Dwarkesh Patel

Robert Caro says the Great Depression played a big role in what Robert Moses was able to build. When Moses needed architects, thousands of out-of-work architects lined up on Fifth Avenue, and he could hire them at very low wages. More than the pay, they were available because they had nothing else to do. He does not remember which project it was. It is in The Power Broker.

He could hire a great many people, and he could hire and fire them in a ruthless way that made the workmen work hard. If they did not keep up, he could fire them. It was the Depression, and they had no other jobs to go to. He used that. Dwarkesh Patel had asked whether those early achievements came from the government bei

Robert Caro says no. He does not want another Robert Moses, a builder who is ruthless and who disregards public opinion. A city still needs someone who will find a way to get things built when many people do not want them built.

Such a person appears only every once in a while. Decades can pass when everything is stalled and it looks as if nothing big will ever be built again. Then someone comes along, and things get built. He calls that political genius. The two he is writing about are Lyndon Johnson on civil rights and Robert Moses on public works. People said those things could not be done, and they did them.

Dwarkesh Patel heard this as a case for a little Moses today. Caro sometimes w

Dwarkesh Patel argues, drawing on Robert Caro's Master of the Senate, that Lyndon Johnson took control of the Senate by turning a job almost nobody wanted into a hold on favors, the floor, and the deals.

The Senate had been set up not to be led. The Library of Congress told Johnson the majority leader had one formal power, the right to be recognized first. The two leaders before him both lost their own seats. He got the job at 44, in a body whose average age was around 60, because others did not want it.

He then collected what senators needed. He could schedule a bill, move it out of committee, call Speaker Sam Rayburn about the House, and mention a colleague to a Texas oilman who gave cam

Dwarkesh Patel says no. Drawing on Robert Caro's biographies, he argues that Lyndon Johnson showed little coherent political ideology for most of his career and fitted himself to the conservative Southern Democrats who held power in the Senate.

When Johnson entered the Senate, the most powerful men there were the Southern Dixiecrats, and the new senator molded himself to them. While writing Master of the Senate, volume three of the Johnson biographies, Caro asked former Georgia senator Herman Talmadge what impression freshman senator Lyndon Johnson gave of the proper relationship between white people and black people. Talmadge answered, master and servant. Caro asked whether Johnson had any

Dwarkesh Patel says, drawing on Robert Caro's The Power Broker, that mayors and governors could not fire Robert Moses. The statute, a stack of separate jobs, and his popularity all stood in the way.

The Triborough Act allowed removal midterm only on stated charges after a formal hearing. Caro writes that Moses was money honest, so there were never charges to bring. Terms ran six years, and he held 12 posts at once, some filled by the mayor and some by the governor. Nobody could strip every post at once, and removing one still left the most powerful man in New York as an enemy. He was popular for most of his career too. In the 1930s and 1940s, New York high schools voted him the most admired

Dwarkesh Patel argues, drawing on Robert Caro's The Power Broker, that Robert Moses got power without winning an election by writing the laws that governed him and hiding new powers inside them.

Moses was known as the best bill drafter in Albany. The Triborough Bridge Authority Act of 1933 was supposed to build two bridges, collect tolls until the bonds were paid, and then shut down. Bonds could run at most 40 years. In a subsection of Section 9 he added one sentence: the authority could, from time to time, pay off old bonds by issuing new ones for any other corporate purpose. A fresh 40-year note could fund a different project. Connecting roads, and facilities not inconsistent with the pro

Dwarkesh Patel says Lyndon Johnson passed a civil rights law by spending seven months on a deal that gave unrelated factions something each of them wanted. He needed a bill that year, even a weak one, because he wanted to become president, and no Southerner with his record of voting against civil rights could hope to win the Democratic presidential nomination in 1960.

Southern states require jurors to be registered voters, and an all-white Southern jury would never convict a white registrar for turning away Black voters. Any version one side could accept, the other would reject, and a less creative legislator would have backed down. Western senators got a vote on a federal dam they had want

Pushkar Ranade says Intel will invest more in chipmaking tools, and would have bought more a year ago if today's AI demand had been clear then.

The limit is not the fabs Intel already runs. It is the tools upstream in the supply chain. Dwarkesh Patel points to NVIDIA locking in memory and other parts with commitments reportedly in the hundreds of billions of dollars, and asks whether Intel should make the same kind of commitment so advanced manufacturing is not left short later. Ranade calls that a valid question. Had Intel known what it knows now, it would have invested more about a year ago. Knowing that now, it will invest more. Otherwise the tools end up elsewhere, including Taiwan, and

Pushkar Ranade, Intel's CTO, says the world's chip tools can about double leading-edge wafer output, not multiply it by ten, even if AI labs want far more chips.

ASML produces fewer than 100 extreme ultraviolet machines a year and does not plan to build significantly more than 100 a year by the end of the decade. That supply is split across Intel, TSMC, and memory makers such as SK Hynix. EUV is the bigger bottleneck, and it is not the only one. Once a flow such as Intel 18A is locked, the fab needs that specific tool from that supplier. Going from 18A to 14A, about 80 percent of the tool set can be reused, and the denser node still needs some new tools. Even with unlimited capital, extra w

Dwarkesh Patel says a gate-all-around transistor slices the silicon channel into thin ribbons and wraps the gate completely around each ribbon, so the gate can stop current from leaking when the switch is off.

A transistor is a switch. A positive voltage on the gate pulls electrons into a channel between the source and the drain, and a thin insulator keeps them off the gate. Remove the voltage and the channel disappears. As transistors shrank, current could still leak when the switch was off. The planar MOS transistor, demonstrated in 1960, touches the channel only from the top. A FinFET squeezes it into a vertical fin so the gate controls three sides. Gate-all-around also covers the bottom

Si Sheppard says the splits among conquistadors were personal rivalries inside one Spanish world, so the Aztecs could not use them to stay in power.

Dwarkesh Patel points to the larger force sent to arrest Cortés, and to Almagro, Pizarro's second-in-command, who came back angry about the spoils and besieged Cuzco. Spaniards fought each other while taking over a much larger empire. Sheppard's distinction is that those fights stayed inside a shared Castilian background. An outside crisis meant to pit them against each other would have failed, because a common enemy would have pulled them back together.

Native rivalries were older and ran between whole societies. After La Noche Triste, Moctez

Si Sheppard says the Aztecs never made bronze, the Inca world did, and the two empires did not pass that knowledge to each other.

In Mesoamerica the tools of war were stone. The Aztecs could smelt silver and gold, which drew the Spaniards in, and they could not smelt iron to hold them off. The Tarascans, to the west, could smelt copper and put it on arrow points, and the first Aztec invasion of their territory was a disaster partly for that reason. Tin exists in the Americas. Mesoamerica still never turned copper and tin into bronze. The Inca did have bronze. Bronze cannot stand up to iron, and it is better than stone or obsidian. Access to it, Sheppard says, would have given the Aztecs a h

Si Sheppard says Francisco Pizarro captured Atahualpa by copying Cortés: seize the person at the top of the hierarchy, in a place where horses and gunpowder could shock fighters who had never seen them.

Pizarro, a distant cousin, met Cortés, took the playbook, and sailed down the west coast of South America into Peru. The Inca realm was in civil war. The previous Sapa Inca had died in 1525, and his sons Atahualpa and Huáscar had turned on each other. Atahualpa had just won and was heading to Cuzco to be enthroned when word came that strangers were seeking him. Sheppard says 168 men had entered the Andes. An envoy warned that they were not men of good faith and could be burned where they sle

Dwarkesh Patel says the Spanish conquest is a real case of what people now picture as an AI takeover: an outside force, whose motives the people already there do not understand, uses diplomatic cunning and better technology to take over a much larger and long-established order.

He is explicit that the analogy is not the subject of the hour. The conversation is about the history. The comparison is that a takeover by an advanced outsider would not be the first time a small arriving force outmaneuvered empires with millions of people. In his telling, a few hundred conquistadors won battles while outnumbered by enormous margins, sometimes without a single casualty on their side, and while most

Dwarkesh Patel says Aztec warriors tried to capture Cortés instead of killing him because their wars were built around taking victims for sacrifice, and that habit threw away the chance to finish him.

Cortés was knocked off his horse and left on the ground. Warriors moved to seize him for a sacrifice rather than kill him where he lay, and he got out. Si Sheppard agrees that Aztec fighting aimed to capture, not to kill, and that there were times when Cortés was in their hands and his companions still pulled him free because the captors were hauling him off. Status came from bringing captives back to Tenochtitlan. Prisoners even had a rank: four Totonacs were worth one Chichimec, because the

Dwarkesh Patel asks how anyone will know alignment is working during recursive self-improvement, and he wants a safety case that alignment is working before each next RSI rung. Noam Brown says there may not be time to know. New frontier models ship at most every two months, sometimes faster, while the models go from week-long tasks to month-long tasks and then to tasks that run for three months. Once an agent can work for three months and the next model still ships every two months, there is no way to test it across the full length of what it can do before that next release.

The gap is not only alignment. The product, the alignment, and the safety can all get worse over a stretch the tests

Dwarkesh Patel says each lab could have enough compute by the end of 2030 to run hundreds of millions of human-level intelligences, if the current rate of progress simply continues.

People are not taking seriously what continued progress implies once systems cross beyond the human horizon. Reasoning about smarter-than-human minds is hard, so the comparison is in human population sizes. A given level of compute supports a three times bigger effective population every single year, and compute keeps growing in the background. The end of 2030 is a late marker; the same scale could arrive much sooner. A few years after that, by the mid-2030s or earlier, each lab would hold many Earths worth of h

Dwarkesh Patel says a fully automated firm can make an infinite copy of its best talent, then spin extra copies down when a task no longer needs them.

AI minds can share context and merge knowledge more seamlessly than humans can. Hiring skips the hunt for new talent, because the firm spins up any number of copies that already have the right knowledge. It can also replicate the most effective parts of an organization, or replicate whole organizations that already work well. He was working from an essay he wrote a couple of years ago on what automated firms of human-level intelligences would look like.

Noam Brown adds one contrast with a human coworker: a firm that wants two copies of a per

John Schulman says a model good enough at learning in context would not need training on finance. It could read the books on the fly and work in the appropriate jurisdiction. Labs might still bake those domain skills into the weights with reinforcement learning so the model is more efficient at runtime. In practice, model providers go domain by domain and strengthen the highest-value domains and the most common skills. He counts that coverage as one reason the models have gotten so much better.

Dwarkesh Patel had asked why a lab would bake in PowerPoint if a model can learn on the job the way a person does. Before Schulman answered, Patel offered two candidates: the models are not there yet

Dwarkesh Patel says the bet AI labs are making is to scale RLVR training across millions of diverse environments and hundreds of domains. What emerges is an agent with skills he calls basic, or less than basic: staying persistent, triaging information and context, and eventually doing end-to-end work with other agents. In-context learning keeps getting scaled so the agent can run for a long time. What comes out functions like a drop-in remote worker over a week or a month. The learning happens inside simulations in a data center. The agent learns how to learn there, then gets deployed without learning from real-world deployment.

Charlie O'Neill says that picture fits the models labs deploy