<html><head><meta http-equiv="content-type" content="text/html; charset=utf-8"></head><body dir="auto"><span style="font-size:-webkit-xxx-large">💯</span><div dir="ltr"></div><div dir="ltr"><br></div><div dir="ltr"><br><blockquote type="cite">On Feb 19, 2024, at 08:41, Dave Touretzky <dst@cs.cmu.edu> wrote:<br><br></blockquote></div><blockquote type="cite"><div dir="ltr"><span>Even GPT 3.5 can answer many questions about the spelling of words, so I</span><br><span>don't think the problem is that the tokenization has removed this</span><br><span>information. I think the problem is that a stack of attention heads</span><br><span>isn't good at applying a novel, arbitrary rule consistently to a long</span><br><span>list of items.</span><br><span></span><br><span>Also, while the average human would be forgiven if they missed a few</span><br><span>entries in the list of US states that do not contain "a" in their name,</span><br><span>no normal human would claim that "Alaska" or "Texas" belongs in this</span><br><span>list, nor would they persist in such an error after it was pointed out</span><br><span>to them, the way ChatGPT 4 and Gemini do.</span><br><span></span><br><span>This isn't an argument against the possibility of AGI. It's an argument</span><br><span>that "attention is all you need" might be a bit of an overstatement.</span><br><span></span><br><span>-- Dave</span><br></div></blockquote></body></html>