What is intelligence? (2024)

Objective definitions and formal models

• Some argue an objective definition of intelligence is impossible; we only ever get “intelligence as defined by N,” not a context-free essence. Others push back, citing formal work (e.g., Hutter/AIXI) as useful idealizations.
• AIXI/Solomonoff-style formalisms define an “ideal” agent that considers all Turing machines consistent with data and weights simpler ones more (Occam + Bayes). This yields a precise but noncomputable standard of inductive intelligence.
• Debate: is parsimony the only general principle, or can “true intelligence” trade some parsimony for efficiency, redundancy, or smoother explanation changes?

Prediction, parsimony, and world models

• Several commenters align with the book’s emphasis on prediction as central: intelligence as modeling likely future states and acting to thrive.
• Others stress that prediction alone (universal function approximation) is too weak; what matters is building structured, causal world models that support simulation and planning, not just mapping past states to next states.

Intelligence, comprehension, and creativity

• One line: intelligence = problem-solving + insight/novel pattern creation; creativity (e.g., new boxing styles) is key, not just extrapolation from data.
• Another line distinguishes intelligence from comprehension and memory: current AIs are seen by some as “pre-calculated idiot savants” lacking on-the-fly comprehension, generalization beyond training, or meta-reasoning over their own goals.

LLMs: pattern matching vs genuine reasoning

• Some treat LLMs as mere next-token predictors with no logic or knowledge, just high-dimensional “repeated information.”
• Others argue they do form internal conceptual representations and even implement logical circuits in their computational graphs, though still data-driven and fallible (e.g., seahorse emoji failure).
• Disagreement persists over whether creative-seeming outputs under unusual constraints reflect real creativity or just human projection onto hallucinated text.

Embodiment, social context, and costs

• One perspective: intelligence is tightly coupled to embodiment, energy costs, and evolutionary/economic pressure to “pay back” execution costs and support future agents; intelligence becomes efficient search under those constraints.
• Simple control systems (kettles, cisterns, governors) are discussed as minimal “goal-seeking” intelligences, though some say their intelligence is really that of their designers.
• Another view emphasizes that artificial systems lack the deeply embodied, evolution-shaped will to persist that characterizes biological intelligence.

Philosophical and epistemic critiques

• Commenters note the difficulty of defining intelligence without clear distinctions among intelligence, comprehension, memory, and action.
• A long critique targets the book’s treatment of Hume and Kant, arguing it misreads the is/ought problem and transcendental idealism and underestimates the role of synthetic a priori knowledge and time in human cognition, which current AIs lack.

Reception of the book and format

• The multimedia book is available free online; some appreciate its breadth and cross-domain synthesis (evolution, computation, cybernetics), treating it as thought-provoking pop science.
• Others find it prolix, rhetorically manipulative (a “yes-set” persuasion sandwich), and light on genuinely novel ideas relative to existing predictive-processing and neuroscience literature.
• The design (scroll hijack, many short sections) and lack of clear central argument make some readers skeptical or unwilling to invest the time.