Towards a Checklist of AGI Implementation – Can a Critic Become a Solutionist?

Janne P. Hukkineninrobotico.com, Helsinki

Describing AGI Agent & Environment

Agency1

  • Spirit: Virtual control/operating system of any agent (cells, animals, humans, families, cities, ecosystems, corporations, nation states, …)
  • Agency: Cybernetic control plant in feedback relation to environment
  • Sentience: Agent discovers itself in and its relationship to the world
    • Consciousness: Agent aware of own attention, able to control it. Creates coherent interpretation. Maintains indexed memory for disambiguation, learning, and reasoning. Mediates knowledge within mind.
  • Self: self-image, 1st person perspective. Content modulated when agent turns intentions to actions. Between discovery of own existence and deconstruction of the self-representations.
  • Emotions: content & expressions are learned, on top of low level bodily valence2
  • Mental models of self & world create subjective reality
  • Generic Reward is enough hypothesis3 & Motivations4

World Knowledge, Representation

  • What do we know about world structure and dynamics?
  • Motor control of body, movement, locomotion
  • Mental models of environment & self21
    • Causal Systems/Networks
  • 3+D physical world (sensory modalities + time)
    • Objects
    • Affordances
  • Social World (human culture), other agents
  • Abstract Concepts
  • Energy
  • Survival
  • Resolution4 of time, space, information channel width, world knowledge, decision making, etc.

Environment

  • Dimension scales have alternatives. Observations : discrete – continuous
    Actions : discrete – continuous
    Time : discrete – continuous
    Dynamics : deterministic – stochastic – chaotic
    Observability : full – partial
    Agency (others’) : single – multiagency
    Uncertainty : certain – uncertain
    Reality : simulated – real-world
  • 3,5,4

Hardware

  • Sensors, Actuators, Signaling
  • Motor Control, multiscale time and space resolutions
  • Embodied Learning: body constrains and modulates learning
  • Implicit Computation by physical & mechanical properties of body

Communication

  • Signaling
  • Signal Combinations
  • Symbols are physical entities on sensory modality; labels for concepts
  • Language as multi-level rule & symbol system: phonology, morphology, universal grammar

Cognitive Capabilities

  • Perception
  • Abstraction, Conceptualization, Objectification
  • Learning
  • Memory (sensory, motor, experiential, episodic, procedural)
  • Mental simulation
  • Reasoning, Planning
  • Navigation
  • Causality: interaction between mental models

Cognitive Architecture of Modules/Agents

  • A set of modules/agents comprise complete AGI agent (society of mind1,6)
  • No sentience, self, consciousness, etc. for sub-modules/sub-agents
  • Divide labor between modules/agents
  • Orchestration
  • Marrian computational levels
    • purpose
    • algorithm, 1 per module/agent
    • implementation/hardware
  • Reward-is-enough framework3

References

(1) Science, Technology & the Future. Joscha Bach - Agency in an Age of Machines, 2022.

(2) Feldman-Barrett, L. How Emotions Are Made: The Secret Life of the Brain; Pan Macmillan, 2017.

(3) Silver, D.; Singh, S.; Precup, D.; Sutton, R. S. Reward Is Enough. Artificial Intelligence 2021, 299, 103535.

(4) Dörner, D.; Güss, C. D. PSI: A Computational Architecture of Cognition, Motivation, and Emotion. Review of General Psychology 2013, 17 (3), 297–317.

(5) Hutter, M. Universal Artificial Intelligence, 2016.

(6) Minsky, M. Society of Mind; Simon; Schuster, 1988.

(7) Rooij, I. van. Psychological Models and Their Distractors. Nature Reviews Psychology 2022, 1 (3), 127–128.

(8) Wang, P. On Defining Artificial Intelligence. Journal of Artificial General Intelligence 2019, 10 (2), 1–37.

(9) Legg, S. Machine Super Intelligence. PhD Thesis, Università della Svizzera italiana, 2008.

Why AGI?

Why Build?

  • Complex system is best understood by modeling it. Building a system reprioritizes and explicates what we don’t understand (mechanisms instead of narratives7)
  • AGI agent needs to be run in the world for alignment testing with world dynamics (aesthetics), which is extrapolated from highest level purposes of civilization1
  • White hat security: Improve security and ethics by trying to break/missuse a working system
  • Not building does not protect us from adversarial and unethical entities using AGI systems against us.

Definition, Criteria

  • Human-level or super-human behavior and adaptation with insufficient knowledge and resources8 in undefined environments and tasks
  • native (system information content) vs. performance intelligence9
  • Computational part of reaching goals adaptively9
  • Hypothesis: generic objective of maximizing reward is enough for AGI3

Why Checklist?

  • Many AGI models exist, but have gaps. Can you find any right now? How would you build AGI?

Goals

  1. Minimize & explicate unknowns.
  2. Help design & evaluation (of functionality, ethics, progress).

AGI Design Thinking

AGI Design Thinking: Modular “Designed Organization”

  1. Define application requiring AGI (cognitive goal/task/problem)
  2. Empathize environment, world-knowledge, and cognitive capabilities required (from human intuitive to explicit technical) by (1)
  3. Create descriptive functional system’s architecture (high-level intuition-pumped human-inspired design narrative aid)
  4. Make an inventory of algorithms and hardware available
  5. Divide labor & orchestrate computational modules
  6. Operationalize cognitive architecture: specify software & hardware
    • goal/task
    • perception (world & self)
    • data processing
    • learning, cognitive scaffolding
    • orchestration
  7. Try to implement.
  8. Iterate
  • Key Problem: Orchestration: How can we know beforehand whether a particular architecture actually works?

AGI Design Thinking: Cybernetic “Constrained Organization”

  1. Identify purposes1/ rewards3/goals on the highest level
  2. Empathize RICH environment for agent(s)3 to facilitate & constrain learning
  3. Make an inventory of algorithms and hardware available. Evaluate with theoretical Universal AI5 [hutter_universal_2005]
  4. Decide: 1 or >=2 agents1,3,5
  5. Divide labor & orchestrate Society of Mind for >=2 agents
  6. Operationalize agent(s): specify software & hardware
    • purpose/goal, generic reward
    • environment (world & self)
    • perception
    • cybernetic agent
    • reinforcement learning
    • algorithms;
  7. Try to implement.
  8. Iterate
  • Training in interaction with environment
  • Reward-is-enough hypothesis: Rich environment and generic reward/goal is enough for AGI. When agent gravitates towards reward, sub-goals/skills are learned implicitly on the way.3
  • Key Problem: How to partition & scaffold learning space? Scaffolding of learning, to monitor orchestration & transparency.

Intuition Pump

  • Pros: Helps in designing architecture.
  • Cons: Illusion of explanation. Only the tip of the ice berg of cognition and the world is visible.

Open Problems

  • Perception, Learning, Architecture1, Orchestration
  • Values & priorities of civilization?1