David Eagleman predicts an AI president within 20 years, citing an Anthropic paper where an AI interviewed 81,000 people.
Eagleman calls the brain a neural parliament and expects AI to use bolted-together architectures that vote.
Eagleman says AI hallucination is often hypothesizing, and LLMs remix well but lack the brain's continuous learning.
Why a neuroscientist expects an AI president within twenty years
David Eagleman thinks we will elect an AI as president within twenty years, and his case starts with how brains work. The brain is a parliament of competing drives that vote, and he expects machine intelligence to be assembled the same way, from architectures bolted together. He is optimistic that intelligence is computationally capturable.
Takeaways
An AI president is inevitable within twenty years. He calls it inevitable and faster than anyone thinks, and asks why voters must choose between flawed humans. It would not chase donors or re-election, and people are more honest with AI than with other people .
The brain is a parliament of competing drives. He calls it a team of rivals and cites Nietzsche's 1880s account of competing drives . He expects AI to be bolted together from architectures that vote, as ChatGPT already sends maths questions to a calculator .
Intelligence is computationally capturable, he argues. He says machines can surpass the human brain, and that creativity is the remixing of ideas, which LLMs do extremely well . What people call hallucination is often the same thing as hypothesising .
Continuous learning separates brains from language models. The brain changes its weights every moment, while LLMs have a training phase followed by question answering . The brain runs on about 20 watts, and a large language model is unbelievably energy intensive .
Science will make giant data centres obsolete. He expects scientific advances to make today's giant data centres obsolete, as the Human Genome Project's machines were quickly outdated, though plenty of compute is still needed for other things .
Neuroscience is still at the foot of the mountain. Nine and a half of the ten unsolved questions he listed in 2006 remain open, and consciousness has no known shape of an answer . He hopes AI clears the list by 2036 .
How an AI presidency would work
Anthropic's March paper had an AI conduct 45-minute interviews with 81,000 people .
Each citizen would be quizzed yearly by a personal agent and could update it after a change of mind .
The agents would converge into something representing what everyone actually wants, which he calls the truest democracy .
It must be open source and transparent, with anyone able to inspect the weights .
Constitutionally the president must be a person born in the United States, so amendments would be needed .
Eagleman's theory of why we dream
The planet is dark half the time, and the visual system is at a disadvantage because you can still hear, smell, taste and touch in the dark .
If wiring breaks, the cortex takes the territory for other uses, so the visual cortex needs defending .
Every 90 minutes an ancient circuit blasts random activity into it, a screensaver for the territory, and plasticity across 25 primate species tracks dream sleep.
