Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.lang.prolog > #14408

Or ChatGPT is better in Program Verification (Re: John Sowa is close with RNT)

Show all headers | View raw

Hi,

Or ChatGPT is better in Program Verification. Means
the "Fine Tuning" that ChatGPT has to deal with
Queries that include programming language fragments

or tasks, and the generation of programming language
fragements or tasks is practically already better and
expected to improve even more:

- Not an Architecture based on a Symbolic HOL language
   and backend to extract Programming Languages.

- Not an Architecture that only supports some academic
   "pure" Languages like ML, Haskell, Scala, etc..

- Rather an Architecture where one can directly interacts
   in the Programming Languages, similar like it deals with
   Natural Language, Pipe Dream Onologies are not seen.

- Rather an Architecture tat can deal with "impure"
   Languages such as JavaScript, Python, etc.. that are
   imperative and have side effects.

- What else?

Bye

Mild Shock schrieb:
> Hi,
> 
> Interestingly, John Sowa is close with RNT to how
> ChatGPT works . In his LLMs bashing videos, John Sowa
> repeatedly showed brain models in his slides that
> 
> come from Sydney Lamb:
> 
> Relational Network Theory (RNT), also known as
> Neurocognitive Linguistics (NCL) and formerly as
> Stratificational Linguistics or Cognitive-Stratificational
> Linguistics, is a connectionist theoretical framework in
> linguistics primarily developed by Sydney Lamb which
> aims to integrate theoretical linguistics with neuroanatomy.
> https://en.wikipedia.org/wiki/Relational_Network_Theory
> 
> You can ask ChatGPT and ChatGPT will tell you
> what parallels it sees between LLM and RNT.
> 
> Bye
> 
> P.S.: Here is what ChatGPT tells me about LLM and RNT
> as a summary, but the answer itself was much bigger:
> 
> While there are shared aspects, particularly the
> emphasis on relational dynamics, ChatGPT models are
> not explicitly designed with RNT principles. Instead,
> they indirectly align with RNT through their ability
> to encode and use relationships learned from data.
> However, GPT’s probabilistic approach and reliance
> on large-scale data contrast with the more structured
> and theory-driven nature of RNT.
> https://chatgpt.com/share/67818fb7-a788-8013-9cfe-93b3972c8114
> 
> I would also put the answer into perspective if one
> include RAG. So with Retrieval Augmented Generation
> 
> things look completely different again.
> 
> Mild Shock schrieb:
>> Hi,
>>
>>  > Subject: An Isabelle Foundation?
>>  > Date: Fri, 10 Jan 2025 14:16:33 +0000
>>  > From: Lawrence Paulson via isabelle-dev
>>  > Some of us have been talking about how to keep
>> things going after the recent retirement of Tobias and
>> myself and the withdrawal of resources from Munich.
>> I've even heard a suggestion that Isabelle would not
>> be able to survive for much longer.
>> https://en.wikipedia.org/wiki/Isabelle_%28proof_assistant%29
>>
>> No more money for symbolic AI? LoL
>>
>> Maybe suplanted by Lean (proof assistant) and my
>> speculation maybe re-orientation to keep up with
>> Hybrid methods, such as found in ChatGPT.
>> https://en.wikipedia.org/wiki/Lean_%28proof_assistant%29
>>
>> Bye
> 

Back to comp.lang.prolog | Previous | Next — Previous in thread | Next in thread | Find similar | Unroll thread

Thread

Traditions die: Another one bites the Dust Mild Shock <janburse@fastmail.fm> - 2025-01-10 22:36 +0100
  John Sowa is close with RNT (Re: Traditions die: Another one bites the Dust) Mild Shock <janburse@fastmail.fm> - 2025-01-10 22:44 +0100
    Or ChatGPT is better in Program Verification (Re: John Sowa is close with RNT) Mild Shock <janburse@fastmail.fm> - 2025-01-10 22:59 +0100
      What is Argonne doing now? (Re: Or ChatGPT is better in Program Verification) Mild Shock <janburse@fastmail.fm> - 2025-01-11 02:11 +0100
        Lemanicus at EPFL Computer Museum [compare to NVIDIA H100 NVL] Mild Shock <janburse@fastmail.fm> - 2025-01-11 02:50 +0100

csiph-web