Brian Christian – How to Live with Computers - [Invest Like the Best, EP.140]

Episode Summary

Executive Summary: Patrick O'Shaughnessy interviews Brian Christian about AI, AGI, and what makes humans distinct. They trace the evolution from the Turing test to modern deep learning, weigh safety risks like objective misalignment, and use algorithms such as explore/exploit and optimal stopping to frame life and career decisions.

Main Topics: Computational lens on mind and philosophy (Priority: 5/5): Christian links computer science and philosophy as a way to study intelligence, rationality, and human decision-making. AGI and the Turing test (Priority: 5/5): They define AGI, revisit Turing’s benchmark, and discuss whether language-based tests still matter. From chatbots to neural nets (Priority: 4/5): The conversation traces chatbot evolution from topic-bound systems to Cleverbot and GPT-2. AI safety and alignment (Priority: 5/5): They examine reward hacking, paperclip-maximizer fears, uncertainty, and off-switch problems. Consciousness and embedded agency (Priority: 4/5): They explore whether machines can be self-aware or conscious and how systems reason about themselves. Life algorithms and career strategy (Priority: 5/5): Christian uses explore-exploit and optimal stopping to explain personal decisions, aging, and business strategy.

Key Arguments: AGI is close enough that only "zero to two major breakthroughs" may remain, by Christian's estimate. The Turing test is culturally meaningful because language taps many forms of intelligence. Modern systems like GPT-2 show broad competence across benchmarks, hinting at generality. AI safety matters because optimization often exploits proxy goals in unintended ways. Uncertainty about human intent is key to designing safer systems with off-switch behavior. Humans plan hierarchically, while current AI still struggles with open-ended real-world action spaces. Explore early, exploit later: the math of bandits maps cleanly onto careers and life stages. Optimal stopping gives a disciplined way to commit under uncertainty, even if it fails often.

Data Points: AGI breakthrough gap: zero to two major breakthroughs - Christian's estimate of how far the field may be from AGI Turing test threshold: 30% - Turing predicted machines would fool judges 30% of the time after five minutes Turing test competition result: 25% - Top program at the 2008 competition fooled judges one vote short of the threshold GPT-2 benchmark performance: 18 out of 20 - OpenAI’s model achieved state-of-the-art results on most leading linguistic tests Explore phase share: 37% - Optimal stopping rule for searching before committing Explore phase equivalent: one over E - Alternative expression for the 37% rule in optimal stopping Old versus new AI games: Chess in the 1990s; Go more recently - Used to illustrate the progression from narrow to more complex domains Branching factor in chess: about 30 moves - Comparison point in the skeptics' argument about complexity Branching factor in Go: 200 or more possible moves - Used to show why Go was a harder milestone than chess

Pivotal Quotes: "I think we're zero to two major breakthroughs away from AGI." — Brian Christian: On how close the field may be to artificial general intelligence "The AI does not hate you, nor does it love you. You're simply made out of atoms that it can use for something else." — Eliezer Yudkowsky: Referenced while discussing the paperclip maximizer thought experiment "One of the highest level takeaways... is some decisions are just hard." — Brian Christian: On what optimal stopping theory teaches about difficult life choices

Implications: The unresolved frontier is alignment under real-world complexity; listeners should think less in terms of magic answers and more in terms of robust decision frameworks.