<div dir="ltr"><div dir="ltr"><div class="gmail_default" style="font-family:arial,helvetica,sans-serif"><br></div></div><br><div class="gmail_quote gmail_quote_container"><div dir="ltr" class="gmail_attr">On Fri, Nov 28, 2025 at 12:17 PM Adrian Tymes via extropy-chat <<a href="mailto:extropy-chat@lists.extropy.org">extropy-chat@lists.extropy.org</a>> wrote:<br></div><div dir="ltr" class="gmail_attr"><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><i><font size="4" face="tahoma, sans-serif"><b>As with other such predictions, this site's timeline has enough<br>
internal contradictions to dismiss its predicted results.<br>The simplest one is: at some point it predicts that China outright<br>
steals Agent-2...and yet, the best Chinese model for that time period<br>
and the one after is well behind Agent-2, despite the operation in<br>
question supposedly rocketing it right up to equivalence.</b></font></i></blockquote><div><br></div><div><font size="4" face="tahoma, sans-serif"><b>I asked Gemini the following question:<span class="gmail_default" style=""> </span> </b></font></div><div><br></div><font size="4" face="tahoma, sans-serif"><b><span class="gmail_default" style="font-family:arial,helvetica,sans-serif">"</span>Somebody told me that the following website has an inconsistency, he claims " it predicts that China outright steals Agent-2...and yet, the best Chinese model for that time period<span class="gmail_default" style=""> </span>and the one after is well behind Agent-2, despite the operation in<span class="gmail_default" style=""> </span>question supposedly rocketing it right up to equivalence". Is there really an inconsistency?<span class="gmail_default" style="">    </span><a href="https://ai-2027.com/">https://ai-2027.com/</a><span class="gmail_default" style="font-family:arial,helvetica,sans-serif">"</span></b></font></div><div class="gmail_quote gmail_quote_container"><font face="arial, helvetica, sans-serif" size="4"><b><br></b></font></div><div class="gmail_quote gmail_quote_container"><font face="arial, helvetica, sans-serif" size="4"><b>And this is what Gemini said<span class="gmail_default" style="font-family:arial,helvetica,sans-serif"> in response</span>:<span class="gmail_default" style="font-family:arial,helvetica,sans-serif"></span></b></font></div><div class="gmail_quote gmail_quote_container"><font face="arial, helvetica, sans-serif" size="4"><b><span class="gmail_default" style="font-family:arial,helvetica,sans-serif">==</span></b></font></div><div class="gmail_quote gmail_quote_container"><p>Yes, based on the logic of how Large Language Models (LLMs) and "weight theft" work, your friend has identified a valid <b>logical inconsistency</b> (or at least a confusing contradiction) often cited in critiques of the "Situational Awareness" paper and the associated <code><a href="http://ai-2027.com">ai-2027.com</a></code> visualizations by Leopold Aschenbrenner.</p><p>Here is a breakdown of why that inconsistency exists and the mechanics behind it.</p><p></p><h3>The Logic of the Inconsistency</h3><p></p><p>To understand why the claim contradicts the graph/prediction, you have to look at the difference between <b>Training</b> and <b>Inference</b>.</p><ol start="1"><li><p><b>The "Theft" Claim:</b> The text (and Aschenbrenner's general thesis) argues that US lab security is currently insufficient and that state actors (like China/CCP) could "outright steal" the model weights.</p></li><li><p><b>What "Stealing Weights" Means:</b> If you steal the weights of "Agent-2" (presumably an AGI-level model in this timeline), you do not need to train the model. You effectively <i>have</i> the model. You only need the compute power to <i>run</i> it (inference), which is a tiny fraction of the compute power required to <i>train</i> it.</p></li><li><p><b>The Resulting Expectation:</b> If China steals Agent-2, China's "Best Model" should immediately become <b>equal to Agent-2</b>.</p></li></ol><p></p><h3>Where the Website Conflicts</h3><p></p><p>The inconsistency arises because the website's visual timelines or specific capability forecasts often depict China as maintaining a <b>persistent lag</b> (e.g., being 1–2 years behind the US).</p><ul><li><p><b>The Text says:</b> China steals the model (Event A).</p></li><li><p><b>The Graph implies:</b> China is still struggling to catch up or is running a model inferior to Agent-2 (Event B).</p></li></ul><p>If China had stolen the weights, the line on the graph representing "Best Chinese Model" should have jumped vertically to meet the US line the moment the theft occurred. If the graph shows China remaining "well behind" Agent-2, it implies they are relying on <b>indigenous development</b> (building it themselves from scratch), which contradicts the text claiming they successfully stole it.</p><p></p><h3>Possible "Steel-man" Explanations</h3><p></p><p>While it looks like an error, proponents of the paper might argue a few nuances, though they are often considered weak defenses against this specific point:</p><ol start="1"><li><p><b>The "Deployment" Lag:</b> They might argue that even with the weights, China lacks the specific infrastructure or system prompts to deploy it effectively immediately. However, for a state actor, this delay would be measured in weeks, not the long durations implied by the charts.</p></li><li><p><b>The "Next Token" Defense:</b> The chart might be tracking the "Ability to Train the <i>Next</i> Model" rather than "Best Available Model." Even if China steals Agent-2, they don't necessarily understand <i>how</i> it was built or have the compute to build Agent-3. They are stuck using a stolen Agent-2 while the US moves on to Agent-3. However, the user prompt specifically notes the Chinese model is "behind Agent-2," which invalidates this defense.</p></li><li><p><b>Inconsistent Narratives:</b> The most likely reality is that the author is presenting two different warnings that conflict when placed on the same timeline:</p><ul><li><p><b>Warning 1:</b> "China will steal our secrets!" (Implying parity).</p></li><li><p><b>Warning 2:</b> "The US must maintain its lead!" (Implying a gap).</p></li><li><p>When visualized, the author accidentally prioritized the "US Lead" visual (Warning 2) over the logical consequence of the "Theft" (Warning 1).</p></li></ul></li></ol><p></p><h3>Verdict</h3><p></p><p><b>Your friend is correct.</b></p><p>If a prediction claims a competitor "outright steals" a specific AI model, but simultaneously predicts that competitor's best model is "well behind" that same stolen model, it is a contradiction. Possession of the weights <i>is</i> possession of the model's capabilities.</p><p><br></p><p><span class="gmail_default" style=""><b style=""><font face="tahoma, sans-serif" size="4">John K Clark</font></b></span><br></p></div><div class="gmail_quote gmail_quote_container"><font face="arial, helvetica, sans-serif" size="4"><b><br></b></font><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><br>
</blockquote></div></div>