What do computers want? - Reverse engineering what motivates computation | Michael Levin

Quick Overview

The core theme of the discussion between Lex Fridman and Michael Levin revolves around defining intelligence not as a fixed set of capabilities, but as the capacity to find novel, efficient means (low complexity) to achieve a fixed goal, contrasting this with simplistic, brute-force approaches.

Key Points: Intelligence is defined by William James as a fixed goal with variable means of achieving it, emphasizing adaptability over rigid processes. Levin suggests that many conventional computational systems, like those following set rules (e.g., turn the crank, time goes forward), are not truly intelligent because they lack reversibility and adaptability. The goal in biological systems, like development, is to find the lowest complexity pathway to a target state, which requires overcoming barriers through novel, intelligent maneuvers. Levin proposes two steps for assessing a system's intelligence: 1) identifying the space it operates in, and 2) determining the goal within that space. If a system is forced to use only one path (like a simple algorithm), it demonstrates low intelligence; true intelligence involves discovering entirely new approaches when old ones fail. The work in bioelectric imaging and reprogramming shows that memory/goals can be encoded in the environment, and resetting the environment can lead to new, goal-directed behavior. Levin contrasts the 'open loop' systems (which fail when conditions change) with goal-directed systems that can adapt their methods.

Context: This video is an episode of the Lex Fridman Podcast featuring Michael Levin, a leading researcher in developmental biology and bioelectromagnetics. The conversation centers on the nature of intelligence, contrasting machine computation and rigid algorithms with the adaptive, goal-directed processes observed in biological systems, particularly concerning morphogenesis and developmental reprogramming.

Detailed Analysis

Lex Fridman and Michael Levin explore the philosophical and scientific definition of intelligence, heavily referencing William James's definition: "Intelligence is a fixed goal with variable means of achieving it." Levin argues that biological systems exhibit this true intelligence because they can navigate complex environments to reach a target state (the goal) by finding novel, low-complexity pathways, even when initial pathways are blocked or reversed. He contrasts this with many computational models, which he describes as 'open loop' systems that follow fixed, non-reversible rules (like turning a crank) and fail when the environment changes unexpectedly. Levin outlines a two-step approach to evaluating intelligence: defining the operational space and then defining the goal within that space. He illustrates that true intelligence is demonstrated when a system can dynamically rewrite its approach (the means) to achieve the same goal, citing evidence from his lab's work in bioelectric signaling, which shows that encoded memories or goals can be reset by altering the physical environment, leading to entirely new developmental trajectories. The discussion implies that biological systems are inherently more intelligent than deterministic algorithms because they prioritize achieving the endpoint over strictly following the initial plan.