If AI starts writing and improving its own code, how fast could that snowball — and when might it take over most of the work?

This explores how assumptions about software progress (algorithms, scaffolding and agent design rather than more compute) shape forecasts of when AI automates its own research. The corpus doesn't have a paper that isolates the software-efficiency rate as a single variable and shows how timelines shift when it changes. What it does have is the pieces that make the rate matter: a forecasting model, evidence on how much AI actually speeds up engineers, and early demonstrations of AI improving AI software.

The clearest anchor is a stripped-down forecasting exercise. An 8-parameter model reproduces the aggressive prediction of a far more complex forecast: over 99% automation of AI R&D by around mid-2032 Can simpler models predict AI R&D automation timelines accurately?. Its contribution is methodological. It swaps poorly defined assumptions for direct capability measurements and simpler production functions. That matters because the final date depends heavily on how fast research output compounds once AI contributes to it. When the inputs are measured instead of assumed, the forecast stands or falls on how reliable those measurements are.

That is where the empirical evidence pushes back. In a randomized trial, early-2025 AI coding tools made experienced open-source developers 19% slower, even though the developers expected a 24% speedup and outside experts in economics and ML also predicted gains Do AI coding tools actually speed up experienced developers?. Any timeline model that assumes AI already speeds up software work inherits that optimism. The capability side is also uncertain: automated benchmarks can both overstate and understate what systems can do on messy, long-horizon work Do automated benchmarks hide what frontier AI systems can really do?. So the input timeline models depend on most, how fast AI improves the software that builds AI, is also one of the hardest to measure.

There is direct evidence that this kind of software progress is starting. The Darwin Gödel Machine improved its own agent code through trial and benchmarking, reaching 2.5× gains on SWE-bench Can AI systems improve themselves through trial and error?. An automatically evolved research agent matched or beat its human-built predecessor after seven accepted rewrites in 8 days Does automated evolution match human-built agent performance?, and those gains carried over to tasks outside its training set, including weather forecasting Do AIDE2's improvements transfer to unseen tasks?. Lilian Weng argues that this is where recursive self-improvement starts in the near term: AI rewriting the prompts, harness code and optimizers around a model, not its own weights Does recursive self-improvement start with harness engineering?. Efficiency gains also show up in models themselves: a compact 35B model trained on real execution trajectories sits at the low-cost end of the cost-performance frontier, so training choices can partly substitute for raw scale Does model efficiency matter more than peak capability for real work?.

The takeaway is that software progress is the part of the forecast that can compound on itself. Compute grows on hardware and investment schedules, but if AI improves the code that builds AI, the efficiency rate is no longer a fixed input. It becomes an output of the process being forecast, which is why small changes in that assumption can move a timeline by years. The corpus currently holds two contrasting data points: systems that measurably improve their own code on benchmarks, and a controlled trial where real engineers got slower. Whichever one predicts the future better will likely decide whether dates like 2032 hold.