autosave: 2026-03-11 06:15:32

Author: tecunningham

…recursive-self-improvement-explosion.qmd …f-improvement-explosion-optimization.qmdposts/2025-09-13-recursive-self-improvement-explosion.qmd renamed to posts/2025-09-13-recursive-self-improvement-explosion-optimization.qmd posts/2025-09-13-recursive-self-improvement-explosion.qmd renamed to posts/2025-09-13-recursive-self-improvement-explosion-optimization.qmd

@@ -117,6 +117,49 @@ We could plot data from a variety of different experiments on this plot, & ask s
 
 
 
+#           Apple Tree Model / Low Hanging
+
+Apple tree model of ideas.
+: 
+    - Some problems, all the fruit are low-hanging.
+    - You have an orchard, you send in an army of midgets and they.
+    - You can list out all the insights, some are low, some are 
+
+Model.
+:   - There's a distribution of apples, each has a height.
+    - Each apple is equally valuable.
+    - An agent can pick apples up to their height.
+    - Can represent state with a scalar: the line below which apples are picked.
+    - There are diminishing returns to effort in picking apples below that line, so not all apples are picked.
+
+Implication: you want tests of *reach* for LLMs.
+:   
+    - You don't want to just see whether LLMs can pick apples, but you want to test for their reach.
+    - Thus you find problems which 
+
+Implication: whether they can stand on each others' shoulders.
+:   - 
+
+--------------------------------------------
+
+Apple Tree Optimization
+
+1. LLMs are suddenly able to optimize algorithms pretty well.
+
+2. AI R&D is an algorithm optimization problem -- model training is a big stack of algorithms which we've been optimizing at around 10X/year.
+
+3. It seems like.
+
+--------------------------------------------
+
+LLMs are suddenly able to optimize algorithms pretty well, & so maybe  recursive-self-improvement has kicked off.
+
+I think the critical thing is to distinguish between shallow & deep optimizations. 
+
+Most models of RSI, as far as I can tell, 
+
+--------------------------------------------
+
 #           Toy model of AI stack
 
 See also `posts/2026-02-10-model-of-labs.qmd`
@@ -252,12 +295,6 @@ Josh You / Epoch (Oct 2025) ["Most of OpenAI’s 2024 compute went to experiment
 
 ##          estimates of efficiency growth
 
-|   |   |
-|---|---|
-|   |   |
-|   |   |
-
-
 Anson Ho (Feb 2026): 10X/year.
 : 
     https://epoch.ai/gradient-updates/the-least-understood-driver-of-ai-progress
@@ -321,7 +358,7 @@ Thompson et al. (2022) ["The Importance of (Exponentially More) Computing Power"
 
 
 
-#           Models
+#           Other Peoples' Models
 
 ##                charts
 
@@ -488,37 +525,6 @@ Natural processes often have explosive *regions*.
 : 
     You 
 
-##    metaphors for RSI
-
-Drawing balls from an urn.
-: 
-    There's some distribution of values, then you get a very nice expression. The expected value of $N$ draws just depends on the extreme value distribution of $f(v)$. This is exactly Kortum (1997).
-
-A tool factory makes better tools.
-: 
-    I think this is Jones' metaphor. *Distinct* from Kortum, because the returns to search now depends on the stock of ideas.
-
-Machine tools and regular tools
-: 
-    1. The Dutch make machine tools.
-    2. The machine tools make regular tools.
-    3. The regular tools make products.
-
-    At some point the machine tools become good enough to make themselves, but it's a discrete jump.
-
-Lego - combining ideas.
-: 
-    Weitzmann's recombinant search is like this: you combine ideas to make new ideas, now you have a larger stock of ideas to combine.
-
-Recipes.
-: 
-    You try out recipes, which are combinations of prior recipes. @jones2023recipes -- something like a reconciliation of Weitzman & Kortum.
-
-Blacksmith.
-: 
-    You have a hammer and you spend time making horseshoes, or working on a new hammer.
-
-    A harder hammer both (1) makes horseshoes faster, or (2) makes your hammer still harder.
 
 #           References
 
@@ -603,6 +609,38 @@ Eric Drexler (Aug 2025) ["The Reality of Recursive Self-Improvement"](https://ai
 
 #           Overflow / Offcuts
 
+##    metaphors for RSI
+
+Drawing balls from an urn.
+: 
+    There's some distribution of values, then you get a very nice expression. The expected value of $N$ draws just depends on the extreme value distribution of $f(v)$. This is exactly Kortum (1997).
+
+A tool factory makes better tools.
+: 
+    I think this is Jones' metaphor. *Distinct* from Kortum, because the returns to search now depends on the stock of ideas.
+
+Machine tools and regular tools
+: 
+    1. The Dutch make machine tools.
+    2. The machine tools make regular tools.
+    3. The regular tools make products.
+
+    At some point the machine tools become good enough to make themselves, but it's a discrete jump.
+
+Lego - combining ideas.
+: 
+    Weitzmann's recombinant search is like this: you combine ideas to make new ideas, now you have a larger stock of ideas to combine.
+
+Recipes.
+: 
+    You try out recipes, which are combinations of prior recipes. @jones2023recipes -- something like a reconciliation of Weitzman & Kortum.
+
+Blacksmith.
+: 
+    You have a hammer and you spend time making horseshoes, or working on a new hammer.
+
+    A harder hammer both (1) makes horseshoes faster, or (2) makes your hammer still harder.
+
 
 
 ##       We need more bottom-up modelling of AI's economic effect [UNFINISHED]