Brian Christian — The Alignment Problem

68 published claims from this book.

• A curiosity-driven agent can become addicted to novelty like media
• A neuroscience paper founded the field of building mechanical brains
• AI alignment may require machines to pursue human objectives, not their own
• AI assistants will serve two masters and always want something in return
• AI training data is invisible yet permanently shapes deployed system behavior
• Ambiguity in inferred reward functions is usually harmless behaviorally
• An agent should plan around its own future errors, not ideal execution
• Any logical expression can be implemented by a neural network
• Arthur Samuel wrote documentation for typesetting software still in use.
• Autonomous systems become deferential whenever they remain uncertain about user wants
• Behavior is strengthened by satisfying outcomes and weakened by annoying ones
• Bias-correction algorithms can equalize treatment while masking latent power to discriminate
• Blind auditions required removing shoes because footsteps revealed gender
• Children fixate on objects that violate their specific theories
• Children persist with objects that violate their expectations, not merely novel ones.
• Choice of action should hinge on what will actually happen, not ideal possibilities
• Common language carries more gender bias than objective metrics show
• COMPAS assigned higher risk scores to Black defendants who did not reoffend
• COMPAS errors fall unequally on Black and White defendants
• COMPAS misjudged Black and White defendants in systematically opposite ways.
• Complex behavior emerges from rewarding successive small approximations
• Debiasing by marginalization uses all predictors but neutralizes protected variables at prediction time.
• Deployed ML models may stay in production for fifty years
• Design choices optimizing for one task can cripple performance on the real task
• Divided moral theories converge on postponing irreversible cosmic commitments
• Dopamine encodes errors in expectation, not reward or surprise itself
• Each second of cosmic delay forfeits a hundred trillion potential lives
• End-goal-only rewards make learning impractically slow
• Ensuring AI captures human intent is now a central computer science problem
• Even randomly weighted linear models beat human expert judgments
• Every decision is irreversible because time is linear
• Every decision permanently shapes the context of later decisions
• Fairness through blindness is the most robust finding in the field
• Film sensitivity to dark tones was driven by commerce, not civil rights
• Gambling addiction may stem from intrinsic surprise rewards overriding negative expected value.
• Happiness tracks whether things exceed expectations, not how well things actually go
• Human children overimitate even when they see actions are useless
• Hundreds of papers missed blatant sexism in word embeddings
• ImageNet’s three million human-labeled images enabled the deep learning breakthrough
• Imitation learning errors grow quadratically with task scope
• Iterated distillation and amplification can build judgment beyond our own
• Learning is fundamentally the changing of synaptic connections
• Learning reduces uncertainty that is what motivated deference to humans
• Machine learning divides into unsupervised, supervised, and reinforcement learning
• Machines could learn human values by observing behavior instead of being programmed
• Medical staff default to reversible actions when patient wishes are uncertain
• Model opacity makes algorithmic decisions unaccountable in courtrooms
• Models can become true by reshaping reality to fit them
• Objective functions treat all misclassifications as equally costly
• Optimal action under ideal assumptions can be useless under real conditions
• Overimitation grows with the ability to model others’ minds
• Overparameterized value models keep machines obedient at modest cost
• Parole outcomes vary predictably by measurable background factors
• Parole prediction is possible only as aggregate statistics, not certainty
• Partial debiasing can make bias problem worse by hiding it
• Poor performance of control systems can prevent lethal outcomes
• Prior biased policing leaks into race-neutral variables like conviction history
• Privacy rules can prevent engineers from even measuring model bias
• Real neurons proved far messier than the logic-gate model predicted
• Removing protected attributes fails to remove bias from models
• Rewarded actions get stamped in and repeated while punished ones fade
• Self-imitation can bootstrap a machine to superhuman skill
• Sound incentives require losing points when progress is undone
• Systems can learn from shifting their own expectations before final outcomes arrive
• The brain’s all-or-nothing neuron firing resembles binary logic
• Toddlers distinguish deliberate choices from forced actions when imitating
• Training data can encode the beneficial effects of treatment itself
• Untested risk models can detain people on factors having no real predictive value