Alison Gopnik
About
Professor of psychology and philosophy at UC Berkeley; expert in child development and learning
Cast within
No topic-region cast yet — this appears once Alison Gopnik's compiled claims are aligned into a topic region's argument tree.
Claims by Alison Gopnik (20 of 82)
Humans Invent Their Own Objectives
Current AI excels at optimizing a specified objective function given examples of input and output, but human beings continually invent their own novel objectives and goals that no one set before, with some unexplained sense of progress toward better goals, and this capacity is not merely a matter of more compute power but a fundamentally different category of behavior.
Knowledge Gap Solved by Compute and Data, Not New Theory
The recent AI spring was not driven by a new idea about how the mind works but by old ideas from the 1980s (neural networks, associative networks) becoming practical once large internet-scale datasets and Moore's-law compute power made them able to scale beyond toy problems, surprising even their designers.
Children Generalize From Little Data, Machines Cannot
Human children are the best example we have of a system that truly works as an intelligence, generalizing from comparatively small amounts of data using innate structure, whereas current AI relies on statistical inference from huge datasets, is fooled by adversarial and superficial features, and suffers catastrophic forgetting when rules change slightly, so nothing we have is even in the ballpark of a four-year-old.
Curiosity-Driven AI Outperforms Reward Maximizers
An algorithm designed to seek out violations of its own model's predictions, treating surprise as interesting rather than as failure, can solve problems that typical reinforcement-learning reward-maximizers cannot, mirroring the insatiable curiosity that drives children to actively experiment on the world rather than passively await data.
Moravec's Paradox: Hard Tasks Easy, Easy Tasks Hard
Moravec's paradox holds that the things that impress us as humans, like arithmetic or chess, are precisely the things computers find easy, while the seemingly trivial things almost any human or child can do, like physically manipulating objects in the real world ('at chess'), are extraordinarily difficult for machines, so menial physical jobs like plumbing are harder to automate than high-cognition jobs like oncology.
Knowledge Problem Is a Search Problem
The hard problem of intelligence is not figuring out what our representations and rules are but solving a search problem: given the vast space of all possible beliefs, solutions, or actions, how do we converge on the relatively few that are true, good, and effective; studying how children narrow this space is the most promising route to progress.
Childhood Solves the Explore-Exploit Tension
Children are actually better than adults at reaching unlikely new solutions because, in Bayesian terms, adults' priors become more peaked and confident, making it harder to conceive of new possibilities; childhood itself can be understood as a solution to the explore-exploit tension, dedicating a protected life-history phase to wide exploration before narrowing in.
Consciousness May Dissolve Like the Life Question
Rather than a single capital-C consciousness problem with one explanatory answer, the productive path is to map specific kinds of phenomenology to specific computations and neural systems; consciousness may dissolve the way the 19th-century 'hard problem' of life and elan vital did, where the question turned out to be misframed and life proved to be complex non-obvious relationships among non-living things.
Computational Theory of Mind Is Best Available
Turing's idea of treating the human mind as a computational system has been an extraordinarily productive framework, enabling accurate predictions about behavior (such as children doing Bayesian inference over structured causal systems); no rival idea makes better predictions, so although biology might add something, the computational theory of mind remains the best hypothesis on the table.
Theory of Mind Develops in Self and Other Together
Children build everyday theories of how minds work using data from both their own behavior and others', and understanding develops in parallel for self and other: a child who cannot grasp that other people hold false beliefs cannot grasp it for themselves either, contradicting the Cartesian picture that we introspect our own minds and then project onto others.
Intersubjectivity Present in Early Infancy
The 'like me' recognition that other agents function as I do is present very early: neonatal imitation from birth implies an innate link between felt inner states and observed expressions, and by three months infants engage in 'primary intersubjectivity,' coordinating body movements in a conversational dance with caregivers, while mirror self-recognition only emerges around 18 months.
Climate Change Is the Larger Near-Term Existential Risk
Worrying is a triage problem, and current science suggests climate change is the larger near-term existential risk, deserving maximum attention now, whereas general AI is at an unknown future point; this does not mean AI researchers should neglect safety, but the relative urgency favors climate, and reassuringly the very scientists building AI, like those who built nuclear weapons, are among the first to insist on keeping the dangers in check.
Eighteen-Month-Olds Do Implicit Statistics
Despite adults being notoriously bad at explicit probability, 18-month-olds can implicitly perform statistical reasoning: shown that one object activates a machine eight out of ten times and another four out of ten times, they will choose the higher-probability object, and they do so even for closer ratios like two-thirds versus eight-tenths.
Reinforcement Learning Is Too Narrow for Exploration
Reinforcement learning, where an agent takes actions and learns from whether they make it better off (more reward), played a key role in solving games like Go, but it is too narrow for genuine world-understanding because an agent solely tracking whether it is better off will not perform the exploration needed to figure out how the world really works.
Human Before Puberty, Human After Menopause
We are most distinctively human up until puberty and after menopause; in between we behave like glorified primates focused on dominance hierarchies, mating, and acquiring resources, while the genuinely human activities like theory of mind, discovery, causal inference, cultural transmission, and large-scale storytelling are concentrated in childhood and elderhood.
Helplessness Is an Adaptive Price of Exploration
Prolonged childhood helplessness is not an accident but an adaptive trade-off: creatures living in changing, unpredictable environments need a protected exploratory period before they can exploit, and during that period they are necessarily bad at doing things and dependent on adults who keep them alive and fed so they can devote energy to learning.
Childhood Is Evolution's Simulated Annealing
For solving high-dimensional problems with many possible solutions, computer science shows the best strategy is to start with a wide, hot, random search and gradually cool into a focused, low-temperature search (simulated annealing); childhood is evolution's way of implementing this, beginning with noisy bouncy exploration before cooling into focused adult exploitation, because the two strategies cannot be done simultaneously and trade off against each other.
Babies Imitate Unlike Other Primates
From birth babies pay special attention to faces and imitate other people in a way that other primates do not, and by nine months they point to communicate and become agitated if an adult fails to follow their point, indicating early understanding that others share their attentional focus.
Children Build Causal Bayes Nets From Data
Using formalisms developed by Judea Pearl such as causal Bayes nets, Gopnik's work demonstrates that children construct causal graphical models from data and then use those models to decide on their interventions, exactly as the causal graphical model framework would predict.
Children Balance Imitation and Innovation
Children are social and cultural learners but do not blindly imitate; they balance what they observe an adult demonstrate against their own beliefs about how something works, and experimental manipulations of how confident or uncertain the demonstrator appears shift children toward more imitation or more independent exploration accordingly.
My Notes
Loading notes...