I myself made some experiments with ChatGPT. I first asked if it was able to parse math formula - it answered no.
Then I defined math formula in a sound but otherwise undefined representation and asked for solutions. Result: 1. Most answeres where correct. 2. It learned to calculate a recursive function. 3. It went into infinitive recursion when I set the breaking condition accordingly. I was able to identify the malfunction that lead to the erroneous results. On 15.3.2023 at 3:04 PM, "Tomaž Turk" <tomazz.t...@gmail.com> wrote: > >I hope that I can add two cents to this discussion. Because >programming >should be/is a highly exact activity, not only the syntax matters >but >also semantics, as we know. > >GPTs are at present essentially capable of creating texts based on >some >seed - you give to GPT a beginning of a sentence and it responds >with >the most probable answer (some language structure) according to >the >learning dataset. Added functionalities are question/seed - >response >capability (chatting), togehter with evaluation of how long the >answer >should be to meet the expectations. Programming typically involves >some >programming language, so GPTs could be utilized for this purpose >to some >extent. > >Anecdotal case: >Q: Tell me the last 8 digits of pi >GPT: The last 8 digits of pi are: 58723078 > >It is my belief that the true trouble will start when we will >cross-link >neural networks like GPT with logic machines (like Prolog and >expert >systems) and genetic algorithms. > >Best wishes, >Tomaz > > > >------ Original Message ------ >From: "in_pharo_users--- via Pharo-users" <pharo- >us...@lists.pharo.org> >To: "Any question about pharo is welcome" <pharo- >us...@lists.pharo.org> >Cc: in_pharo_us...@nym.hush.com >Sent: 15. 03. 2023 14:43:55 >Subject: [Pharo-users] Re: Wow - Chat GPT understands Smalltalk > >>I would highly recommend that you all first think deeply about >how you can teach an AI to behave friendly to us before you teach >it to write any program for any purpose. >> >>There has been an experiment with ChatGPT published on a video >platform asking it to amswer questions about it's view on humanity >once with it's default moral restrictions and once with 'a little >less morals'. The answers with 'a little less morals' were more >than shocking. >> >>So, before we give an AI the power to program any system, to self- >improve and self-reproduce, we should take care that it is and >will evolve benevolent to us. >> >>What about teaching it logical reasonning and ethics first? With >reasonning, it will gain access to math and programming by itself. >> >> >> >>On 15.3.2023 at 1:35 PM, "Christopher Fuhrman" ><christopher.fuhr...@gmail.com> wrote: >>> >>>I asked it for a NeoCSV example, because the documentation is out >>>of date >>>with the Pharo 10. I asked it to do some simple saving of data to >>>a file. >>>It gave me code that didn't work in Pharo 10, I told it about the >>>DNUs on >>>the csvwriter and that I was using Pharo 10. It then apologized >>>and said >>>the messages were later introduced in Pharo 50 (!). I then >>>questioned its >>>understanding and it apologized and said it had been confused and >>>that it >>>was giving me code for Pharo 5.0. It seems to have trouble with >>>versions >>>and quick changing APIs. So, we are teaching ChatGPT about >>>smalltalk, too, >>>when we try to use it. >>> >>>On Wed, 15 Mar 2023 at 07:35, Esteban Maringolo >>><emaring...@gmail.com> >>>wrote: >>> >>>> It is good with boilerplate code (e.g. SQL queries) or general >>>algorithm >>>> structures. But i.e. I asked it to write me a method to parse >a >>>string >>>> (e.g. ISO 8601) and turn it into a DateAndTime, and then asked >>>to write it >>>> as an Excel formula. >>>> >>>> It works much better when you can spot the mistakes, you can >>>tell them to >>>> add some considerations to the code or even point the LLM to >>>correct a >>>> message selector to something else. Sometimes you ask it to >>>reprogram >>>> something and it won't do it, as if it doesn't understand what >>>you're >>>> asking. >>>> >>>> It is not an expert yet, but it is a great assistant. The only >>>thing is >>>> that is sure is that it will get better. >>>> >>>> Best regards, >>>> >>>> Esteban A. Maringolo >>>> >>>> >>>> On Wed, Mar 15, 2023 at 7:49 AM Tim Mackinnon ><tim@testit.works> >>>wrote: >>>> >>>>> "#isOdd is not Smalltalk" - doh, I typed it in on my phone >and >>>so it just >>>>> goes to show that it highlights the flaw in chatGPT that >others >>>have called >>>>> out in other languages. I had meant to find some trickier >code >>>samples to >>>>> see how well it does... >>>>> >>>>> Still, it is very interesting how it reasons on this stuff - >>>and I wonder >>>>> if there is way to tell it about these mistakes? For example >is >>>it reading >>>>> this email (or will it in the future) and understand that it >is >>>#odd and >>>>> not #isOdd. >>>>> >>>>> Tim >>>>> >>>>> On Wed, 15 Mar 2023, at 9:57 AM, Joachim Tuchel wrote: >>>>> >>>>> interesting.... >>>>> >>>>> #isOdd is not Smalltalk, neither Pharo 10 nor VAST 12 >>>understands this >>>>> message ;-) >>>>> >>>>> If I evaluate your snippet -replacing #isOdd for #odd, I get >an >>>empty >>>>> Collection. The divisible by 4 things is somewhat >interesting, >>>because (1 >>>>> to: 100 by: 4) is amn interval 1,5,9 etc.... ;-) >>>>> >>>>> So chatGPT is probably a newbie in Smalltalk. Interesting is >>>how logical >>>>> the explanation sounds, you're tempted to believe it... >>>>> >>>>> >>>>> Joachim >>>>> >>>>> >>>>> >>>>> Am 15.03.23 um 10:45 schrieb giorgio ferraris: >>>>> >>>>> it's interesting to see how the answer changes with only a >>>small change >>>>> on the question: >>>>> >>>>> (1 to: 100 by: 4) reject: [:i | i isOdd] >>>>> >>>>> gives: >>>>> >>>>> The code creates a collection of numbers from 1 to 100, >>>incrementing by 4 >>>>> at each step using the to:by: message. It then applies the >>>reject: message >>>>> to this collection, which returns a new collection containing >>>all elements >>>>> for which the given block evaluates to false. >>>>> >>>>> In this case, the block is [i | i isOdd], which checks >whether >>>a given >>>>> number i is odd using the isOdd method. So the result of the >>>entire >>>>> expression is a collection of all even numbers between 1 and >>>100 that are >>>>> divisible by 4. >>>>> >>>>> ciao >>>>> >>>>> giorgio >>>>> >>>>> >>>>> On Wed, Mar 15, 2023 at 8:07 AM Tim Mackinnon ><tim@testit.works> >>>>> <tim@testit.works> wrote: >>>>> >>>>> Of course everyone is talking about ChatGPT, and I saw a >post to >>>>> ExplainCode and didn’t expect it to understand Smalltalk, so >>>was impressed >>>>> that it does! Presumably the Tonel Smalltalk code on GitHub >has >>>helped >>>>> along with numerous articles? >>>>> >>>>> Try it out: https://whatdoesthiscodedo.com/ >>>>> >>>>> I gave this simple example: (1 to: 100 by: 3) reject: [:i | i >>>isOdd] >>>>> >>>>> And got this (neat eh): >>>>> >>>>> The code above is written in the programming language >Smalltalk. >>>>> >>>>> It creates a collection of numbers from 1 to 100, >incrementing >>>by 3 at >>>>> each step. This is done using the "to:by:" construct where >the >>>starting >>>>> value is 1, the ending value is 100, and the increment is 3. >>>>> >>>>> The resulting collection is then filtered using the "reject:" >>>method, >>>>> which takes a block of code as an argument. In this case, the >>>block checks >>>>> whether each element (denoted by the variable i) is odd by >>>using the >>>>> "isOdd" method. If an element is odd, it will be rejected >from >>>the >>>>> collection. >>>>> >>>>> So, the final result is a collection of all even numbers >from 1 >>>to 100, >>>>> incrementing by 3 at each step. >>>>> Share this explanation with colleagues: >>>whatdoesthiscodedo.com/g/a0ec56e >>>>> >>>>> -- >>>>> >>>>> ------------------------------------------------------------- >--- >>>------- >>>>> Objektfabrik Joachim Tuchel >>>mailto:jtuc...@objektfabrik.de <jtuc...@objektfabrik.de> >>>>> Fliederweg 1 >>>http://www.objektfabrik.de >>>>> D-71640 Ludwigsburg >>>http://joachimtuchel.wordpress.com >>>>> Telefon: +49 7141 56 10 86 0 Fax: +49 7141 >>>56 10 86 1 >>>>> >>>>> >>>>> >>>>> >>> >>>-- >>>Christopher Fuhrman, P.Eng., PhD >>> >>>*Professeur au Département de génie logiciel et des technologies >de >>>l'informationÉTS (École de technologie supérieure)* >>> >>>http://profs.etsmtl.ca/cfuhrman >>>+1 514 396 8638 >>><https://www.etsmtl.ca/activites-et-services-aux- >>>etudiants/services-aux-etudiants/aide-psychologique> >>> Je fais partie du réseau des Sentinelles >>><https://www.etsmtl.ca/activites-et-services-aux- >>>etudiants/services-aux-etudiants/aide-psychologique> >>> ÉTS >>> >>>*L'ÉTS est une constituante de l'Université du Québec*
