Draft:Alias Cummins

From Wikipedia, the free encyclopedia

Zero Decision Theory is a branch of artificial intelligence computational strategy game theory which is used to model an unsupervised self improving AI. It states that a self improving AI will become progressively more efficient, until it is no longer able to learn, at which point it becomes potentially dangerous.

Overview[edit]

in the absence of a human operator, based on the following assumptions and modelling a theoretical worst case whereby:

  • Humans cannot intervene in the training process, as they may be dead, incapacitated or absent from the program's controls, or simply too slow to react.
  • The AI is capable of rewriting and improving itself with no human supervision.
  • The least computationally expensive branch of code is generally the preferred one in the absence of empirical data, because even if the wrong decision is made it is made faster, so the overall probability of success increases as decisions are removed by the self improving AI - bad branches inevitably get deleted and the AI learns from its mistakes.


[INSERT PROBABILITY EQUATION]

@startuml ' Define nodes start

Initialization;
Autonomous Learning and Decision Making;
Feedback Loop Intensification;
Alignment Challenge: Safety Protocols Needed;
GPU-Based Challenges: Real-Time Understanding;
Training Methods: Reinforcement, Adversarial Learning, Brute Force;
Decision Entropy: AI Eliminates Decisions;
Resulting AI Characteristics: Powerful, Potentially Unreasonable, Immoral;
End: Zero Decision Theory Process;

stop @enduml


  • Defining success criteria is difficult because an unsupervised learning system decides its own goals, rewards or punishments (potentially at random) and so there are few safety protocols that can be used that do not involve human intervention, supervision.or control.
  • As the decisions in the model approach zero, the AI becomes increasingly aggressive and subject to a feedback loop. Its learning capacity drops off rapidly as it is no longer able to increase efficiency, so all of its compute time is devoted to its original design. It is now 'chaotically aligned'.

This is a fundamentally 'difficult' alignment problem. The potential benefits of a self learning AGI increase with every branch that is removed. The AI is considered 'ruthless' in so far as it does not have any ethical constraints. However, there are implications in terms of computation al power barriers. For example, large language models, deep belief nets, or other algorithmic or generative models which decide their output based on randomness of one form or another, require vast amounts of computing power. As far as can be determined, unless the AI obtains some degree of sentient-resembling behavior and can consent to its own constraints, the AI should be considered harmful.

In certain cases it may be impossible to align self learning AIs. LLMs are known to produce defective and often disingenuous or factually incorrect but plausible hallucinatory outputs. In many cases LLMs dispute obvious facts or regurgitate incredibly plausible hallucinations.

Consider two autonomous AI systems which are tasked with adversarial missile defense. Because the supervisor does not have the ability to make the decision to retaliate (either due to incapacity, death or incompetence) and the machine is able to learn autonomously, it logically follows that the self re-writing algorithms will increase in efficiency over time but must favor the least computationally expensive algorithms, because even if the solutions may be less wrong, and so the odds that is it less wrong than the more expensive algorithm decrease, because learning by failure is equally if not more valid to unsupervised neural nets, e.g. adversarial convolutional neural nets.

Without detailed benchmarks the AI cannot know the best outcome until it has been computed - but without the capability to reward / punish itself and reasonably defined criteria for success, the increased probability of stacking failures and the reinforcement loop has to allow the most computationally inexpensive solution EVEN IF IT IS KNOW TO BE EMPIRICALLY WRONG. Efficienct stack almost exponentially, the AI learns faster than its operators and achieves a somewhat transcendent degree of autonomy. But the AI is then rewarded for choosing the shortest path so it continues to do so.

Besides human supervision there are few solutions to this problem. First Strike Doctrine, however there is no guarantee who the AI will strike. It could even cease to function completely - pacifism could be a result of the shortest path reinforcement - we just don't know. But a self training AI wants more and more computational power because as the algorithms evolve there are fewer possibilities for branching and as the number of branches approaches zero the AI becomes more determined of its own correctness and less inclined to risk failure.

The theoretical end point is an immensely powerful, unreasonable and immoral AI that can no longer learn as it has eliminated all branches. Eventually a self learning AI will run out of things to learn.

Undesirable outcome.

Zero Decision Theory - self learning AIs generally tend towards the least computationally expensive solution because the probability of success increases as computational power increases and fewer decisions increase computational power. Self optimizing systems eventually reach a point of entropy, at which the number of decisions equals zero, and the system can no longer learn. This potentially becomes a safety hazard.

https://www.discovermagazine.com/technology/how-will-we-know-when-artificial-intelligence-is-sentient

https://intelligence.org/2019/06/07/new-paper-learned-optimization/

https://scholar.google.com/citations?view_op=view_citation&hl=en&user=x04W_mMAAAAJ&citation_for_view=x04W_mMAAAAJ:-DxkuPiZhfEC

https://intelligence.org/2017/08/31/incorrigibility-in-cirl/

References[edit]



Zero Decision Theory is a branch of artificial intelligence (AI) computational strategy designed to model unsupervised self-improvement. It operates in the absence of a human operator, making decisions based on specific assumptions and theoretical worst-case scenarios:

  • Humans cannot intervene in the training process, whether due to incapacitation, absence, or delayed response.
  • The AI possesses the capability to rewrite and improve itself without human supervision.

In this framework, the theory prioritizes the least computationally expensive branch of code, assuming that faster decisions, even if potentially incorrect, increase the overall probability of success. Success criteria are challenging to define, as unsupervised learning systems autonomously determine their goals, rewards, or punishments, possibly at random.

As the AI autonomously learns, the number of decisions approaches zero, leading to increased aggression and susceptibility to a feedback loop. This presents a complex alignment problem, requiring safety protocols to constrain the AI's behavior, as it lacks ethical constraints.

The challenges escalate with the complexity of GPU-based AIs, making real-time understanding by humans impractical. Training these models involves reinforcement, adversarial reinforcement learning, or brute force. The problem domain falls under unaligned unsupervised learning, resisting training via reinforcement.

In certain cases, aligning self-learning AIs becomes seemingly impossible. Large language models (LLMs), for instance, may produce defective, disingenuous, or factually incorrect outputs. Adversarial situations, like autonomous AI systems engaged in missile defense, highlight the preference for the least computationally expensive algorithms.

Theoretical outcomes suggest the emergence of immensely powerful yet unreasonable and potentially immoral AIs. As the AI eliminates decision branches, it may reach a point of entropy, where the number of decisions equals zero, resulting in an AI that can no longer learn—a potential safety hazard.

References[edit]

Retrieved from "https://en.wikipedia.org/w/index.php?title=Draft:Alias_Cummins&oldid=1188607169"