like OpenAI and Anthropic are seeing AI be used more in their own workflows, it's a far cry from full automation of the AI researcher's job.

Is there actual evidence for RSI?

When you look at the public data, there isn't a "smoking gun" for a self-sustaining acceleration loop. JS Denain pointed out that OpenAI's own blog posts showed a 2X monthly increase in Codex spending by their internal researchers. While that proves the researchers are getting significant value out of the tools, it doesn't necessarily mean the AI is rewriting its own architecture to become smarter. It's more likely a sign of productivity gains rather than a looming intelligence explosion in the next six months.

Why is there a gap between internal lab views and public data?

There's always a suspicion that people inside the frontier labs are seeing something we aren't, which is why some internal updates sound more urgent or "scary" than the public benchmarks. However, Denain suggests we shouldn't assume there's some secret, massive leap happening behind closed doors. A lot of the "acceleration risk" chatter might be more about how lab cultures evolve over time rather than a specific, measurable breakthrough that warrants panic.

What does the "jaggedness" of AI capabilities mean?

The conversation touched on the "jagged frontier," which is the idea that AI can perform a complex task (like writing a specific piece of code) but fail at a seemingly simpler task that requires a different kind of reasoning. This inconsistency makes it hard to track whether a model is actually "improving" its general intelligence or just getting better at the specific benchmarks we use to measure it. If the capabilities are jagged, a spike in Codex usage doesn't automatically translate to a general leap in research ability.

The role of robotics and post-training

The discussion also looped back to how robotics fits into this acceleration. If AI can't improve itself through software alone, the physical world becomes the next bottleneck. Additionally, the "recipe" for frontier post-training remains a closely guarded secret, but the trend is clearly moving toward more sophisticated reinforcement learning and synthetic data loops.

Ultimately, we're seeing a massive increase in how AI is used to build AI, but we aren't at the point where the human is out of the loop. The "acceleration" is real, but the "self-sustaining" part is still a hypothesis.