RSI

Author: tecunningham

_freeze/posts/2025-09-13-recursive-self-improvement-explosion-optimization/execute-results/html.json

@@ -9,7 +9,7 @@
 <link>tecunningham.github.io/</link>
 <atom:link href="tecunningham.github.io/index.xml" rel="self" type="application/rss+xml"/>
 <description>{{&lt; meta description-meta &gt;}}</description>
-<generator>quarto-1.5.57</generator>
+<generator>quarto-1.8.25</generator>
 <lastBuildDate>Sat, 07 Mar 2026 08:00:00 GMT</lastBuildDate>
 <item>
   <title>When You Overtake More Runners than You’re Overtaken by</title>

docs/index.html

@@ -3,6 +3,7 @@
     "listing": "/index.html",
     "items": [
       "/posts/2024-10-27-from-citlali.html",
+      "/posts/2026-03-19-heebie-jeebies.html",
       "/posts/2023-08-13-runner-percentile.html",
       "/posts/2026-01-29-knowledge-creating-llms.html",
       "/posts/2025-12-30-llm-verification.html",

docs/index.xml

@@ -2,14 +2,14 @@
 <urlset xmlns="http://www.sitemaps.org/schemas/sitemap/0.9">
   <url>
     <loc>tecunningham.github.io/index.html</loc>
-    <lastmod>2023-07-26T14:50:16.883Z</lastmod>
+    <lastmod>2025-10-29T15:46:32.992Z</lastmod>
   </url>
   <url>
     <loc>tecunningham.github.io/posts/2023-01-31-social-media-suspensions-data.html</loc>
     <lastmod>2025-03-10T06:05:27.346Z</lastmod>
   </url>
   <url>
-    <loc>tecunningham.github.io/posts/2025-09-13-recursive-self-improvement-explosion-optimization.html</loc>
-    <lastmod>2026-03-21T12:33:54.246Z</lastmod>
+    <loc>tecunningham.github.io/posts/2026-03-19-heebie-jeebies.html</loc>
+    <lastmod>2026-03-20T03:28:02.301Z</lastmod>
   </url>
 </urlset>

docs/listings.json

@@ -18,11 +18,49 @@ execute:
 
 #           Summary
 
+
+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.
+
+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.
+
+    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 3X/year increase in algorithmic efficiency seems to be coming from a roughly 2X/year increase in researchers.
+
+AI speedups will cause a loop, but unclear how strong.
+: 
+    Historically algorithmic progress has been .
+    $$
+    \xymatrix@C=1.4em@R=1.4em{
+        & *++[F]{R\&D_t}\ar[d] & *++[F]{R\&D_{t+1}}\ar[d] & *++[F]{R\&D_{t+2}}\ar[d] \\
+        \ar[r]& *++[F]{Algorithms_t}\ar[r] & *++[F]{Algorithms_{t+1}}\ar[r]\ar[ur]|{??} & *++[F]{Algorithms_{t+2}}
+    }
+    $$
+
+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.
+
+
+
+
+
+
+
+
+#           Summary (OLD)
+
 1. **Baseline model:**
+    - Frontier model capability is growing at 9X/year (measured in effective compute)
     - Frontier training compute is growing at 3X/year
     - Algorithmic efficiency is growing at 3X/year
-    - Frontier model capability is growing at 9X/year (measured in effective compute)
-    - R&D staff growing at 2X/year.
+    - R&D staff is growing at 2X/year.
 
 2. **Two ways we can get RSI:** (1) augmentation of AI R&D; (2) automation of AI R&D.
 
@@ -120,17 +158,14 @@ Q: where is AI likely to help?
 
 
 
-#          Data
+#          Data on Compute Growth
 
-My best estimates.
-
-1. Training compute expenditure ($) has been growing around 3X/year, but will slow to 1.1X/year over 2026-2030.
-2. Training compute (FLOP) has been growing around 4X/year, but will 
-3. Algorithmic efficiency has been growing around 3X/year, not clear if it will slow down or accelerate.
-
-Takeaway: the outside-view seems roughly like `4-5X/year` for frontier training compute, `~3.5X/year` for training cost, `~2.3X/year` for installed compute stock, and `~1.37X/year` for hardware price-performance.
+Best estimates.
+: 
+    1. Training compute expenditure has been growing around 3X/year, but will gradually fall to 1.1X/year over 2026-2030.
+    2. Training compute (FLOP) has been growing around 4X/year, but will fall to around 1.5X/year.
+    3. Algorithmic efficiency has been growing around 3X/year, it is hard to forecast future trends.
 
-### Training compute growth
 
 | source                                | scope                                 | growth     | years             | quantity                   |
 | ------------------------------------- | ------------------------------------- | ---------- | ----------------- | -------------------------- |
@@ -223,9 +258,6 @@ Takeaway: the outside-view seems roughly like `4-5X/year` for frontier training
 
     The interactive graph currently shows about 4.6x/year growth in FLOPs of notable models, over 2020 - July 2025 (the latest datapoint).
 
-@epoch2026canaicompaniesprofitable "Can AI companies become profitable?"
-: 
-    Useful mainly as a reminder that frontier-model economics are not just final training compute.
 
 @you2025openaicomputespend "Most of OpenAI's 2024 compute went to experiments"
 : 
@@ -234,12 +266,12 @@ Takeaway: the outside-view seems roughly like `4-5X/year` for frontier training
     - only a minority of R&D compute appears to have gone to the final training runs of released models
     - GPT-4.5 final training run was only a modest share of the total R&D bucket
 
-This seems relevant to the RSI story: training runs are scaling fast, but frontier labs are also spending enormous amounts on experiments, unreleased models, and inference.
-
 
-##          Estimates of LLM Algorithmic Progress
+#          Data on Algorithmic Progress
 
-These estimates are also not all measuring the same thing. Some are narrow pre-training efficiency estimates; others are all-in software-progress estimates; others are small-scale case studies.
+Best estimates.
+: 
+    - Around 4X/year, including the entire stack (GPU, pretraining, posttraining, elicitation).
 
 | source                                  | scope                                                     | progress     | years          | quantity                                                  |
 | --------------------------------------- | --------------------------------------------------------- | ------------ | -------------- | --------------------------------------------------------- |