<div dir="ltr"><div dir="ltr">You're making the mistake of thinking it understands any of it... it doesn't.  There isn't a KB in the sense you're probably expecting to be there, only the meaning it's extracted in terms of the relationships of the text (in the form of encoded symbols) ingested.  It only appeared to understand how Smalltalk strings work because its training data included plenty of Smalltalk code, in which string manipulation is abundant.  These models are trained on hundreds of gigabytes of text from the usual sources (wikipedia, stackoverflow, open source books, reddit etc.) around the Internet.</div><div dir="ltr"><br></div><div dir="ltr">Sure it's doing far more than simply regurgitating the text it's seen before back to you, but it's a whole lot less than any actual understanding.  Think of it as ELIZA NG++.  Its output often has a sort of fractals with text quality in that it's spitting out something that looks nice (i.e. reasonable) since that's what it was rewarded for doing during training.  I must admit I find the dynamic style transfer endlessly entertaining:<div><br></div><div><div class="gmail-sc-3e5e0208-1 gmail-bMNNoh">me: 'hello' copyFrom: 3 to: 4 would return the string 'l'... answer in the style of shakespeare</div><div class="gmail-sc-3e5e0208-2 gmail-jIBPh"><div class="gmail-sc-3e5e0208-5 gmail-bkjcuL"><p>gpt: Aye,
 'tis true, for the 'hello' string hath a length of five characters, and
 from index three to index four is the specified range, thus the 
resulting substring shall contain only the character at index three, 
which is the letter 'l'.</p><div><div class="gmail-sc-3e5e0208-2 gmail-jIBPh"><div class="gmail-sc-3e5e0208-5 gmail-bkjcuL"><div><div class="gmail-sc-3e5e0208-1 gmail-bMNNoh">-------</div><div class="gmail-sc-3e5e0208-1 gmail-bMNNoh"><br></div><div class="gmail-sc-3e5e0208-1 gmail-bMNNoh">me: 'hello' copyFrom: 3 to: 4 would return the string 'l'... answer in the style of yoda</div><div class="gmail-sc-3e5e0208-2 gmail-jIBPh"><div class="gmail-sc-3e5e0208-5 gmail-bkjcuL"><p>gpt: 'l', the string returning, 'hello' copyFrom: 3 to: 4 would.</p><div><div class="gmail-sc-3e5e0208-1 gmail-bMNNoh">-------</div><div class="gmail-sc-3e5e0208-1 gmail-bMNNoh"><br></div><div class="gmail-sc-3e5e0208-1 gmail-bMNNoh">me: 'hello' copyFrom: 3 to: 4 would return the string 'l'... answer in the style of edgar allan poe</div><div class="gmail-sc-3e5e0208-2 gmail-jIBPh"><div class="gmail-sc-3e5e0208-5 gmail-bkjcuL"><p>gpt: Ah,
 'tis a question of such simplicity, yet a riddle of such intrigue. 'Tis
 true that a single letter may be copied from one point to another. 'Tis
 that single letter which stands between two points, with the answer to 
this conundrum standing between them. Aye, 'twould be precisely that 
single letter 'l' that would be returned.</p></div></div></div></div></div></div></div></div></div></div></div></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Thu, Jan 12, 2023 at 11:07 AM Lauren P <<a href="mailto:drurowin@gmail.com">drurowin@gmail.com</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">Chris, all,<div dir="auto"><br></div><div dir="auto">Very interesting.  Even while looking for the inconsistencies you pointed out, it was still very hard to follow since each answer was so verbose.</div><div dir="auto"><br></div><div dir="auto">My biggest surprise was that it knew to treat '12345' as a Smalltalk from-1 string.</div><div dir="auto"><br></div><div dir="auto">And then proceeded to return 2 characters even after stating the to: term is inclusive.  Something in it's KB was really insistent on the to: term being exclusive, even though the reasoning engine got the correct answer while showing its work.</div><div dir="auto"><br></div><div dir="auto">I had a similar experience with the art AI.  My dad said he got some really cool results from it, so I tried asking it for a lemur playing chess with a penguin and could not get a penguin at all.  All of my queries were so disappointingly met, and I made a lot.</div><div dir="auto"><br></div><div dir="auto">I guess people have forgotten AI winter is still going strong.</div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Mon, Jan 9, 2023, 13:46 tim Rowledge <<a href="mailto:tim@rowledge.org" target="_blank">tim@rowledge.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">Hey, maybe I should ask it about my limits problem :-)<br>
<br>
> On 2023-01-09, at 12:09 PM, Craig Latta <<a href="mailto:craig@blackpagedigital.com" rel="noreferrer" target="_blank">craig@blackpagedigital.com</a>> wrote:<br>
> <br>
> <br>
>     Tim writes:<br>
> <br>
> > Unfortunately I suspect we are going to see a lot of 'content'<br>
> > produced by this system that will be treated as Serious Information To<br>
> > Base Policy On, whilst actually being as valuable as "this bloke down<br>
> > the pub told me..."<br>
> <br>
>     I think there's a huge market opportunity for signing content and verifying its provenance, in ways which are usable and well-known. We haven't been able to trust mediated experiences for a very long time; now that more is digital, we can use tools to identify those who will vouch for it.<br>
> <br>
>     Also, the current stuff being passed off as "AI" is far from it. There's no comprehension or understanding, only glorified brute-force auto-completion, and social engineering via human-written code (backpedaling caveats, mostly). I enjoyed Ezra Klein's recent interview with Gary Marcus about it:<br>
> <br>
>     <a href="https://overcast.fm/+oiPWgxJVM" rel="noreferrer noreferrer" target="_blank">https://overcast.fm/+oiPWgxJVM</a><br>
> <br>
> <br>
> -C<br>
> <br>
> --<br>
> Craig Latta        :: research computer scientist ::<br>
> Black Page Digital ::     Berkeley, California    ::<br>
> 663137D7940BF5C0AF ::    C1349FB2ADA32C4D5314CE   ::<br>
> <br>
> <br>
> <br>
<br>
<br>
tim<br>
--<br>
tim Rowledge; <a href="mailto:tim@rowledge.org" rel="noreferrer" target="_blank">tim@rowledge.org</a>; <a href="http://www.rowledge.org/tim" rel="noreferrer noreferrer" target="_blank">http://www.rowledge.org/tim</a><br>
Strange OpCodes: SLTMDL: Shift Left, Test Mask and Dim the Lights<br>
<br>
<br>
<br>
</blockquote></div>
<br>
</blockquote></div></div>