rsi

Author: tecunningham

posts/2025-09-13-recursive-self-improvement-explosion-optimization.qmd

@@ -16,20 +16,35 @@ execute:
   cache: true # caches chunk output
 ---
 
-#           Summary
+TL;DR.
+: 
+    1. AI capabilities are growing very rapidly, in large part due to algorithmic discoveries.
+    2. We are now beginning to see signs of feedback: AI is itself accelerating progress in algorithmic efficiency.
+    3. The strength of the feedback loop depends on the mapping from capabilities to effective R&D, but there's a great deal of uncertainty about that.
+    4. Best guess: (A) increasing human efficiency by 50%; (B) picking the low-hanging fruit.
+    5. 
+
 
+AI capabilities are growing very rapidly.
+: 
+    - There's no consensus for a *scale* in AI capabilities, however we see fairly steady growth across many metrics: time horizon, average benchmark scores (ECI), effective compute, and predictive loss.
+    - We can quantify algorithmic progress with compute efficiency, i.e. the reduction in cost required to reach a given capability score.
 
-Capabilities are growing rapidly, but they're hard to quantify.
-: There's no consensus for a *scale* in AI capabilities. We do see fairly steady growth across many metrics: time horizon, average benchmark scores (ECI), effective compute, and predictive loss.
+Much of the growth in capabilities is due to R&D progress.
+: 
+    - AI researchers have been making a very consistent series of discoveries, typically estimated at increasing compute efficiency by around 4X/year (with many qualifications, discussed below).
+    - The 4X/year increase in algorithmic efficiency seems to be coming from a roughly 2X/year increase in researchers.
 
-Algorithmic progress seems to be around 4X/year.
-:   We can quantify algorithmic progress with compute efficiency, i.e. the reduction in cost required to reach a given capability score.
+We are now seeing signs of AI capabilities accelerating R&D input.
+: 
+    Until recently AI R&D was mostly done without help by LLMs, but we now see evidence for two channels:
 
-    AI researchers have been making a very consistent series of discoveries, typically estimated at increasing compute efficiency by around 4X/year (with many qualifications, discussed below).
+    1. _Augmenting AI researchers:_ AI researchers self-report big efficiency gains, e.g. @anthropic2025claude_work self-report approximately 50% productivity gains.
+    2. _Automating AI research:_ E.g. AlphaEvolve, TTT-Discover, autoresearch.
 
-    The 4X/year increase in algorithmic efficiency seems to be coming from a roughly 2X/year increase in researchers.
+    Both of these effects are hard to measure, & we have a great deal of uncertainty.
 
-AI speedups will cause a loop, but unclear how strong.
+The strength of the feedback depends on TH->uplift.
 : 
     Historically algorithmic progress has been a function of human R&D, but we now expect the AI to itself increase the rate of algorithmic progress, through increaseing effective R&D:
     $$
@@ -49,15 +64,6 @@ AI speedups will cause a loop, but unclear how strong.
 
     A lot of the nearby literature is actually written in terms of returns parameters like $r=\lambda/\beta$ or $r=\lambda \alpha / \beta$ rather than $\beta$ itself. That matters because $r$ is informative about long-run dynamics, but it does not separately identify $\beta$ unless the authors also pin down $\lambda$ and sometimes $\alpha$ @erdil2024estimating; @davidson2025howquickandbigwo.
 
-AI systems are now contributing to algorithmic progress.
-:   Until recently most AI R&D was done without help by LLMs, but we now see evidence for two channels:
-
-    1. _Augmenting AI researchers:_ AI researchers self-report big efficiency gains, e.g. @anthropic2025claude_work self-report approximately 50% productivity gains.
-    2. _Automating AI research:_ E.g. AlphaEvolve, TTT-Discover, autoresearch.
-
-    Both of these effects are hard to measure, & we have a great deal of uncertainty.
-
-