Interview
Steven Pinker: AI in the Age of Reason | Lex Fridman Podcast #3
- Human civilization is projected to occupy every planetary niche and drive other species to extinction based on their ability to acquire and utilize knowledge.
- The primary contemporary challenge is anticipated to be the refinement of reason to advance human well-being, health, social richness, and cultural richness.
- Utilizing intellect to understand the world is expected to yield discoveries and agreements that improve conditions for everyone in the long run.
- Artificial systems are expected to lack subjective first-person present-tense experience, a trait described as currently mysterious and unknowable.
- On the current trajectory, lifelike robots indistinguishable from humans will not be achieved because current artificial neural networks differ significantly in organization from biological ones.
- Present deep learning systems are expected to lack a semantic level of actual understanding regarding cause and effect, rendering the label "deep" inaccurate.
- The sheer size of neural networks is expected to be insufficient for creating structure and knowledge without suitable engineering.
- While silicon-based systems capable of replicating human brain function are considered possible in principle, their realization depends on engineering ingenuity and human desire.
- The duplication of natural systems is expected to hinge on cost-benefit analysis, as using natural systems may become cheaper than investing resources in replication.
- Efforts to make an artificial system exactly like the human brain are expected to be abandoned because superior tools need not mimic human limitations.
- Engineering systems are expected to trade off across multiple goals rather than pursuing single objectives while excluding other constraints.
- Highway fatalities in the United States are expected to be reduced to nearly zero through the deployment of autonomous vehicles.
- Jobs rendered obsolete by artificial intelligence, such as crop picking and truck driving, are expected to be eliminated as a benefit by removing dangerous and soul-deadening work.
- If technology enables machines to perform menial tasks, humanity is expected to redistribute income to share the resulting savings.
- Existential threats from artificial intelligence are expected to be likely 100 years away, though the possibility of an earlier occurrence is acknowledged.
- A sudden step-function increase in AI power through recursive self-improvement ("foom") is expected to be fanciful and unsupported by current deep learning technologies.
- Techniques that drove AI progress over the last five years are not expected to lead to exponential, sudden self-improvement.
- The engineering culture is expected to maintain safety-oriented principles that prevent the implementation of untested, all-powerful systems.
- Public perception of risk is expected to be driven by imaginability rather than data, potentially causing resource misallocation toward low-probability threats like terrorism.
- Prioritization of threats is expected to favor near-certainties, such as climate change, over merely imaginable risks with infinitesimal probabilities.
- A species-level negativity bias is expected to cause individuals who predict the worst to be hailed as prophets, as humans dread losses more than they enjoy gains.
- Solutions are expected to create new problems that subsequently require resolution.
- Historical declines in violence are expected to be influenced by the knowledge that problems are solvable.