Grazie del commento Rob. Io non sono un cognitivista quindi mi esprimo un'opinione consapevole che possa essere estremamente ingenua. Trovo il confronto con una teoria della mente molto fuorviante, dal momento che le menti che conosco sono menti che posseggono una semantica mentre ChatGPT possiede solo capacità sintattiche approssimando quelle semantiche mediante la statistica. Per non parlare poi degli aspetti di consapevolezza... Ciao, Enrico Per non parlare poi Il 15/03/2023 14:26, Roberto Dolci ha scritto:
Ciao Enrico,
le principali differenze tra il fanciullo e ChatGPT, di cui notiamo comunque essere il primo robot a raggiungere la performance di un 9nne sulla scala di teoria della mente (era 3enne solo a novembre scorso):
1. il consumo energetico. Il nostro cervello non arriva ad una lampadina, mentre per ogni richiesta di ChatGP consumiamo una quantita' immonda di energia.
2. la creazione. Il bambino crea la spiegazione, e poi la testa immediatamente, validandola o creadone un altra. ChatGPT parte dallo scibile disponibile per inferire e quindi imparare, ovviamente non crea mai nulla di effettivamente nuovo.
3. la plasticita' degli organoidi cerebrali rispetto ai chip sotto ChatGPT e l'utilizzo della chimica nei processi cerebrali. La gestione della "potatura" delle memorie inutili nel nostro cervello e' ordini di grandezza piu' efficiente ed efficace di quanto disponibile con AI. Su Zafferano di sabato
Un abbraccio Roberto ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ *Da:* nexa <nexa-bounces@server-nexa.polito.it> per conto di Enrico Nardelli <nardelli@mat.uniroma2.it> *Inviato:* martedì 14 marzo 2023 17:30:08 *A:* Nexa *Oggetto:* [nexa] a proposito dell'articolo di Chomsky sul NYT
Si tratta di questo articolo: https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html (mi scuso qualora se ne fosse già discusso in lista)
Ne stavo discutendo con colleghi di vari paesi in un'altra mailing list e in risposta ad alcune critiche fatte all'articolo ho scritto quanto riporto sotto (scusatemi se lo lascio in inglese ma siccome coinvolge l'uso di ChatGPT nell'analisi di un testo in inglese traducendolo diventava un'altra cosa)
...
I find your viewpoint highly interesting, even if I somehow disagree on some of your criticism to Chomsky's article.
I find an analogy between the way ChatGPT learned to master the language in comparison to human beings and how Alpha Zero learned to master board games in comparison to human beings.
The deep-learning based approach exploited in both cases (almost) all available data, either through their availability on the Internet (in the case of the language) or through sophisticated systematic search through the strategy space (in the case of board games).
Human beings become master (in both fields) leveraging a much smaller information space: the number of examples a child is exposed to while learning language is, just by considering time and space constraints, much smaller than what ChatGPT used. I do not know of any estimation of this (which would be an interesting thing to do) but I would say the size is inferior of some orders of magnitude (which to me implies the "language learning algorithm" inside human brain is exponentially more efficient then those used in DL - but that's a different story).
Moreover human beings know the "meaning" of their action and are aware of the fact that they know such meaning.
(See the note at the end supporting my statements.)
This appears to me to be an important difference between the two approaches and I think is the reason why Chomsky wrote that ChatGTP is not going to teach us anything useful about how human beings learn and use language.
One more interesting discussion of the Chomsky viewpoint, this time defending it, I suggest (in case you've not already read it) is the piece of Gary Marcus https://garymarcus.substack.com/p/caricaturing-noam-chomsky
Best, Enrico
NOTE: To test my statements I asked ChatGPT to analyse a paragraph of my mail. You can read the prompt and the answer further below.
The answer is at best confusing.
ChatGPT fails to identify the text has 2 sentences and not just 1.
It does a wrong analysis of the first sentence, which contain two main clauses and not one main and one subordinate.
That's particularly interesting since right after the wrong analysis it "explains" that "a main clause is an independent clause that stands alone as a complete sentence". So, it writes that it "knows" what a main clause is right after having shown it doesn't know.
And the rest contains a number of additional errors.
So, even if the examples ChatGPT has been exposed to surely contain many descriptions of grammar rules, it has not been able to get any understanding of them. A secondary high school student, having received much less data and much less training is able to do a much better job.
CVD :-)
PROMPT Please perform the analysis of clause and sentence structure of the text enclosed below between «»: «Human beings become master (in both fields) leveraging a much smaller information space: the number of examples a child is exposed to while learning language is, just by considering time and space constraints, much smaller than what ChatGPT used. I do not know of any estimation of this (which would be an interesting thing to do) but I would say the size is inferior of some orders of magnitude (which to me implies the "language learning algorithm" inside human brain is exponentially more efficient then those used in DL - but that's a different story).»
ANSWER (ChatGPT release of Feb.13 - tested on Mar.14th) The text contains one sentence:
"Human beings become master (in both fields) leveraging a much smaller information space: the number of examples a child is exposed to while learning language is, just by considering time and space constraints, much smaller than what ChatGPT used. I do not know of any estimation of this (which would be an interesting thing to do) but I would say the size is inferior of some orders of magnitude (which to me implies the "language learning algorithm" inside human brain is exponentially more efficient then those used in DL - but that's a different story)."
The sentence contains two clauses:
Main clause: "Human beings become master leveraging a much smaller information space" Subordinate clause: "the number of examples a child is exposed to while learning language is, just by considering time and space constraints, much smaller than what ChatGPT used"
The main clause is an independent clause that stands alone as a complete sentence. It contains a subject ("human beings") and a predicate ("become master leveraging a much smaller information space").
The subordinate clause is a dependent clause that cannot stand alone as a complete sentence. It provides additional information to the main clause by describing the size of the information space that humans use to become masters in both fields. The subordinate clause contains a subject ("the number of examples a child is exposed to while learning language") and a predicate ("is much smaller than what ChatGPT used"). The phrase "just by considering time and space constraints" acts as an adverbial modifier that further clarifies the subordinate clause.
The text also contains two embedded clauses:
"which would be an interesting thing to do" is an embedded clause that acts as an object of the verb "know" in the main clause. "which to me implies the 'language learning algorithm' inside human brain is exponentially more efficient then those used in DL - but that's a different story" is an embedded clause that acts as the subject of the verb "is" in the subordinate clause.
Additionally, the text contains a parenthetical remark: "but that's a different story," which is not syntactically connected to the rest of the sentence.
END OF CHATGPT ANSWER
--
-- EN
https://www.hoepli.it/libro/la-rivoluzione-informatica/9788896069516.html ====================================================== Prof. Enrico Nardelli Presidente di "Informatics Europe" Direttore del Laboratorio Nazionale "Informatica e Scuola" del CINI Dipartimento di Matematica - Università di Roma "Tor Vergata" Via della Ricerca Scientifica snc - 00133 Roma home page: https://www.mat.uniroma2.it/~nardelli blog: https://link-and-think.blogspot.it/ tel: +39 06 7259.4204 fax: +39 06 7259.4699 mobile: +39 335 590.2331 e-mail: nardelli@mat.uniroma2.it online meeting: https://blue.meet.garr.it/b/enr-y7f-t0q-ont ======================================================
--
-- -- EN https://www.hoepli.it/libro/la-rivoluzione-informatica/9788896069516.html ====================================================== Prof. Enrico Nardelli Presidente di "Informatics Europe" Direttore del Laboratorio Nazionale "Informatica e Scuola" del CINI Dipartimento di Matematica - Università di Roma "Tor Vergata" Via della Ricerca Scientifica snc - 00133 Roma home page: https://www.mat.uniroma2.it/~nardelli blog: https://link-and-think.blogspot.it/ tel: +39 06 7259.4204 fax: +39 06 7259.4699 mobile: +39 335 590.2331 e-mail: nardelli@mat.uniroma2.it online meeting: https://blue.meet.garr.it/b/enr-y7f-t0q-ont ====================================================== --
