A New Direction in AI Research

Monica Anderson of Syntience

Why are so many things that are so easy for humans so difficult for computers ?

Computers use models
Science use models
people don’t
So AGI should not

Split things into reasoning and understanding
Reasoning -conscious, logic based deliberation
Understanding – subconscious intuition

Conscious reasoning has been overemphasized.

People receive 10 megabits/sec from the eyes and 1 megabits/sec from the skin and short change from elsewhere.
100 bits/second go the conscious layer.

reasoning is a paint thin layer on top of understanding

Introspection fallacy

Wishful thinking
I am a logical thinker
computers are logical
intelligence is logical

all Human Experience

most of experience is mundane (survival skills) (between rational and mystical – very small segments)

no miracles, no reductionist science

Intuitive understanding

Intuition is an algorithm
Invisibile to introspection
but not a mystical process
it is fallible (infallibilty is a selective memory)

Very difficult domains

1. chaotic systems that cannot be predicted long term
– deep complexity
-nonlinear responses
-state (memory)

2. Irreducible systems
-Open systems
-combinatorial explosion
-time variance

3. Amibiguous input data
4. Emergence

bizarre domains where models cannot be built

the world is bizarre
Life is bizarre
People are bizarre
language is bizarre
the mundane is bizarre

Technology holistic duplicates of scanned brains are not models
they will not provide the closed, compact

cannot restrict AGI implementation to the rational world

model free methods

Science creates models

Low level observation ==> Scientific intuition ==> Cross line to rational science with a scientific model

Observation ==> Engineer intuition ==> check preconditions and select a scientific model

model free methods subscientific intuition

low level observation ==>Experience & Intuition ==> Intuitive prediction
(gain from hanging out in pool halls)

Stronger to Weaker models

full scientific models
Partial models
Simulations
Statistics
Non-parametric models (weak model free method)
Pseudomodels (weak model free method)
Trial and Error (weak model free method)

Evolution is a strong model free method

model Free method zoo

Language – consultation, google
Evolution – genetic algorithms
Adaptation – PID controller / MFA controllers
Learning – Table lookup
Recognition – Repeat success
Discovery – Trial and Error

You are not guaranteed success using any of them.

No models given from outside
Gathering experience (learning is common
work everywhere – even in bizarre domains
require no intelligence to use
Disadvantages
– may not provide portable results

Artificial Intuition
* A very powerful model free method
* a fusion of other MFMs
* Provides understanding, learning, saliency, abstractions, semantics, and novelty
* 15000 lines of code

sensory input
recog
patterns

Emergent reductionism

Singularity and Skynet

Opinion and speculation

Intelligence/intuition is for prediction
the limits to prediction accuracy, precision, and time are not technological
limited by the bizarreness of the world
arrival of AGI may increase bizarreness
logic based machines have no advantage in the bizarre world

Logic based godlike infallible AGI is impossible
humans may already be almost as intuitive as anything can ever get
IMO at most we get machines just a little more intuitive than humans
A trillion that are not different

Enforced diversity

Create AGIs with enforced diversity by giving each one a unique education

An ecosystems of intelligences
intelligences may increase but slowly
no single agi to rule than all

RELATED SITES

syntience.com
artificial-intuition.com
monicasmind.com