<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=Windows-1252">
<style type="text/css" style="display:none;"> P {margin-top:0;margin-bottom:0;} </style>
</head>
<body dir="ltr">
<div class="elementToProof" style="font-family: Aptos, Aptos_EmbeddedFont, Aptos_MSFontService, Calibri, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
Dear Dave,</div>
<div class="elementToProof" style="font-family: Aptos, Aptos_EmbeddedFont, Aptos_MSFontService, Calibri, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
We have heard many expressive examples and failure examples about ChatGPT. They are intuitive but not systematic. </div>
<div class="elementToProof" style="font-family: Aptos, Aptos_EmbeddedFont, Aptos_MSFontService, Calibri, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
If my writing in the last CDS Newsletter applies to ChatGPT, ChatGPT (along with all other LLMs) is the luckiest fitter for its training set F, among many similar fitters. The luckiest fitter has not been systematically tested on a new test T yet. Its
human developers are keeping improving ChatGPT, by hand-tuning its ever-increasing number of hyperparameters. </div>
<div class="elementToProof" style="font-family: Aptos, Aptos_EmbeddedFont, Aptos_MSFontService, Calibri, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
The outputs from ChatGPT are from humans assisted by computers, not really from computers.</div>
<div class="elementToProof" style="font-family: Aptos, Aptos_EmbeddedFont, Aptos_MSFontService, Calibri, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
Best regards,</div>
<div class="elementToProof" style="font-family: Aptos, Aptos_EmbeddedFont, Aptos_MSFontService, Calibri, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
-John </div>
<div id="appendonsend"></div>
<hr style="display:inline-block;width:98%" tabindex="-1">
<div id="divRplyFwdMsg" dir="ltr"><font face="Calibri, sans-serif" style="font-size:11pt" color="#000000"><b>From:</b> Connectionists <connectionists-bounces@mailman.srv.cs.cmu.edu> on behalf of Dave Touretzky <dst@cs.cmu.edu><br>
<b>Sent:</b> Monday, February 19, 2024 11:21 AM<br>
<b>To:</b> Connectionists <connectionists@cs.cmu.edu><br>
<b>Subject:</b> Re: Connectionists: Statistics versus “Understanding” in Generative AI.</font>
<div> </div>
</div>
<div class="BodyFragment"><font size="2"><span style="font-size:11pt;">
<div class="PlainText">Even GPT 3.5 can answer many questions about the spelling of words, so I<br>
don't think the problem is that the tokenization has removed this<br>
information. I think the problem is that a stack of attention heads<br>
isn't good at applying a novel, arbitrary rule consistently to a long<br>
list of items.<br>
<br>
Also, while the average human would be forgiven if they missed a few<br>
entries in the list of US states that do not contain "a" in their name,<br>
no normal human would claim that "Alaska" or "Texas" belongs in this<br>
list, nor would they persist in such an error after it was pointed out<br>
to them, the way ChatGPT 4 and Gemini do.<br>
<br>
This isn't an argument against the possibility of AGI. It's an argument<br>
that "attention is all you need" might be a bit of an overstatement.<br>
<br>
-- Dave<br>
</div>
</span></font></div>
</body>
</html>