explainx / blog
Key takeaways from a Naval podcast founder roundtable with Gary Tan (Y Combinator), Daniel (Able Police), and Farbood (A-List): 90,000x compute scaling, GStack, OpenClaw cost cuts, and the ASI debate.
Jun 14, 2026
Two numbers determine how rich a startup founder gets: the growth rate and how long it continues. Paul Graham's Oxford Union talk shows the math that makes billion-dollar outcomes not just possible, but predictable.
Jun 3, 2026
In a stunning reversal, Sam Altman admits he was 'pretty wrong' about AI wiping out entry-level jobs. Dario Amodei quietly shifts from '50% of white-collar jobs eliminated' to focusing on augmentation. The timing? Both companies eyeing IPOs as enterprise AI costs spiral out of control.
Jul 3, 2026
GitHub Copilot now offers Moonshot AI's Kimi K2.7 Code as a selectable open-weight model — the first in Copilot's model picker. Pro plans first; Business and Enterprise require admin enablement. Here's how to turn it on.
Three founders sat down on the Naval podcast for the kind of conversation everyone in AI is quietly having anyway — the one where you try to figure out how far off the ground you actually are before you say something dumb into a microphone. Gary Tan (Y Combinator), Daniel (Able Police, which turns police body-cam footage into compliant reports and now handles translation and citizen reporting), and Farbood (A-List, a health super app) traded notes on compute scaling, agent costs, the ASI question, and whether AI-written prose is a status signal or just... writing now.
None of it will be current by the time it airs — that's the joke about covering AI. But the framework underneath the specific numbers holds up, so here's what's worth keeping.
| Question | Answer from the roundtable |
|---|---|
| Is inference compute really going up 90,000x? | Gary Tan's estimate, over 24-36 months, driven by chip and data center buildout already committed — "we might be off by a couple orders of magnitude, but magnitude, friends." |
| What's the real bottleneck — intelligence or cost? | Cost, according to Daniel. His team cut OpenClaw + Opus spend from ~$100/user/month to $2.84/user/month over 3-4 months of harness and eval work. |
| Is open source catching closed models? | Yes, and faster than the labs would like — via legal pre-training on uncopyrighted web data, distillation of closed-model outputs, researcher mobility, and leaked weights. |
| Will AI replace all white-collar work? | Disputed. The founders agree AI removes the last mile of "mundane" execution but disagree on whether creative and judgment work survives, and for how long. |
| Is AI-written content acceptable to publish? | Split down the middle — one camp says it disrespects the reader, the other says with a big enough personal corpus and cross-model evals it's already indistinguishable. |
Gary Tan opened with the number that's been making the rounds in YC circles: inference compute could grow roughly 90,000x from today to about three years out, driven by chip production and data center construction that's already underway, not speculative.
"A lot of people say that Nvidia is way overpriced, but what if it's way way underpriced by like several orders of magnitude, right?"
His framing: every order of magnitude increase in compute has historically unlocked new capabilities, not just more of the same. That's the argument for staying long on infrastructure even after a run-up — the payoff isn't linear with the spend.
Gary also described building GStack, the open-source Claude Code skill pack that became, in his words, "one of the top 100 open-source packages for how people vibe code" — built during a stretch of sleeping three hours a night for four months, going from not coding at all to teaching others how to do it. He also mentioned switching between Claude Code and OpenClaw stacks within 24 hours of meeting Brex's Pedro, and now using both.
The most concrete data point in the episode came from Daniel, describing how Able Police runs an agent fleet at scale:
"Every single person in our app has their own OpenClaw running for them... when we started with OpenClaw and Opus, it was $100 a month per person. We spent three, four months driving that down to $2.84."
That's a 35x cost reduction achieved purely through eval harness work and elastic fleet management — no change in the underlying model. Daniel's takeaway was blunt: "I don't think intelligence is the bottleneck. Cost is the bottleneck right now." He also predicted the next fight in the industry is over the agent harness — whether it's Codex, Claude Code, OpenClaw/Hermes, or something else — becoming the layer everyone standardizes on by 2027.
This tracks with what explainx.ai has covered on agent harness engineering and why AI companies want you using agents in the first place — the token economics of running fleets, not raw model IQ, decide whether a product is viable at scale.
Gary Tan framed the one open question he still cares about: do scaling laws keep compounding until models exceed the smartest humans through recursive self-improvement (ASI), or does progress plateau at "expert at everything, but not creatively beyond the training distribution"?
He pointed to progress on hard math/proof problems (referred to in the conversation as "the Airish math problem") as evidence that's harder to dismiss than people expected two years ago — while still maintaining he hasn't seen "broad general creativity" emerge. The chess analogy came up too: centaur chess (human + engine beating engine alone) was real for a while, until it wasn't. Whether the same trajectory applies to knowledge work is, per the group, the actual stakes-defining question — bigger than open-vs-closed or nationalized-vs-private.
The panel laid out — without invoking conspiracy — several concrete reasons Chinese open-weight models (GLM, DeepSeek, MiniMax and peers) keep narrowing the distance to closed US labs:
The group's read: hardware is commoditized and largely owned by China; software is commoditized by AI; the only remaining non-commoditized layer is AI research itself, and that's concentrated in a small number of labs with massive compute and proprietary data. For more on the open-vs-closed dynamic, see explainx.ai's coverage of closed source vs. local open source alternatives and the China AI playbook of free models and cheap compute.
The sharpest disagreement wasn't about compute or geopolitics — it was about whether it's acceptable to publish AI-written text.
Camp one: publishing raw AI output is a disservice to the reader. If you're not compressing your own thinking down to what's worth someone's time, you're outsourcing not just the writing but the thinking, and eventually "their AI is going to end up reading your AI and neither of you are in the loop."
Camp two: with a large enough personal corpus — one founder described maintaining ~400,000 markdown files of his own writing, emails, and messages, run through a custom retrieval system he calls "LSD mode" (lateral sarcastic drift) that cross-ranks ideas across multiple frontier models — AI-assisted output is already close to indistinguishable from human writing, and will fully close the gap "in less than 9 months."
Both camps agreed on one thing: out-of-the-box AI writing is currently bad — verbose, clinical, over-hedged. The disagreement is about how much a custom eval harness and skill file can fix that, not whether raw model output is good enough today. This is a live debate on explainx.ai too — see why AI companies want you using agents for the token-economics angle on why verbosity gets rewarded by default.
On labor displacement, the group split the difference between optimism and honesty about the transition speed:
This article summarizes a roundtable conversation as of July 4, 2026. Cost figures, compute projections, and product details reflect what was stated on the podcast and may change.
