Building One with Tomer Cohen

The episode delivers a solid cluster of actionable frameworks - kill criteria, information gain per dollar, the monkey/pedestal metaphor, and the 7%/93% learning thought experiment - but the pace is uneven, with meaningful filler (the Waymo ride anecdote, the glasses tangent, a lengthy host recap) and Astro's talking points feel rehearsed rather than discovered in conversation.

“If 7% of the time is what it would take for you to rebuild the solution, then 93% of the work was figuring out what the solution was and 7% was building it.”

“you take this moonshot story hypothesis, huge problem, science fiction sounding product or service, breakthrough technology. As soon as you have one of those high five, good for you. You did it. We have another one of our investigations. Here's $20,000. Here's $50,000. Go see how fast you can kill it because there is a 99% chance you're wrong.”

The card counting, 10x vs 10%, and monkey/pedestal frameworks are well-worn Astro Teller public-talk staples that have circulated for nearly a decade; the episode adds genuine freshness only in the specific project stories (Loon→Tara, the fusion physics bet) rather than in the overarching theory of innovation.

“We're trying to be the card counters of innovation.”

“If I ask you to do something to make a phone that runs on the same battery the same number of milliamp hours for 10% longer, at some unconscious level, you will start with the phone that already exists.”

Astro Teller is a genuine top-tier practitioner - architect of Waymo, Google Brain, Wing, Loon, and Tara at operational scale over 15+ years - not a career podcast guest, and he draws directly from hands-on project decisions rather than theory, though his polish suggests these are well-rehearsed narratives.

“We actually did almost 1000 plasma shots before we shut the experiment down.”

“eight years later, the Tara team, which has now graduated from EX and is an independent company, is moving more data every like two or three minutes in 20 countries around the world than Loon did in its entire nine year history”

The episode has a strong density of concrete specifics - named dollar thresholds, kill criteria routes, plasma shot counts, Tara's comparative throughput - though some of the most important figures remain approximate and a few claims (carbon nanotubes, synthetic genomics) are left abstract.

“we said, here are 10 different 100 mile routes around the Bay Area...for you to do each of these routes at least once without a human touching the wheel or the brake of the gas pedal”

“call it a trillion dollar opportunity if it works, maybe a 1% chance that we actually get there...these people said, we can ask and answer this question for like under $10 million”

The host is clearly well-prepared and occasionally presses meaningfully - notably asking Astro to critique his own learning framework and challenging the Glass narrative - but he frequently answers his own questions, validates rather than probes, and lets significant claims (1% hit rate, cultural independence) pass without follow-up pressure.

“Astro, if you were to critique the focus on learning just in kind of in a intellectually honest way because everything has a shadow effect, do you see times when it becomes an excuse?”

“Um, that sounds like an amazing mandate. It also sounds like a recipe for disaster a little bit, uh, at least in other companies.”