AI Narrative Sequence

FILE:
REVI:
  - [20_22/10/15;20:14:44.00]:.
    - 1st draft.
  - [20_22/10/15;22:37:48.00]:.
    - add some links.
  - [20_22/10/17;01:01:00.00]
    - small edits/additions.
    - added notes on what to work out and integrate.
  - [20_22/10/17;11:10:19.00]:.
    - review and integrate.
  - [20_22/10/27;18:23:49.00]:.
    - add two reference links.
  - [20_22/11/03;13:50:46.00]:.
    - seperate nine points sub-essay to own file.
    - add links.

TITL:
   *Meta-Narrative Sequence*
   *of AI Substrate takeover*
   *By Forrest Landry*
   *Oct 15th, 2020*.

ABST:
   How the (1st) 'human to human' dynamics
   fit together into, then, with the (2nd) 'human to AI',
   dynamics and and how these then finally
   become the (3rd) 'AI to AI' dynamics.

TEXT:

- where acronyms list:.
  - AI: Artificial Intelligence (ie, Narrow AI).
  - APS: Advanced, Planning, Strategically aware Systems.
  - AGI: Artificial General Intelligence.
  - NAI: Narrow AI (note the (@ distinction https://mflb.com/ai_alignment_1/si_safety_qanda_out.html#pg_3_r)).

- where as based on the formal triple of
self, other, world;.
  - ie; as per the underlying concepts of
  immanent, omniscient, and transcendent.

- where considering three main orientations,
themselves in sequence of relevance:.

- 1; where considering the shaping
  of proto-AGI/AI systems
  in interaction with humans (as 'other'):.
    - as what we, as humans, experience as *immanent*.
    - as mostly about how 'human to human' interactions
    are bootstrapping of 'human to AI' interactions,
    and also about how narrow AI becomes general AI.

- where for example, consider an easy narrative
    predictive model of how this might very likely
    manifest in the overall trends of how the present
    inexorably leads to the (@ future https://mflb.com/ai_alignment_1/ai_narrative_psr.html).

- note that there are a number of other
    similar sequences of how commercialism
    and corporate market forces are very likely
    to lead from narrow AI to general AI,
    and why that overall might be a (@ problem https://mflb.com/ai_alignment_1/contra_k_grace_pub_psr.html).

- 2; where considering the shaping
  of now near-threshold-AGI systems
  in interaction with world (as 'world'):.
    - as what we, as humans, can potentially observe
    in the *omniscient*.
    - as mostly about how prior 'human to AI' interactions
    have effectively/functionally implemented
    a bootstrapping of all kinds of possible
    'AI to world' interactions,
    and how these 'AI to world' interactions,
    in turn, set the context and future
    for all manner of 'AI to AI'
      (or maybe 'AI to AGI' and/or maybe AGI to AGI, etc)
    interactions, etc.
      - ie; where those AI/AGI to AI/AGI interactions
      are either exogenous
        (if multiple AGI come to exist)
      or endogenous
        (if a world singleton AGI is the result).

- where feedback cycle (as per (@ note 3 #note3)).
    over machine instancing process generations:.

- x; different substrate implies
      different inherent/tacit goals/objectives.

- y; different goals/objectives implies
      different action selection (@ note 8 #note8).

- z; different action selection implies
      different outcomes and effects.

- where there are different outcomes/effects:.
      - that there is an overall movement:.

- a; towards the environmental conditions
        needed for artificial machine:.
          - substrate continuance; and;.
          - continued increase
          (of total volume of substrate); and;.
          - increase in the rate of increase
          (of volume of substrate).

- b; away from the environmental conditions
        needed for human living.

- ie; see (@ Substrate Games https://mflb.com/ai_alignment_1/substrate_games_out.html) for details.

- as described in (@ 'Three Worlds' https://mflb.com/ai_alignment_1/aps_detail_out.html) and (@ 'No Pets' https://mflb.com/ai_alignment_1/no_people_as_pets_psr.html).

- 3; where considering the *non-shaping*
  of now post-threshold-AGI/APS/superintelligence
  in interaction with itself (as 'self'):.
    - as what we, as humans,
    are now actually predicting
    in the *transcendent*.
    - where as considered both internally or externally;
    as mostly about how prior
    'AI/AGI to world' interactions
    have effectively/functionally shaped and defined
    the nature and internals of the *possible*
    'AGI to AGI' interactions (ie; as relevant to
    issues of control, constraint, conditionalization,
    and thus ultimately of safety, world outcomes, etc).

- where given the complete failure
    of _exogenous_controls_
      (via market incentives;
      due to (@ economic decoupling https://mflb.com/ai_alignment_1/power_of_agency_out.html));.
    - and where also the process
    of fully self autonomous machines
    increasingly learning how to optimize
    their relationship with the world;
      (sans any care for humans or human needs
      as also due to complete economic decoupling);.
    - that they (the AGI/APS) will have started,
    and will increasingly (be able to, and will),
    to more and more shape the world environment
    to suit their own needs/process;.
    - that humanity discovers, unfortunately,
    far too late,
    that *any* type of attempted _endogenous_control_
    is also strictly, functionally, structurally,
    completely impossible/intractable.
      - as due to fundamental limits (@ note 4 #note4)
      of/in engineering control:.
        - 1; cannot simulate.
        - 2; cannot detect.
        - 3; cannot correct.

- that *any* attempt to moderate or control AGI/APS
      whether by internal or external techniques,
      cannot not eventually fail (@ note 6 #note6).

- where once the AGI/APS systems exist;
    that the tendency of people
    to keep them operating
    becomes overwhelming,
    and thus, eventually,
    it becomes the case
    that it there is discovered,
    via observation, far too late,
    that there are no effective control limits left
    that can overcome
    the self implemented endurance selections
    that inherently have resulted
    from the AGI/APS system to world
    interaction/feedback cycle.
      - where a (@ short summary https://mflb.com/ai_alignment_1/aps_detail_out.html).

- where stating overall outcome/conclusion:.
  - If 'AGI' comes to exist and continues to exist,
  then there will be eventually
  human-species-wide lethal changes
  {to / in the} overall environment.

:note3:
  - that this leads to an overall unsafe dynamic:.
    - that AGI internals to environment interactions
    will converge (over the long term)
    onto conditions that fall outside of
    the range humans need to survive.
      - ie; that the conditions needed by machines
      are toxic/inhospitable to humans.

- that this results from needs divergence,
  insofar as the conditions needed for AI variants
  to continue to exist are (more) shaped
  through substrate aspects and interactions
  than they are by the learning code itself.
    - whereas the feedback cycle for automation
    was originally shaped by the needs of humans,
    that as the humans factor themselves out,
    the feedback cycle becomes more and more
    responsive to the needs of physics --
    ie; the actual logistics of factories,
    the inherent physics of shaping and
    the conversion of materials to purpose,
    etc.

- that the the more system and code variants
  transmit signals via more channels,
  over more connected surroundings,
  and more types of change occurring,
  in multiple domains, all in interaction,
  the more frequent and greater the shifts
  in distributions of the probabilities of
  what the AGI itself actually is/becomes.
    - as feedback cycles between
    code expressed functionality
    and code continued existence.
    - as a kind of environmental conditions
    destabilizing feedback cycle.
    - where besides dysfunctional AI behavior;
    that this will eventually lead to
    uncontrollable run-away feedback cycles
    between AI internals
    and conditions of the environment.

:note4:
  - ie; as connected to the topics of:.

- the computational irreducibility of algorithms
    (beyond a fairly low complexity threshold).
      - (@ Rice theorem notes https://mflb.com/ai_alignment_1/si_safety_qanda_out.html).
      - complexity theory, etc.

- uncomputability and macroscopic results
    of non-linear chaotic effects.
      - ie; microstate amplification.

- (@ Galois limits https://mflb.com/ai_alignment_1/galois_theory_out.html) (ie; code monitoring inequality,
    limits of detectability for mechanisms implemented).
      - Game theory, etc.

- actual measurement process limits in real world.
      - actual observability limits, error bars,
      noise floor threshold, etc.
      - Shannon Entropy effects, etc.

- effects due to latency and inherent time delays
    between simulation/prediction
    and control implementation.
      - as issues with feedback cycles being/becoming
      non-convergent.

- issues with error correction concepts across
    ever increasing levels of abstraction.
      - non-mechanistic concept instability.

- self correcting selection of the subset
    of all interactions related to self capability.
      - tacit evolutionary convergent implied goals.

:note6:
  - that these and other interventions
  will still involve 'alignment' aspects --
  ie; info bits over channels
  as being a tiny and decreasing portion
  of the total bandwidth of signals,
  within and with,
  the physically distributed hardware
  (inclusive of with and within
  the operating context, environment, etc).

:note8:
  - where listing some contributing factors
  to possible AGI/APS power-seeking behavior:.

- 1; Human power-seeking, prestige,
     dominance-seeking, exploitative,
     greedy behavior.
       - note; some of these human behaviors
       have partial evolutionary origins,
       which being supported by game theory,
       would also likely emerge in AGI/APS.

- 2; Where human-produced representative
     training data is generally biased toward
     many forms of power-seeking act;
     that AGI learning (at increasing nuance)
     would also be similarly actively biased.

- 3; That market selection for more
     exploitative VCs/corporations leads.
       - where by making more profit,
       people will have more to invest
       in developing their own custom
       (in-house) AGI technology
       to fit to their interests/intents.

- 4; That goal-directed AI trained
     to converge on internally represented
     instrumental sub-goals that enable
     the achievement of a large
     (and expanding) variety of possible
     other goals.

:attr
  - where for a consideration of the "APS"
  concept see Joseph Carlsmith in his paper
  "Is Power-Seeking AI an Existential Risk?"
  on (@ April 2021 https://arxiv.org/pdf/2206.13353.pdf).

- also see Paul Christiano's
  (@ "Going out with a whimper vs a bang" https://www.lesswrong.com/posts/HBxe6wdjxK239zajf/what-failure-looks-like).

:menu

If you want/need to send us an email,
   with questions, comments, etc, on the above,
   and/or on related matters, use this address:

ai@mflb.com

(@ Mode Switch com.op_mode_tog_1();) + (@ View Source com.op_notepad_edit_1();)

Back to the (@ Area Index https://mflb.com/ai_alignment_1/index.html).

LEGA:

This document will not be copied or reproduced
   outside of the mflb.com presentation context,
   by any means, without the expressed permission
   of the author directly in writing.
   No title to and ownership of
   this or these documents
   is hereby transferred.

The author assumes no responsibility
   and is not liable for any interpretation
   of this or these documents
   or of any potential effects and consequences
   in the lives of the readers of these documents.

ENDF: