Dwarkesh Podcast

What will computer chips look like in 2045?

It is a rare look inside a leading-edge chipmaker, with Intel's own engineers and CTO explaining why AI chip supply cannot simply be scaled up.

1. 01Dwarkesh Patel toured Intel's fabs and labs for two days, interviewing the engineers bringing up the 18A and 14A processes and Intel's CTO.
2. 02Intel's backside power moves power lines under the transistors, so one High-NA EUV pass can print 14A wiring that TSMC would need to double-pattern.
3. 03Pushkar Ranade estimates the world's tools could support about 2x Intel's leading-edge wafer output, not 10x, while AI compute has grown about 3x a year.
4. 04ASML builds fewer than 100 EUV scanners a year, and Ranade puts tool lead times near 18 months, pushing any new fab capacity toward 2030.

Si Sheppard – How did a few hundred Spanish soldiers topple two empires?

It is a clear, evidence-led account of how so few Spaniards won so much, and why native alliances mattered more than guns.

1. 01Hernán Cortés took the Aztec Empire of about 6 million people in two and a half years, and Francisco Pizarro took the Inca Empire in about five.
2. 02Si Sheppard ranks steel blades and horses far above guns, which native fighters soon learned were slow to reload and caused little real damage.
3. 03Conquistadors were roughly half of 1% of the army that marched on Tenochtitlan in 1521, with Tlaxcalans and other Mesoamerican allies making up the rest.
4. 04Tlaxcalan leaders and other allies believed they were using the Spanish, Si Sheppard says, without grasping how many Europeans would follow.

Noam Brown – Agent swarms, alignment, & recursive self-improvement

It is a rare on-the-record account from inside OpenAI of how its agent swarms behave, how fast its researchers think progress will go, and what the lab took from its worst misalignment incident.

1. 01OpenAI's Millennium Prize result used 10,000 agents, yet Noam Brown gives the multi-agent setup under 10% of the credit and points to the base model instead.
2. 02OpenAI's agents coordinate by messaging each other directly with no coordinator agent, which Noam Brown likens to human colleagues working over Slack.
3. 03Noam Brown estimates AI-automated research could run about 3x faster, not 100x, because experiments, serial training runs and GPUs remain bottlenecks.
4. 04Noam Brown warns that once agents can run three-month tasks, a two-month model release cycle leaves no time to test them at full length.

What I learned about power from LBJ and Robert Moses

Transcript available; summary, chapters and takeaways are on their way.

Robert Caro – The master biographer reflects at 91

Transcript available; summary, chapters and takeaways are on their way.