apple picking

Author: tecunningham

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

@@ -409,15 +409,16 @@ <h1>Apple-Picking Model / Low Hanging</h1>
 <li><p>LLMs are suddenly able to optimize algorithms pretty well — maybe recursive-self-improvement has just kicked off.</p></li>
 <li><p>But the critical question is whether the fruit that it’s picking are low-hanging. If all the optimizations are routine, then it’ll run out pretty quickly.</p></li>
 </ol>
-<section id="apple-picking-model" class="level2">
-<h2 class="anchored" data-anchor-id="apple-picking-model">Apple-Picking Model</h2>
+<section id="bernoulli-apple-picking-model" class="level2">
+<h2 class="anchored" data-anchor-id="bernoulli-apple-picking-model">Bernoulli Apple-Picking Model</h2>
 <p><strong>Motivation:</strong></p>
 <ul>
-<li>Agents are starting to autonomously optimize real-world functions, beating the human state-of-the-art.</li>
-<li>In most models of RSI, agents only automate subtasks in R&amp;D, but now they seem to be doing the whole thing autonomously.</li>
-<li>Intuitively they’re only doing <em>shallow</em> optimizations, not deep optimizations.</li>
-<li>We want a model so that we can define what it means to only do shallow optimizations, and test whether agents obey these predictions.</li>
-<li>Testable predictions: (A) agents asymptote to a lower value than humans; (B) agents are relatively more useful for problems which have had less human time on them; (C) agents will asymptote to higher values if they are given better human starting points.</li>
+<li><p>Agents are showing the ability to autonomously optimize AI training, beating the human state-of-the-art.</p></li>
+<li><p>Where does this end? If you can pay some tokens to improve the training function by <span class="math inline">\(x\)</span>% better, then ….</p></li>
+<li><p>In most models of RSI, agents only automate subtasks in R&amp;D, but now they seem to be doing the whole thing autonomously.</p></li>
+<li><p>Intuitively they’re only doing <em>shallow</em> optimizations, not deep optimizations.</p></li>
+<li><p>We want a model so that we can define what it means to only do shallow optimizations, and test whether agents obey these predictions.</p></li>
+<li><p>Testable predictions: (A) agents asymptote to a lower value than humans; (B) agents are relatively more useful for problems which have had less human time on them; (C) agents will asymptote to higher values if they are given better human starting points.</p></li>
 </ul>
 <p><strong>Setup:</strong></p>
 <ul>