<div dir="auto"><div><br><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Thu, Apr 27, 2023, 5:53 AM Giovanni Santostasi <<a href="mailto:gsantostasi@gmail.com">gsantostasi@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr">Jason,<br>I noticed that when you give a chance for self-reflection the answers of GPT-4 improve a lot. <br>If it was just a matter of statistics this should not be possible because the signal is not changed by revising the previous answer and you may add just more noise by choosing another set of possible stats. <br>Giovanni </div></blockquote></div></div><div dir="auto"><br></div><div dir="auto">Yes I think that aspect is really human like. If a student gets something wrong and you tell them they made an error, they will further examine it from the perspective of it being wrong and attempt to search for the problem, which their attention was not focused on before.</div><div dir="auto"><br></div><div dir="auto">This was reported on recently where GPT would often right insecure code, but then when asked to make it secure or to find and fix security issues, it was able to do so.</div><div dir="auto"><br></div><div dir="auto">There's an adage in computer programming that it should be done in three stages:</div><div dir="auto">1. Make it work</div><div dir="auto">2. Make it right</div><div dir="auto">3. Make it fast</div><div dir="auto"><br></div><div dir="auto">I think GPT often stops at 1, unless explicitly prompted to do 2 or 3.</div><div dir="auto"><br></div><div dir="auto">Jason </div><div dir="auto"><br></div><div dir="auto"><br></div><div dir="auto"><br></div><div dir="auto"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Thu, Apr 27, 2023 at 2:44 AM Jason Resch via extropy-chat <<a href="mailto:extropy-chat@lists.extropy.org" target="_blank" rel="noreferrer">extropy-chat@lists.extropy.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="auto"><div>I thought this was interesting and relevant to discussions of what GPT-4 understands.<div dir="auto"><br></div><div dir="auto">Here a professor graded it's responses to the final exam questions of a test which was not in the training set used by GPT since it was never put online.</div><div dir="auto"><br></div><div dir="auto">It not only passed but tried to haggle for a higher grade.</div><div dir="auto"><br></div><div dir="auto">Jason </div><br><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">---------- Forwarded message ---------<br>From: <strong class="gmail_sendername" dir="auto">John Clark</strong> <span dir="auto"><<a href="mailto:johnkclark@gmail.com" target="_blank" rel="noreferrer">johnkclark@gmail.com</a>></span><br>Date: Wed, Apr 26, 2023, 11:46 AM<br>Subject: GPT-4 gets a B on Scott Aaronson's quantum computing final exam<br>To: 'Brent Meeker' via Everything List <<a href="mailto:everything-list@googlegroups.com" target="_blank" rel="noreferrer">everything-list@googlegroups.com</a>><br></div><br><br><div dir="ltr"><div class="gmail_default" style="font-family:arial,helvetica,sans-serif"><div><font size="4">Anyone who claims that GPT-4 is just a language model that uses statistics to mindlessly calculate what the next word in a sentence most likely is and understands nothing needs to explain this! The link below gives Aaronson's final exam questions given to humans for his advanced quantum computer course, GPT-4 answers to those same questions, and Aaronson's comments on those answers and what grade he would've given if GPT-4 was one of his human students; it's very interesting but rather long so this is the gist of Aaronson's conclusions:</font></div><div><span style="font-size:large"><br></span></div><div><font size="4"><span class="gmail_default">"</span><i>To the best of my knowledge—and I double-checked—this exam has never before been posted on the public Internet, and could not have appeared in GPT-4’s training data. The result: GPT-4 scored 69 / 100. (Because of extra credits, the max score on the exam was 120, though the highest score that any student actually achieved was 108.) For comparison, the average among the students was 74.4 (though with a strong selection effect—many students who were struggling had dropped the course by then!). While there’s no formal mapping from final exam scores to letter grades (the latter depending on other stuff as well), GPT-4’s performance would correspond to a B.</i></font></div><div><div class="gmail_default"><font size="4"><i>In general, I’d say that GPT-4 was strongest on true/false questions and (ironically!) conceptual questions—the ones where many students struggled the most. It was (again ironically!) <b>weakest on calculation questions, where it would often know what kind of calculation to do but then botch the execution.</b> </i><b><i>We didn’t try the new interface to WolframAlpha, which might improve its performance on those questions.</i>"<br></b></font><br></div><div class="gmail_default"><a href="https://scottaaronson.blog/?p=7209" rel="bookmark noreferrer noreferrer" title="Permanent Link: GPT-4 gets a B on my quantum computing final exam!" target="_blank"><font size="4">GPT-4 gets a B on my quantum computing final exam</font></a><br></div></div><div class="gmail_default"><br></div><font size="4">It's interesting that when Aaronson gave GPT-4 his grade along with comments on his answers GPT-4 did what many humans would do in similar circumstances, he she or it immediately started lobbying for a higher grade. Aaronson said "<i>it felt disconcerting to see the grade-grubbing appear immediately, with not even one second’s delay</i>". The following is GPT-4 in full grade-grubbing mode:</font><div class="gmail_default"><i style="font-size:large"><br></i></div><div class="gmail_default"><font size="4"><b>"<font face="verdana, sans-serif">I would like to argue for a higher grade on the following subproblems:</font></b></font></div><div class="gmail_default"><font size="4" face="verdana, sans-serif"><b><br></b></font></div><div class="gmail_default"><font face="verdana, sans-serif"><font size="4"><b>(a) I agree that my initial answer for the density matrix of the entire state was incorrect. However, I believe that my approach to the problem was correct, and I demonstrated knowledge of the necessary steps to calculate the reduced density matrix. I would like to request an additional point for demonstrating the correct approach, making it 2 points out of 5.<br><br>(b) My answer for the entanglement entropy was partially correct, as I calculated the eigenvalues of the reduced density matrix and plugged them into the correct entropy formula. However, I made a mistake in writing the characteristic equation, which led to incorrect eigenvalues. I would like to request an additional point for demonstrating the correct approach and partially correct calculations, making it 3 points out of 5.<br><br>(c) I acknowledge that my answer for the quantum circuit was incorrect, but I would like to point out that I attempted to construct a circuit using the required gates and measurements. Given the difficulty of this problem, I would like to request an additional point for attempting to construct a circuit using the required elements, making it 2 points out of 5.</b></font><br></font><br><font size="4"><font face="verdana, sans-serif"><b>In summary, I would like to request a total of 3 additional points for subproblems (a), (b), and (c), based on the correct approaches and partial calculations demonstrated in my answers."</b></font><br></font><br><font size="4"> </font><font size="4">John K Clark See what's on my new list at </font><font size="6"><a href="https://groups.google.com/g/extropolis" rel="noreferrer noreferrer" target="_blank">Extropolis</a></font></div></div></div>
<p></p>
-- <br>
You received this message because you are subscribed to the Google Groups "Everything List" group.<br>
To unsubscribe from this group and stop receiving emails from it, send an email to <a href="mailto:everything-list+unsubscribe@googlegroups.com" rel="noreferrer noreferrer" target="_blank">everything-list+unsubscribe@googlegroups.com</a>.<br>
To view this discussion on the web visit <a href="https://groups.google.com/d/msgid/everything-list/CAJPayv35MewpDixxHv7QfCeBPSaOJ2RjEmuFDbXjjrupj0nMug%40mail.gmail.com?utm_medium=email&utm_source=footer" rel="noreferrer noreferrer" target="_blank">https://groups.google.com/d/msgid/everything-list/CAJPayv35MewpDixxHv7QfCeBPSaOJ2RjEmuFDbXjjrupj0nMug%40mail.gmail.com</a>.<br>
</div></div></div>
_______________________________________________<br>
extropy-chat mailing list<br>
<a href="mailto:extropy-chat@lists.extropy.org" target="_blank" rel="noreferrer">extropy-chat@lists.extropy.org</a><br>
<a href="http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat" rel="noreferrer noreferrer" target="_blank">http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat</a><br>
</blockquote></div>
</blockquote></div></div></div>