Jim Keller: The Future of Computing, AI, Life, and Consciousness | Lex Fridman Podcast #162

Jim Keller returns to Lex Fridman's podcast to discuss the future of computing, from the craftsmanship behind great processor design to why instruction sets barely matter. He explains Tenstorrent's bet on graph-native hardware built to run neural networks the way they are actually written, versus GPUs that emulate graphs with pixel-shader machinery. The conversation ranges across Moore's law as scaling-by-quantity, autonomous driving and software 2.0, and how brains, consciousness, and dreams might be re-engineered. It closes on intensely human territory: Steve Jobs and Elon Musk's leadership, benzodiazepine withdrawal and Jordan Peterson, love as a functional source of novelty, and advice to avoid groupthink.

• Keller claims x86 is 'arguably the worst architecture on the planet' yet one of the most popular, arguing instruction sets evolve slowly and 'don't matter very much.'
  00:11:55
• Contrarian take: 'the future of computing is inefficiency not efficiency' because scaling across many computers beats per-computer efficiency.
  00:46:34
• Tenstorrent built hardware that natively runs graphs rather than emulating them like GPUs do, the core thesis of the company.
  00:53:18
• Keller predicts AI will not just learn physics but 'understand physics in ways that we can't understand,' a step beyond Wolfram's physics-by-equation.
  01:25:59
• He frames consciousness as a half-second-behind, single-threaded post-hoc narrative running atop a massively parallel brain.
  01:50:46
• Personal revelation: Keller became 'an expert in benzo withdrawal,' detailing the suffering his brother-in-law Jordan Peterson endured.
  02:00:10
• On GameStop/Robinhood, he argues 'we know five percent of what really happened' and the people who made the most money aren't the ones being discussed.
  02:08:31
• Keller reveals he is writing a book about love, framing love as a 'function that generates newness and novelty.'
  02:25:07

• On JavaScript: doing 'the crappy thing fast' and iterating by listening to developers often beats waiting for perfect.
  00:09:19
• Intel won partly by going open early (licensing x86 to ~seven makers) while rivals like Z80 stayed proprietary to one source.
  00:14:32
• On AMD's Zen, defining all module interfaces before writing RTL let teams test pieces independently and dramatically improved quality.
  00:38:48
• 'A beautiful design can't be bigger than the person doing it', good design matches human cognitive limits to module size.
  00:39:18
• A Tesla chose a tiny computer over a resistor to measure battery current because the computer was cheaper than the resistor.
  00:47:34
• Modern self-driving trains an arbitrarily large network that works, then refactors it down to what fits and is affordable to ship.
  01:17:08
• Keller preps his mind before sleep with problems he wants to solve, then directs and remembers his dreams to work on them.
  01:42:31
• He cites a statistic that roughly one-third of ocean bacteria die every day from viral invasion.
  02:06:29
• On fear of humiliation: 'you're the only person on the planet' who cares about your embarrassment; everyone else forgets it.
  02:20:26
• Love is functional: we habituate to everything except what we love, which keeps it new and makes us notice problems instantly.
  02:26:08