Roman Yampolskiy: AI Can’t Be Controlled — and We’re Building It Anyway - #558

Roman Yampolskiy has spent two decades trying to prove that superintelligent AI can be controlled. He couldn’t. I invited him on to make his case. Subscribe if you want science with evidence, not speculation.

Roman is a professor of computer science at the University of Louisville and one of the earliest researchers in AI safety. His book AI: Unexplainable, Unpredictable, Uncontrollable started as an attempt to solve the alignment problem. After decades of work, it became a proof that the problem cannot be solved. Not difficult. Mathematically impossible.

I push back hard. We go after the Einstein test: can a large language model trained only on pre-1911 physics reproduce what Einstein did with the same data? We ran that experiment. It failed. Roman and I disagree about what that means. We also get into the halting problem and what it actually tells us about predicting smarter-than-human behavior, whether value alignment is a real problem or a well-funded category error, the case for a government moratorium on frontier model development, and why Roman thinks giving an AI agent access to your computer is the dumbest thing a smart person can do.

What you’ll hear:

Whether AI control is mathematically impossible or just unsolved

Why Roman thinks all current AI safety work is security theater

What the halting problem actually means for superintelligence

The alignment problem: real issue or well-funded category error

Why Roman wants a moratorium on frontier model development

What to tell your kids about careers in a world where Roman might be right

If you listen to other people, the best you can become is average.

CHAPTERS

00:00 Creating a mind without an off switch

01:34 Solving problems beyond our own intelligence

04:08 Einstein’s epiphany and the limit of AI intuition

08:18 Assessing the Einstein test: Why the experiment failed

12:22 Path dependency: Are LLMs and GPUs our QWERTY?

16:10 The barriers preventing AI from solving physics

21:54 Safety vs. Capability: Why toddlers are safe but teens are not

23:06 The halting problem: Predicting agents smarter than us

25:58 The impossibility of a system proving its own integrity

28:18 Regulation: Genuine safety or a gift to oligarchs?

33:28 Is human cognition non-computable? Penrose vs. the field

39:00 Ethical duties: Must we treat AI with humanity?

43:00 From internet memes to monsters: Decoding the book cover

46:22 Customized realities: Can everyone have their perfect world?

49:50 Von Neumann probes and the panspermia hypothesis

55:02 Categorizing AI: The one version that should terrify you

58:22 Pause AI: The movement for a development moratorium

59:58 Career advice for kids in a post-professional world

01:07:58 Cross-examining Sam Altman

01:15:48 Roman’s dream debate

01:19:50 Lessons for a younger self

Substack: https://briankeating.substack.com

Get the transcript, fascinating bonus content, and my Monday M.A.G.I.C. Message: https://briankeating.com/yt

Have a .edu email and live in the USA? You automatically win a meteorite: https://BrianKeating.com/edu

Subscribe: https://www.youtube.com/DrBrianKeating?sub_confirmation=1

Support Into the Impossible on Patreon, get my weekly M.A.G.I.C. Message, unfiltered bonus content, and live monthly Office Hours with me: https://www.patreon.com/drbriankeating

Join this channel for perks, monthly Office Hours, and your name in the Member Roster at the end of every episode: https://www.youtube.com/channel/UCmXH_moPhfkqCk6S3b9RWuw/join

Featured Guest:

Roman Yampolskiy on Twitter/X: https://x.com/romanyam?lang=en

AI: Unexplainable, Unpredictable, Uncontrollable: https://www.romanyampolskiy.com/books/

My books:

Losing the Nobel Prize (memoir): http://amzn.to/2sa5UpA

Think Like a Nobel Prize Winner: https://a.co/d/03ezQFu

Focus Like a Nobel Prize Winner: https://a.co/d/hi50U9U

Galileo’s Dialogue (first-ever audiobook): https://a.co/d/iZPi9Un

Twitter/X: https://x.com/BrianKeating

Substack: https://briankeat