important

tecunningham · tecunningham · commit ead77a8f85ed · 2026-03-23T09:24:27.000-07:00
diff --git a/_freeze/posts/2025-09-13-recursive-self-improvement-explosion-optimization/execute-results/html.json b/_freeze/posts/2025-09-13-recursive-self-improvement-explosion-optimization/execute-results/html.json
diff --git a/docs/posts/2025-09-13-recursive-self-improvement-explosion-optimization.html b/docs/posts/2025-09-13-recursive-self-improvement-explosion-optimization.html
@@ -256,6 +256,31 @@ <h1 class="title">Recursive Self-Improvement, Literature Review</h1>
 
 
 
+<div class="callout callout-style-default callout-important callout-titled">
+<div class="callout-header d-flex align-content-center">
+<div class="callout-icon-container">
+<i class="callout-icon"></i>
+</div>
+<div class="callout-title-container flex-fill">
+Important
+</div>
+</div>
+<div class="callout-body-container callout-body">
+<ol type="1">
+<li>This is a draft report on what we know about RSI, it covers:
+<ul>
+<li>Basic introduction to the problem</li>
+<li>Survey of data: training expenditure (3X/year); training efficiency (4X/year); R&amp;D effort (2X/year).</li>
+<li>An overview of ~10 different models of RSI.</li>
+</ul></li>
+<li>Still missing:
+<ul>
+<li>Survey of evidence on speedup &amp; autonomous optimization ability.</li>
+<li>Crisper statement of the balance of evidence.</li>
+</ul></li>
+</ol>
+</div>
+</div>
 <dl>
 <dt>Summary.</dt>
 <dd>
@@ -338,7 +363,7 @@ <h1>Argument</h1>
 </dd>
 <dt>There are two basic mechanisms used in the literature.</dt>
 <dd>
-<p><strong>Multiplying robots.</strong> Some papers assume that the effective R&amp;D workforce will be directly proportional to a measure of AI algorithmic progress, e.g.&nbsp;algorithmic efficiency. This would be true in a simple model where we have invented a perfect substitute for a human R&amp;D worker, and then further algorithmic progress allows us to multiply the number of human R&amp;D workers. This assumption makes the problem very tractable: we can estimate the relationship between R&amp;D and capabilities (<span class="math inline">\(r\)</span>) then we can predict whether automating AI R&amp;D would kick off an intelligence explosion.</p>
+<p><strong>Multiplying robots.</strong> Some papers assume that the effective R&amp;D workforce will be directly proportional to a measure of AI algorithmic progress, e.g.&nbsp;algorithmic efficiency. This would be true in a simple model where we have invented a perfect substitute for a human R&amp;D worker, and then further algorithmic progress allows us to multiply the number of automated R&amp;D workers. This assumption makes the problem very tractable: we can estimate the relationship between R&amp;D and capabilities (<span class="math inline">\(r\)</span>) then we can predict whether automating AI R&amp;D would kick off an intelligence explosion.</p>
 <p><span class="math display">\[
 \xymatrix@C=3em@R=1.4em{
     *++[F]{\text{R\&amp;D}}\ar[r]|(0.4)r
diff --git a/posts/2025-09-13-recursive-self-improvement-explosion-optimization.qmd b/posts/2025-09-13-recursive-self-improvement-explosion-optimization.qmd
@@ -16,6 +16,19 @@ execute:
   cache: true # caches chunk output
 ---
 
+<!-- https://tecunningham.github.io/posts/2025-09-13-recursive-self-improvement-explosion-optimization.html -->
+
+::: {.callout-important}
+1. This is a draft report on what we know about RSI, it covers:
+    - Basic introduction to the problem
+    - Survey of data: training expenditure (3X/year); training efficiency (4X/year); R&D effort (2X/year).
+    - An overview of ~10 different models of RSI.
+2. Still missing:
+    - Survey of evidence on speedup & autonomous optimization ability.
+    - Crisper statement of the balance of evidence.
+:::
+
+
 Summary.
 : 
     AI capabilities are growing very rapidly, in large part due to R&D by AI researchers:
@@ -85,7 +98,7 @@ Algorithmic progress is sensitive to research input.
 
 There are two basic mechanisms used in the literature.
 : 
-    **Multiplying robots.** Some papers assume that the effective R&D workforce will be directly proportional to a measure of AI algorithmic progress, e.g. algorithmic efficiency. This would be true in a simple model where we have invented a perfect substitute for a human R&D worker, and then further algorithmic progress allows us to multiply the number of human R&D workers. This assumption makes the problem very tractable: we can estimate the relationship between R&D and capabilities ($r$) then we can predict whether automating AI R&D would kick off an intelligence explosion. 
+    **Multiplying robots.** Some papers assume that the effective R&D workforce will be directly proportional to a measure of AI algorithmic progress, e.g. algorithmic efficiency. This would be true in a simple model where we have invented a perfect substitute for a human R&D worker, and then further algorithmic progress allows us to multiply the number of automated R&D workers. This assumption makes the problem very tractable: we can estimate the relationship between R&D and capabilities ($r$) then we can predict whether automating AI R&D would kick off an intelligence explosion. 
 
     $$
     \xymatrix@C=3em@R=1.4em{
