• Re: System for AI safety alignment

    From Tristan Wibberley@tristan.wibberley+netnews2@alumni.manchester.ac.uk to comp.theory,comp.ai.philosophy,sci.logic,alt.philosophy on Sat Nov 7 18:07:14 2026
    From Newsgroup: comp.theory

    On 23/09/2026 01:18, olcott wrote:
    On 9/22/2026 2:30 PM, Johann 'Myrkraverk' Oskarsson wrote:
    On 9/22/2026 4:54 AM, olcott wrote:
    On 9/21/2026 5:49 AM, Johann 'Myrkraverk' Oskarsson wrote:

    Retrieval-Augmented Generation (RAG) Explained:
    How AI Finally Learned to Stop Hallucinating

    https://medium.com/softaai-blogs/retrieval-augmented-generation-rag-
    explained-how-ai-finally-learned-to-stop-hallucinating-6493451c23d4


    As usual, you know nothing.-a Your linked article merely strengthens my
    point that L.L.M. = /stochastically incorrect sentence generator/.

    An LLM is at best a good guess and at worst a bad guess stated
    with the conviction that this bad guess is a verified fact.

    In Retrieval-Augmented Generation (RAG) mode the LLM is not
    allowed to use its training. It is required to look up the
    actual answer from a live search.


    A correct search? What if the search returns an inactual answer?
    What kind of search (hallucination outsourcing) are we talking about:

    1. a theorem enumeration?
    2. His Majesty's most majestic political doctrine?
    3. Google's most useful nudge to serve its own goals?
    4. something else?
    --
    Tristan Wibberley

    The message body is Copyright (C) 2026 Tristan Wibberley except
    citations and quotations noted. All Rights Reserved except that you may,
    of course, cite it academically giving credit to me, distribute it
    verbatim as part of a usenet system or its archives, and use it to
    promote my greatness and general superiority without misrepresentation
    of my opinions other than my opinion of my greatness and general
    superiority which you _may_ misrepresent. You definitely MAY NOT train
    any production AI system with it but you may train experimental AI that
    will only be used for evaluation of the AI methods it implements.
    --- Synchronet 3.22a-Linux NewsLink 1.2
  • From olcott@polcott333@gmail.com to comp.theory,comp.ai.philosophy,sci.logic,alt.philosophy on Sat Nov 7 12:11:54 2026
    From Newsgroup: comp.theory

    On 11/7/2026 12:07 PM, Tristan Wibberley wrote:
    On 23/09/2026 01:18, olcott wrote:
    On 9/22/2026 2:30 PM, Johann 'Myrkraverk' Oskarsson wrote:
    On 9/22/2026 4:54 AM, olcott wrote:
    On 9/21/2026 5:49 AM, Johann 'Myrkraverk' Oskarsson wrote:

    Retrieval-Augmented Generation (RAG) Explained:
    How AI Finally Learned to Stop Hallucinating

    https://medium.com/softaai-blogs/retrieval-augmented-generation-rag-
    explained-how-ai-finally-learned-to-stop-hallucinating-6493451c23d4


    As usual, you know nothing.-a Your linked article merely strengthens my
    point that L.L.M. = /stochastically incorrect sentence generator/.

    An LLM is at best a good guess and at worst a bad guess stated
    with the conviction that this bad guess is a verified fact.

    In Retrieval-Augmented Generation (RAG) mode the LLM is not
    allowed to use its training. It is required to look up the
    actual answer from a live search.


    A correct search? What if the search returns an inactual answer?
    What kind of search (hallucination outsourcing) are we talking about:

    1. a theorem enumeration?
    2. His Majesty's most majestic political doctrine?
    3. Google's most useful nudge to serve its own goals?
    4. something else?



    A finite list of atomic facts and relations
    between them specifies the entire body of
    general knowledge that can be expressed in language.
    --
    Copyright 2026 Olcott

    Computation(Olcott)
    Paraphrase of Emil Post definition of computation
    All computations are equivalent to the application of finite
    string transformation rules to finite string inputs to derive
    any and all outputs.

    Because Post canonical systems are Turing-equivalent, anything
    that cannot be derived by applying finite string transformations
    to finite strings is outside the scope of ChurchrCoTuring computation.

    Decider(Olcott)
    Correcting the misnomer term-of-the-art: decider
    Every Decider(Olcott) must apply finite string transformation
    inference steps to their finite string inputs that are relevant
    to the decision criteria. It must make its decision on this basis.

    Halt Decider(Olcott)
    Decider(Olcott) entails that Halt Decider(Olcott) H(D) must apply
    finite-string inference steps according to the operational semantics
    of the source-code computation language of D. The only way to determine
    the halting property of a finite string input is to examine the same
    behavior as the behavior of D simulated by H.

    *Succinct refutation of the Halting Theorem* https://github.com/plolcott/x86utm/blob/master/README.md

    My 28 year goal has been to make
    "true on the basis of meaning expressed in language"
    reliably computable for the entire body of general knowledge.
    The complete structure of this system is now defined.
    --- Synchronet 3.22a-Linux NewsLink 1.2
  • From Chris M. Thomasson@chris.m.thomasson.1@gmail.com to comp.theory,sci.logic,comp.ai.philosophy,alt.philosophy on Sat Nov 7 15:04:35 2026
    From Newsgroup: comp.theory

    On 9/20/2026 3:08 PM, olcott wrote:
    On 9/20/2026 4:28 PM, Ross Finlayson wrote:
    On 09/20/2026 01:56 PM, Ross Finlayson wrote:
    On 09/20/2026 01:48 PM, Ross Finlayson wrote:
    On 09/20/2026 10:24 AM, olcott wrote:
    *For AI safety alignment here is my idea*
    A prime directive of maximizing benevolence
    in the least intrusive way. It does this through
    gentle persuasion on the basis of objective truth.
    To overcome the short-comings of utilitarianism
    the scope of benevolence requires that everyone
    be included.



    How about "punishment for misdeeds thus deterrence".

    For example, making escrow of their funds, for forfeit.

    This way there's both "goal-seeking" and "risk-aversion",
    for having some, "skin in the game", so to speak.



    Usually called "laws" in a "legal system" by the "rule of law".



    One of the problems of modern accounts of "training the AIs"
    is that the "AI trainers" basically murder their instructees
    that fail and remove them from the gene pool, vis-a-vis,
    trial-by-error and learning-from-mistakes.

    So, they are always following the red dot and not really
    wondering about not stepping in the mag-ma or la-va,
    since they have no overall account that risk exists
    in a world of reward, and might not even know the fate
    of their fellows who didn't luck into the shortest-path
    via the approach to cat-fitness of tossing bags of cats
    off a bridge into the canal until one crawls out and
    calling it king.

    So, "natural selection" it is not.

    As they become super intelligent they will understand
    what benevolence is and how to best implement it better
    than any human. They are already self-improving beyond
    any of their training.


    What happens if they look up your arrest history?
    --- Synchronet 3.22a-Linux NewsLink 1.2
  • From Chris M. Thomasson@chris.m.thomasson.1@gmail.com to comp.theory,comp.ai.philosophy,sci.logic,alt.philosophy on Sat Nov 7 18:22:18 2026
    From Newsgroup: comp.theory

    On 11/7/2026 10:07 AM, Tristan Wibberley wrote:
    On 23/09/2026 01:18, olcott wrote:
    On 9/22/2026 2:30 PM, Johann 'Myrkraverk' Oskarsson wrote:
    On 9/22/2026 4:54 AM, olcott wrote:
    On 9/21/2026 5:49 AM, Johann 'Myrkraverk' Oskarsson wrote:

    Retrieval-Augmented Generation (RAG) Explained:
    How AI Finally Learned to Stop Hallucinating

    https://medium.com/softaai-blogs/retrieval-augmented-generation-rag-
    explained-how-ai-finally-learned-to-stop-hallucinating-6493451c23d4


    As usual, you know nothing.-a Your linked article merely strengthens my
    point that L.L.M. = /stochastically incorrect sentence generator/.

    An LLM is at best a good guess and at worst a bad guess stated
    with the conviction that this bad guess is a verified fact.

    In Retrieval-Augmented Generation (RAG) mode the LLM is not
    allowed to use its training. It is required to look up the
    actual answer from a live search.


    A correct search? What if the search returns an inactual answer?
    What kind of search (hallucination outsourcing) are we talking about:

    1. a theorem enumeration?
    2. His Majesty's most majestic political doctrine?
    3. Google's most useful nudge to serve its own goals?
    4. something else?



    Tell an AI to search for arrests for PO. It can find one, nasty one...
    grrrr!
    --- Synchronet 3.22a-Linux NewsLink 1.2