<html><head><meta http-equiv="content-type" content="text/html; charset=utf-8"></head><body dir="auto"><div dir="ltr"></div><div dir="ltr">GPT-4 is surely not reliably accurate; I gave many examples and pointed to Mitchell’s recent work as an example. The two empirical examinations Tom pointed to also note inaccuracies, as do many studies by Yejin Choi, Subbarao Kambhampati, and so on.</div><div dir="ltr"><br></div><div dir="ltr">OpenAI regularly updates their systems with undisclosed mechanisms and addresses some publicly discussed examples. (it is also stochastic, so even on questions it gets right it sometimes makes errors).</div><div dir="ltr"><br></div><div dir="ltr">But their patches are rarely if ever robust; it is absurd to suggest that GPT-4 is accurate, full stop, with no qualification. </div><div dir="ltr"><br><blockquote type="cite">On Feb 17, 2024, at 3:47 PM, Fellous, Jean-Marc - (fellous) <fellous@arizona.edu> wrote:<br><br></blockquote></div><blockquote type="cite"><div dir="ltr">
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
<meta name="Generator" content="Microsoft Word 15 (filtered medium)">
<!--[if !mso]><style>v\:* {behavior:url(#default#VML);}
o\:* {behavior:url(#default#VML);}
w\:* {behavior:url(#default#VML);}
.shape {behavior:url(#default#VML);}
</style><![endif]--><style>@font-face { font-family: "Cambria Math"; }
@font-face { font-family: Calibri; }
@font-face { font-family: Tahoma; }
p.MsoNormal, li.MsoNormal, div.MsoNormal { margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: "Times New Roman", serif; }
a:link, span.MsoHyperlink { color: blue; text-decoration: underline; }
a:visited, span.MsoHyperlinkFollowed { color: purple; text-decoration: underline; }
p { margin-right: 0in; margin-left: 0in; font-size: 12pt; font-family: "Times New Roman", serif; }
p.msonormal0, li.msonormal0, div.msonormal0 { margin-right: 0in; margin-left: 0in; font-size: 12pt; font-family: "Times New Roman", serif; }
span.EmailStyle20 { font-family: Calibri, sans-serif; color: rgb(31, 73, 125); }
.MsoChpDefault { font-size: 10pt; }
@page WordSection1 { size: 8.5in 11in; margin: 1in; }
div.WordSection1 { page: WordSection1; }</style><!--[if gte mso 9]><xml>
<o:shapedefaults v:ext="edit" spidmax="1026" />
</xml><![endif]--><!--[if gte mso 9]><xml>
<o:shapelayout v:ext="edit">
<o:idmap v:ext="edit" data="1" />
</o:shapelayout></xml><![endif]-->
<div class="WordSection1">
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">Hi Gary,<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">Yes, I meant the GPT-4 output. If I understand your answer, it is accurate.<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">Then my point is that whether GPT ‘understands’ or fakes an understanding may be secondary to the fact that it made us understand (to some good? extent) your
point of view on AI. It may not ‘have’ an understanding, but it generates one.<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">That is amazing.<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">What would be good now, is if there was a (relatively safe) feedback mechanism that would allow for a (human) reinforcement of whatever lead to that output. Moving
from a purely feedforward GPT (query->answer) to a recurrent one (answer + human->GPT inner workings)…
<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">Best,<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">Jean-Marc<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
<div>
<div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0in 0in 0in">
<p class="MsoNormal"><b><span style="font-size:11.0pt;font-family:"Calibri",sans-serif">From:</span></b><span style="font-size:11.0pt;font-family:"Calibri",sans-serif"> Gary Marcus <gary.marcus@nyu.edu>
<br>
<b>Sent:</b> Friday, February 16, 2024 9:16 AM<br>
<b>To:</b> Fellous, Jean-Marc - (fellous) <fellous@arizona.edu><br>
<b>Cc:</b> Terry Sejnowski <terry@snl.salk.edu>; connectionists@cs.cmu.edu<br>
<b>Subject:</b> Re: [EXT]Re: Connectionists: Fwd: Statistics versus “Understanding” in Generative AI.<o:p></o:p></span></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<p align="center" style="text-align:center"><strong><span style="color:red">External Email</span></strong><o:p></o:p></p>
<div>
<p class="MsoNormal">On your first question: Is what accurate? There is no substantive claim to be tested in Sejnowski’s schoolyard taunt. <o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">As for my position on GPT-4, one place to start: All my predictions of GPT-4 were poublished before its public release (and before I had access); I think all were correct; I don’t think I revised any. You can read them here: <a href="https://urldefense.proofpoint.com/v2/url?u=https-3A__garymarcus.substack.com_p_what-2Dto-2Dexpect-2Dwhen-2Dyoure-2Dexpecting&d=DwMGaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=Tk9Tsq2lRHVKW_v2w_dSieqtGjS5HRGpyvPdkcTb8w_iY7lnlguK4xlwUl1ueOA_&s=D-WgPlGV1AzE_9ICCZIaS-C9yGeU0zTkuBChEperbZ8&e=">https://garymarcus.substack.com/p/what-to-expect-when-youre-expecting</a><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<p class="MsoNormal" style="margin-bottom:12.0pt">On Feb 16, 2024, at 08:10, Fellous, Jean-Marc - (fellous) <<a href="mailto:fellous@arizona.edu">fellous@arizona.edu</a>> wrote:<o:p></o:p></p>
</blockquote>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<p class="MsoNormal"><span style="font-family:"Tahoma",sans-serif"></span> <o:p>
</o:p></p>
<div>
<p class="MsoNormal">Is it accurate though?<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">Some of us sure could us a concise summary of everyone's position/opinion!<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">Thanks,<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">Jean-Marc<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div class="MsoNormal" align="center" style="text-align:center">
<hr size="2" width="98%" align="center">
</div>
<div id="divRplyFwdMsg">
<p class="MsoNormal"><b><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:black">From:</span></b><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:black"> Connectionists <<a href="mailto:connectionists-bounces@mailman.srv.cs.cmu.edu">connectionists-bounces@mailman.srv.cs.cmu.edu</a>>
on behalf of Gary Marcus <<a href="mailto:gary.marcus@nyu.edu">gary.marcus@nyu.edu</a>><br>
<b>Sent:</b> Friday, February 16, 2024 5:11 AM<br>
<b>To:</b> Terry Sejnowski <<a href="mailto:terry@snl.salk.edu">terry@snl.salk.edu</a>><br>
<b>Cc:</b> <a href="mailto:connectionists@cs.cmu.edu">connectionists@cs.cmu.edu</a> <<a href="mailto:connectionists@cs.cmu.edu">connectionists@cs.cmu.edu</a>><br>
<b>Subject:</b> [EXT]Re: Connectionists: Fwd: Statistics versus “Understanding” in Generative AI.</span>
<o:p></o:p></p>
<div>
<p class="MsoNormal"> <o:p></o:p></p>
</div>
</div>
<div>
<div>
<p align="center" style="text-align:center"><strong><span style="color:red">External Email</span></strong><o:p></o:p></p>
<div>
<p class="MsoNormal">Always with the childish ad hominem. <o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">This field is something else.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<p class="MsoNormal" style="margin-bottom:12.0pt">On Feb 15, 2024, at 11:09 PM, Terry Sejnowski <<a href="mailto:terry@snl.salk.edu">terry@snl.salk.edu</a>> wrote:<o:p></o:p></p>
</blockquote>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span style="font-family:"Tahoma",sans-serif"></span> GPT-4 is a moving target that may eventually replace Gary.<br>
<br>
Terry<br>
<br>
-------<o:p></o:p></p>
</div>
</blockquote>
</div>
<div>
<div>
<p class="MsoNormal"><3ttu1d7S0cx4Qvpk.png><o:p></o:p></p>
</div>
</div>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</div>
</blockquote>
</div>
<div>
<div>
<p class="MsoNormal"><uGPqolaZPAByhZbS.png><o:p></o:p></p>
</div>
</div>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal"><br>
The request you've made is similar to one I've already fulfilled. The image generated earlier depicts an empty room without any elephant present, adhering to your specific instructions. If you need any further adjustments or have another request, feel free
to let me know! <o:p></o:p></p>
<div>
<div>
<p>------<o:p></o:p></p>
</div>
</div>
<div>
<div>
<div>
<div>
<div>
<div>
<div>
<p class="MsoNormal"><img border="0" width="24" height="24" style="width:.2452in;height:.2452in" id="_x0000_i1026" src="https://s.gravatar.com/avatar/de1fef2d6398439dfb2d5f831a6cc89a?s=480&r=pg&d=https%3A%2F%2Fcdn.auth0.com%2Favatars%2Fte.png" alt="User" data-unique-identifier=""><o:p></o:p></p>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
<div>
<div>
<div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><br>
Write a critique of ChatGPT in the style of Gary Marcus<o:p></o:p></p>
</div>
</div>
</div>
</div>
<div>
<div>
<div>
<div>
<div>
<div>
<div>
<p class="MsoNormal" style="background:#AB68FF">ChatGPT<o:p></o:p></p>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
<div>
<div>
<div>
<p>Writing a critique of ChatGPT in the style of Gary Marcus, a scientist known for his clear-eyed scrutiny of artificial intelligence's capabilities and limitations, presents an intriguing challenge. Marcus, with his deep understanding of both the potential
and pitfalls of AI, often emphasizes the importance of grounding AI advancements in robust, generalized learning principles, rather than overreliance on data and statistical patterns.<o:p></o:p></p>
<p><strong>Critique of ChatGPT in the Style of Gary Marcus:</strong><o:p></o:p></p>
<p>ChatGPT, while a marvel of contemporary AI technology, epitomizes the quintessential limitations that besiege deep learning models. At its core, ChatGPT is a linguistic colossus, built upon the edifice of the transformer architecture, yet it is fundamentally
constrained by the inherent limitations of its training methodology and underlying principles. The brilliance of generating coherent, contextually appropriate responses is overshadowed by its occasional lapses into confabulation, a glaring testament to its
lack of true understanding.<o:p></o:p></p>
<p>Gary Marcus would likely point out that ChatGPT, for all its linguistic dexterity, lacks the nuanced grasp of the world that humans possess. It can mimic the form of human knowledge and discourse, but it doesn't truly comprehend the content in the way a
human does. This distinction is crucial; it is the boundary between a sophisticated parrot and an entity capable of genuine thought and understanding. ChatGPT's responses, no matter how convincing, are ultimately derived from patterns in data it has been fed,
not from an intrinsic understanding of those patterns.<o:p></o:p></p>
<p>Furthermore, Marcus might criticize ChatGPT's reliance on sheer volume of data as a substitute for a more nuanced, rule-based understanding of language and the world. This data-driven approach can lead to impressive performances in many scenarios but falls
short when faced with novel situations or when required to apply common sense reasoning that hasn't been explicitly encoded in its training data. The model's occasional incoherence or its propensity to generate plausible but factually incorrect responses are
symptomatic of this deeper issue.<o:p></o:p></p>
<p>Another point of critique would likely focus on the opacity of ChatGPT's decision-making process. Despite recent advances in making AI models more interpretable, understanding how ChatGPT arrives at a particular response remains a challenge. This opacity
complicates efforts to debug or improve the model, especially when it produces erroneous or nonsensical outputs. For AI to truly advance, Marcus might argue, it must not only perform well but do so in a manner that is transparent and understandable to human
overseers.<o:p></o:p></p>
<p>In the spirit of Marcus's advocacy for a hybrid approach to AI development, one could argue that integrating ChatGPT's statistical prowess with more structured, rule-based systems could ameliorate some of its shortcomings. By grounding the model's responses
in a framework of logical reasoning and factual correctness, one might achieve a more reliable, genuinely intelligent system that better approximates human cognition.<o:p></o:p></p>
<p>In conclusion, while ChatGPT represents a significant achievement in natural language processing, its limitations underscore the need for a more comprehensive approach to AI. An approach that incorporates the strengths of both deep learning and symbolic
AI, coupled with a commitment to transparency and understandability, may pave the way for the development of AI systems that are not only more capable but also more aligned with the nuanced complexities of human intelligence and reasoning.<o:p></o:p></p>
</div>
</div>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<p class="MsoNormal"><br>
-------- Forwarded Message -------- <o:p></o:p></p>
<table class="MsoNormalTable" border="0" cellspacing="0" cellpadding="0">
<tbody>
<tr>
<td nowrap="" valign="top" style="padding:0in 0in 0in 0in">
<p class="MsoNormal" align="right" style="text-align:right"><b>Subject: <o:p></o:p></b></p>
</td>
<td style="padding:0in 0in 0in 0in">
<p class="MsoNormal">Connectionists: Statistics versus “Understanding” in Generative AI.<o:p></o:p></p>
</td>
</tr>
<tr>
<td nowrap="" valign="top" style="padding:0in 0in 0in 0in">
<p class="MsoNormal" align="right" style="text-align:right"><b>Date: <o:p></o:p></b></p>
</td>
<td style="padding:0in 0in 0in 0in">
<p class="MsoNormal">Tue, 13 Feb 2024 09:54:48 -0800<o:p></o:p></p>
</td>
</tr>
<tr>
<td nowrap="" valign="top" style="padding:0in 0in 0in 0in">
<p class="MsoNormal" align="right" style="text-align:right"><b>From: <o:p></o:p></b></p>
</td>
<td style="padding:0in 0in 0in 0in">
<p class="MsoNormal">Gary Marcus <a href="mailto:gary.marcus@nyu.edu"><gary.marcus@nyu.edu></a><o:p></o:p></p>
</td>
</tr>
<tr>
<td nowrap="" valign="top" style="padding:0in 0in 0in 0in">
<p class="MsoNormal" align="right" style="text-align:right"><b>To: <o:p></o:p></b></p>
</td>
<td style="padding:0in 0in 0in 0in">
<p class="MsoNormal">Weng, Juyang <a href="mailto:weng@msu.edu"><weng@msu.edu></a><o:p></o:p></p>
</td>
</tr>
<tr>
<td nowrap="" valign="top" style="padding:0in 0in 0in 0in">
<p class="MsoNormal" align="right" style="text-align:right"><b>CC: <o:p></o:p></b></p>
</td>
<td style="padding:0in 0in 0in 0in">
<p class="MsoNormal"><a href="mailto:connectionists@mailman.srv.cs.cmu.edu">connectionists@mailman.srv.cs.cmu.edu</a><o:p></o:p></p>
</td>
</tr>
</tbody>
</table>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><br>
...<o:p></o:p></p>
<div>
<p class="MsoNormal">A few sample visuals from those below (read the essays for sources). The argument in brief is that the systems do fine in statistically canonical situations, but lack depth in their comprehension of complex syntax, negation, unusual circumstances,
etc. Much more context (and many more examples) in the essays themselves.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt">Gary<o:p></o:p></p>
</div>
</div>
</div>
</div>
</blockquote>
</div>
<div>
<div>
<p class="MsoNormal"><a0ET6JyMvGOjno0L.png><o:p></o:p></p>
</div>
</div>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><o:p> </o:p></p>
<div>
<p class="MsoNormal">Gary<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</div>
</div>
</div>
</blockquote>
</div>
<div>
<div>
<p class="MsoNormal"><https__substack-post-media.s3.amazonaws.com_public%2Fimages_ab5345a6-2f04-463d-a53b-789115a1fdb7_851x1024.jpeg><o:p></o:p></p>
</div>
</div>
<div>
<div>
<p class="MsoNormal"><IMG_5914.jpeg><o:p></o:p></p>
</div>
</div>
<div>
<div>
<p class="MsoNormal"><IMG_3818.jpeg><o:p></o:p></p>
</div>
</div>
<div>
<div>
<p class="MsoNormal"><IMG_3811.jpeg><o:p></o:p></p>
</div>
</div>
</div>
</div>
</blockquote>
</div>
</div></blockquote></body></html>