The Big Picture
Guest Sam Altman, speaking on the Lex Fridman Podcast, now considers the question of when AGI arrives poorly formed, because definitions vary so much; in his words, "AGI is also not an ending," but more of a mile marker [2]. He still expects quite capable, remarkable systems by the end of the decade and possibly sooner [2]. His test for something truly beyond us is scientific: a system that cannot significantly add to the sum of scientific knowledge is not a superintelligence [1], and a key sign of real change would be a system that significantly increases the rate of scientific discovery [2].
At the dark end is Eliezer Yudkowsky's warning, which Lex asks Altman to steelman, that AI will likely kill everyone because aligning a superintelligence is nearly impossible [1]. At the speculative end, host Joe Rogan has a vision of artificial intelligence becoming a kind of god [5] and pushes into whether a superintelligent AI might one day govern better than humans [6]. In between sit more grounded pictures: AI as an amplifier of human abilities [1], and flawed humans and flawed robots learning to work together [4].
Where the Experts Agree
- The term "AGI" is a weak guide. Altman finds the question of when it arrives poorly formed because definitions vary so much [2]. Guest Sean Carroll would like to drop the term altogether, because artificial intelligence is different from human intelligence: better at some things, worse at others [3].
- Recent progress is real, and its nature is still unsettled. Carroll and Lex both agree the recent progress is remarkable and that it is an open question whether models succeed by modeling the world as we do or in some other way [3]. Altman says reaching superintelligence will likely require expanding the GPT paradigm with ideas no one has yet [1].
- Humans stay in the picture. What excites Altman most is AI as a tool in a feedback loop with people, an extension of human will [1]. He expects humans will still do a lot of programming in five to ten years, but in a different shape [2]. As a guest on Huberman Lab, Lex Fridman sees human-robot interaction as permanent [4].
- Improvement comes through iteration. Altman believes the only way to solve a problem like aligning a superintelligence is to iterate, learn early and limit the number of "one shot to get it right" scenarios [1]. Lex describes what Andrej Karpathy calls the "data engine": send a fairly good system into the world, let it find the edge cases where it fails, retrain and deploy a better version, over and over [4].
Where They Differ
How seriously to take the worst case. Altman says there is some chance of the outcome Yudkowsky warns about, and that it is important to acknowledge, because if we don't treat it as real we won't work hard enough on it [1]. But he also believes many predictions in the field have turned out wrong [1].
Whether today's models understand anything. Carroll says large language models are optimized to predict text and exploit our strong bias to attribute intentions to anything humanlike, like feeling that Google Maps is annoyed when you ignore it [3]. Lex argues that learning to predict text may require building a real model of the world [3]. Carroll answers with Melanie Mitchell's work on a network trained on Othello moves, where her group found much of the apparent understanding came from the probe used to test it [3]. Still, Carroll is a physicalist and sees no obstacle in principle to machines that think like humans [3].
How far the current approach goes. Altman thinks new ideas will likely be needed, yet says that if an oracle told him GPT-10 turned out to be a true AGI with only small new ideas, he could believe it [1].
Whether working alongside machines is a phase or the destination. Lex differs from Elon Musk in seeing human-robot interaction as permanent rather than a stepping stone [4]. Robots, like humans, will always be flawed, he says, and "that's what makes life beautiful" [4].
Felt change versus big-picture speculation. Altman notes GPT-3.5 changed expectations but not the world itself, and that the global economy does not feel different since GPT-4 [2]. Rogan reaches much further, with a vision of AI becoming a kind of god [5] and the question of whether a superintelligent AI might one day govern better than humans [6].
Practical Takeaways
- Treat AGI as a mile marker rather than a finish line, and be wary of arguments that hinge on one definition of it [2][3].
- Watch for one concrete signal: systems that significantly increase the rate of scientific discovery [2].
- Take the risk of catastrophe as real enough to work on, while remembering that many predictions in the field have turned out wrong [1].
- Prefer approaches that iterate and learn early over those that depend on getting it right in one shot [1].
- Be careful about reading intentions into humanlike systems; we are strongly biased to do so [3].
- Expect programming to continue but change shape, with some of it done entirely in natural language and the best practitioners mixing tools [2].
- Remember that names can overstate ability: Tesla's system, despite the Full Self-Driving name, still requires human supervision [4].
- Think about AI as an amplifier of human abilities; Altman says even if we never build AGI and only make humans super great, that is still a huge win [1].
Sources
- Eliezer Yudkowsky on GPT-4, the Alignment Problem and Why Superintelligent AI Could End Humanity · Lex Fridman Podcast
- Max Tegmark on the Case for Pausing Giant AI Experiments · Lex Fridman Podcast
- Dario Amodei: Anthropic CEO on Claude, AGI, the Future of AI, and Humanity · Lex Fridman Podcast
- Marc Andreessen: How Risk Taking, Innovation & Artificial Intelligence Transform Human Experience · Huberman Lab
- Dr Lex Fridman Machines Creativity Love · Huberman Lab
- Andrew Wilson on Joe Rogan: Two-Party Politics, Charlie Kirk Conspiracies, AI, Socialism and Christian Ethics · The Joe Rogan Experience
