When AI speeds up its own research, does feedback decide whether it happens, and starting productivity decide how fast?
How do baseline productivity and recursive feedback separately affect amplification speed?
This explores two separate levers in AI-accelerated research: how fast the system already works (baseline productivity) and how strongly each improvement feeds back into making the next one easier (recursive feedback). It asks which of the two decides whether progress speeds up, and which decides how fast.
This explores two separate levers in AI-accelerated research: how fast the system already works (baseline productivity) and how strongly each improvement feeds back into the next one (recursive feedback). The corpus's main point is that the two do different jobs. Recursive feedback decides *whether* self-amplification happens at all. Baseline productivity mostly decides how fast things move once you know which regime you're in.
The clearest version of this comes from a model that borrows the idea of a reproduction number from epidemiology What determines whether AI self-improvement actually compounds?. An epidemic grows when each case causes more than one new case. In the same way, AI-assisted research compounds when the strength of recursive feedback beats the rate at which research problems get harder ("research hardening"). The surprising part is that this tipping point doesn't depend on any particular capability level. A very productive system whose gains don't feed back strongly enough will just be fast and linear. A modest system can still cross into compounding. The crossing can also happen *before* anyone sees visible acceleration, so watching productivity charts may tell you nothing until the shift is already underway.
A second modeling note looks at the same question from another angle Are AI feedback loops strong enough to sustain recursive self-improvement?. Net acceleration depends on the *product* of how responsive each link in the feedback chain is: better AI improves researcher productivity, which improves the systems, and so on. Because the links multiply, one weak link limits the whole loop, however strong the others are. Its reading of current data is that the loops are getting stronger but aren't yet self-sustaining. In other words, productivity is rising, but the feedback has not yet tipped over.
Why is the feedback term so hard to push up? The empirical notes help explain it. Pure self-improvement tends to stall because of the gap between generating answers and verifying them, collapsing diversity, and reward hacking. Methods that actually work bring in external anchors such as judges, tools, or user corrections Can models reliably improve themselves without external feedback?. Where cheap, objective checking exists, as in AlphaEvolve's automated evaluators, the loops run long enough to produce real discoveries Can machine feedback sustain discovery at test time?. That suggests verification quality may be the hidden variable that sets feedback strength. Direct evidence of sustained compounding is still thin. One reported run of seven successive self-rewrites gives no size or timing for each gain, so it can't tell us whether returns held up or faded Does recursive self-improvement sustain gains or hit diminishing returns?.
The takeaway: if you want to know whether AI research will run away, the speed of today's systems is the wrong thing to watch. The ratio of feedback strength to problem hardening matters more, and the best lever on that ratio seems to be how cheaply and reliably improvements can be checked. The corpus doesn't offer a clean empirical split of the two effects. The separation comes from theoretical models, not measured data.
Sources 5 notes
A recursive reproduction number RAI = χ/aσ determines whether AI-assisted R&D self-amplifies, comparing recursive feedback strength against research hardening rate. The transition can occur before visible acceleration and is independent of any particular capability threshold.
Back-of-the-envelope modeling shows recursive improvement loops depend on the product of elasticities across feedback pathways. Current loops remain too weak for self-sustaining acceleration, though they appear to be strengthening based on data on researcher productivity and system benchmarking trends.
Pure self-improvement stalls due to the generation-verification gap, diversity collapse, and reward hacking. Reliable improvement methods succeed by smuggling in external anchors: past model versions, third-party judges, user corrections, or tool feedback.
AlphaEvolve demonstrates that automated evaluators can sustain evolutionary loops long enough to produce real discoveries—faster algorithms, optimized hardware designs, and improved training methods. The key is that cheap, objective verification closes the generation-verification gap where discovery becomes computationally feasible.
The paper reports seven successive improvements in an 8-day run but provides neither the magnitude of each gain nor their timing. Without score trajectories and longer-horizon data, the evidence supports only that improvements transferred, not that recursive self-improvement sustains returns against diminishing curves.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
- The Economics of Recursive Self-Improvement
- Recursive Criticality of AI Self-Improvement
- Dream-RSI: Recursive Self-Improvement through Evolving Worlds
- Self-Improvements in Modern Agentic Systems: A Survey
- Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models
- The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses