Intelligence Isn’t the Bottleneck. Cost Is.

An edited transcript of the Naval podcast — Naval in conversation with Garry Tan (Y Combinator), Daniel (Able), and Farbood (A-List) — on riding AGI, AI anxiety, open vs. closed source, China and hardware, the future of work, and the case for a Universal Basic Robot.

YouTube · NavalRiding AGI, AI Anxiety, Who Funded COVID, Defending Taiwan, and California EmpireWatch on YouTube →

TL;DR

• Intelligence isn't the bottleneck — cost is. The models are already smart enough for a huge amount of real work; price per token is what's rationing deployment, and it's collapsing.
• ~90,000× more inference in 24–36 months. Several orders of magnitude, "baked in" by the chips and data centers being built. Each order of magnitude unlocks new emergent capabilities.
• $100/month → $2.84. One founder drove per-user AI cost down by ~35× in a few months. The leverage is in the harness and the engineering, not just the model.
• Open source is 5–10× cheaper and closing fast — and China is shipping most of the open weights. The gap went from ~12 months to ~9 to ~6, and some say 3.
• Both hardware and software are now commoditized. What isn't: AI research itself. That concentrates power in a few labs — "two kings" — with open source as the counterweight.
• The problem was never displacement, it's the speed of transition. Farms went from ~50% of US labor to ~2% without 48% unemployment. Today, people using AI are working more, not less.
• From UBI to UBR — Universal Basic Robot. Automate the work nobody wants; humans become "AI handlers" and "robot trainers." Practical robotics: 2–10 years, nobody says never.

Intelligence isn’t the bottleneck — cost is

Naval: The models are already smart enough to do a huge amount of real work. What’s rationing deployment isn’t intelligence — it’s cost. Price per token is the constraint, and it’s collapsing.

Daniel: We felt that directly. We drove per-user AI cost from about $100/month to $2.84 in a few months — roughly 35×. The leverage was in the harness and the engineering, not waiting for a better model.

Naval: And there’s far more compute coming. Something like ~90,000× more inference over the next 24–36 months is effectively baked in by the chips and data centers being built. Every order of magnitude tends to unlock new emergent capabilities.

The AI-writing fight

Naval: Here’s a thing I feel strongly about. When you write something meant for humansand it’s clearly AI-generated, that’s a disservice — you’re wasting their time. Everything the AI wrote should be compressed by you, to respect the reader. Otherwise their AI reads your AI and neither of you is in the loop. And good writing and good speaking are the output of good thinking — if you never work the muscle, you lose it.

Garry: My counter: I can now have a high-bandwidth conversation with a smart model all the time.

Naval: Sure — talk to the model all you want. Brainstorm with it: give me synonyms, tangential ideas, argue with me. I just don’t use it to produce the final output a human is expected to sit and read.

Garry: Out of the box the quality is bad, I agree. But if you build a big enough corpus, do cross-modal evals, and tune a “skill file” that captures your style and diction, you can get it indistinguishable — even beforethe next model. I built a retrieval tool over my whole corpus — every email, Slack, DM, plus ~400,000 markdown files of anything I’ve ever thought or read. To predict the future, live in the future and work backwards.Everyone should be “AI-maxing” just to figure out where it’s going.

Naval: I’m still a little better defended on that hill — I write very short, and models are bad at that. Good writing is novelty; it’s unexpected.Anything guessing the next token from a regression struggles to be truly original. Code’s different: it’s meant to be consumed by a computer, so I’m happy to have a computer write it.

Open source vs. closed source

Naval: So — how good are the open-source models?

Farbood: Really unbelievable. Five to ten times cheaper. Some of them are mind-bendingly good. They need a good harness, but with one, they work really well.

Naval: How far behind the frontier?

Farbood: Closing fast — and China is shipping most of the open weights. The gap went from ~12 months to ~9, to ~6. Some say 3.

Naval: So hardware and software both commoditize. What doesn’t is the researchitself — that concentrates in a couple of labs. Call it “two kings,” with open source as the counterweight.

The future of work

Garry: People hear these numbers and panic about jobs.

Naval: The problem was never displacement — it’s the speed of the transition. US farm labor went from ~50% of the workforce to ~2%, and we never saw 48% unemployment; the economy re-absorbed it. The risk is doing that in years instead of generations.

Farbood: And right now the people using AI are working more, not less — it raises the ceiling on what one person can do.

From UBI to a Universal Basic Robot

Naval: The endgame people reach for is UBI. I’d reframe it: Universal Basic Robot.Automate the work nobody wants, and humans move up the stack — “AI handlers” and “robot trainers.”

Garry: Timeline on practical robotics?

Farbood: Somewhere between 2 and 10 years. Nobody serious says never anymore.