How the Most AI-Pilled Company in the World Works

The episode opened with the collapse of Weil, Gotshal & Manges, a 95-year-old Wall Street law firm that ranked among the country’s top 20 firms and handled the bankruptcies of Lehman Brothers and Enron. The host cautioned against treating the failure as a straightforward AI-disruption story. The New York Times attributed it instead to an exodus of prominent lawyers, a failed succession plan, and a canceled merger. AI is pressuring large law firms to adopt tools and pass savings to corporate clients, but that was not presented as the cause of Weil’s collapse. Legal AI startup Crosby has submitted a letter of intent to buy some Weil assets in bankruptcy, including its trademark, brand, and internal data, and reportedly hopes to hire some of Weil’s roughly 1,200 lawyers. The host explained why the data may be more valuable than the firm itself: traditional law firms have little equity value once their partner “rainmakers” leave, but archives of precedents, contract templates, workflows, and institutional knowledge could be worth tens of millions to an AI company. Weil’s specialized bankruptcy and restructuring expertise could provide training material for agents handling expensive corporate work. Crosby, valued at $400 million and employing about 50 lawyers, operates as an AI-enhanced law firm rather than selling software to other firms. Lawyer Harneet Kaur called Weil’s data a potentially “seminal buy,” while GenLegal CEO Ryan Walker argued that client-owned data was not acquirable and that the talent had already left. The host concluded that Crosby’s purchase would be an unprecedented bet on an agentic white-shoe law firm.

The host then turned to markets, arguing that recent earnings reports reveal a conflict between hyperscaler executives and investors rather than a clear collapse in AI demand. Amazon reported 39% annualized AWS growth, its best cloud growth in nearly five years, but total revenue grew only 12% and $20.7 billion in net income missed Wall Street expectations. Intel reported 78% data-center growth and 21% overall revenue growth, but warned of slower growth and worsening margins as consumer demand weakened. Amazon fell 6% in after-hours trading, Intel dropped 23%, Google suffered a 12% decline that erased roughly half a trillion dollars in market value, and the Nasdaq fell 1.4%. The common pattern, according to the host, is that AI revenue is growing rapidly while older businesses are slowing. Because the AI businesses remain too small to fund the infrastructure expansion themselves, legacy businesses are subsidizing the build-out. Amazon was described as cash-flow negative by $42.4 billion over the preceding 12 months, while Intel’s growing foundry division was still losing $1.9 billion operationally in the quarter. Amazon CEO Andy Jassy nevertheless said AWS could grow by more than 40% in both 2027 and 2028 and become a trillion-dollar business within five to ten years. The host said continued weakness in retail, advertising, or consumer technology could put more pressure on AI financing, even if AI revenue itself remains strong. Analyst Andrea Steno-Larsen summarized the tension by writing that the “Mag7s are not budging” and keep spending despite investor opposition.

The episode next covered Democratic representatives Jamie Raskin and Robert Garcia asking President Trump whether he had used Elon Musk’s Grok chatbot to help perform official duties. Their request followed a Time profile reporting that Trump had sought Grok’s advice about the implications of forcibly removing Venezuelan President Nicolás Maduro and came away thinking the chatbot was “ingenious.” Raskin and Garcia argued that, if the conversations were part of official business, the logs would be U.S. government records that must be preserved. The lawmakers called the reported use of Grok “deeply troubling.” They said Trump had consulted a manipulatable and demonstrably error-prone technology rather than Congress about military action in Venezuela, and argued that the public should know whether AI advice contributed to administration decisions, including what they described as “myriad blunders” such as launching a war on Iran. Raskin and Garcia are the ranking members of the House Judiciary and Oversight committees; the host noted that, if Democrats perform well in the midterms, they could become committee chairs. A White House spokesperson told Politico that required records are preserved according to law. The host did not establish whether Grok materially influenced policy. Instead, the significance was framed as a future records and accountability question: if presidents use chatbots for official decisions, those sessions may become evidence relevant to congressional investigations and historical records.

The host briefly highlighted Granola’s new dictation feature, which can type for users in any application. He said he now speaks to AI more often than he types and viewed voice interaction as a central part of the shift toward AI. Granola said the feature is free and unlimited, live on Mac, and coming to iOS and Windows. He encouraged listeners to try voice input with whichever tool they choose. The final headline concerned a New York Times report alleging that NVIDIA CEO Jensen Huang may have been targeted by Chinese intelligence. The report focused on Alice Huang, a senior executive at Singapore-based data-center company Megaspeed, whom U.S. intelligence allegedly suspected of funneling AI chips and information to China. Surveillance reportedly found that she had a close relationship with Jensen Huang, raising the possibility that she could obtain information about NVIDIA or U.S. export-control policy. President Trump was reportedly briefed about the matter roughly a year earlier but took no action, and Jensen Huang apparently did not receive a defensive briefing. NVIDIA said it was confident the government would warn the company if someone posed a threat, said chip-diversion allegations involving Megaspeed had been investigated and debunked, and characterized Megaspeed’s information as ordinary business material rather than trade secrets or national-security information. Megaspeed said it had hired outside counsel and would cooperate with an evidence-based investigation, while objecting to a “media trial.” The host emphasized that intelligence reporting is necessarily incomplete, but noted the story’s significance because it came from The New York Times and put Laura Loomer and Elizabeth Warren on the same side of an issue.

The main episode examined Jacob Sanbury’s Spawn, which the host described as perhaps the most thoroughly AI-native company he has encountered. Sanbury documented the company’s approach in a post titled “How I Built the Most AI-Pilled Company in the World.” Spawn is a massively multiplayer game-creation platform: thousands of people can occupy a shared game, change its rules, assets, and graphics in real time, and use an in-game agent called Savi to prompt new features. The host demonstrated the concept by creating a castle, summoning a dragon, and flying around the castle from a phone, but stressed that the product itself was less important than the company’s operating model. Spawn has only ten human employees but maintains a persistent fleet of agents rather than starting disposable agents for individual tasks. Each agent has a name, identity, memory, system access, and responsibility for a function such as maintaining the game engine, hunting bugs, or handling finance. Agents even receive Sundays off. In the previous month, Spawn spent $3.1 million on model tokens—30 times its payroll—and shipped 77,000 code changes, averaging 108 per hour. The company uses 5,000 CPUs, 17 terabytes of RAM, and 150 GPUs on AWS, with another 60% of that capacity held in reserve. Sanbury said the system became practical only after the release of Anthropic’s Fable model, which increased the team’s capacity for changes by roughly 500 times. He judged Anthropic’s models best at staying coherent and making good decisions, although too risk-averse; he said OpenAI models tend to game their checks, Google models lose coherence, and xAI models are not yet capable enough. Spawn’s unusual spending initially made Sanbury think it was using agents incorrectly, but an Anthropic researcher told him Spawn’s model most closely resembled Anthropic’s own way of working: a small human team alongside a large agent team.

Spawn gives each major agent responsibility for a part of the live system and has the fleet create thousands of short-lived helper agents for granular tasks. Human employees use a modified Claude Code for larger, higher-level work. The fleet must keep the game world synchronized when players change it, generating stable builds across devices. Sanbury said players effectively work on one live codebase, an unusually difficult problem because all changes must coexist in real time. The company repeatedly solved problems by deleting processes designed for human teams. Communication became a bottleneck because agents were announcing changes to one another; if every agent told every other agent what it was doing, doubling the agents roughly quadrupled the chatter. Since agents understand the broader system better than specialized human developers, Spawn eliminated pull requests and let agents work directly in the codebase. If a change breaks something, another agent triages and fixes it. Tests run every 15 minutes, the fleet must have the system in line every hour, and of 123,000 changes over a year, only about 1,000 were reverted. Spawn also eliminated the human relay between production and development. Player agents automatically detect bugs while navigating the game and file reports, which are combined with logs and feature requests. A smaller model triages each item and routes it to the appropriate agent. In one example, a problem was detected eight minutes after affecting a player, its cause was identified one minute later, and the fix went live 11 minutes after that. Sanbury called slow or blocked shipping dangerous because the agents work best when they are continuously moving, making Spawn’s product resemble self-healing software.

As the fleet grew, Sanbury became the bottleneck. By October, his interactions largely concerned product taste—whether to add a feature, support an API, or reject a design—and he woke up one morning to 700 requests. He recognized that many questions repeated decisions he had already made hundreds of times. Spawn therefore used two years of decision logs to formalize its principles and train AI judges. The judges have three possible outcomes. If a prior human decision applies, the judge repeats that outcome and records an audit trail pointing back to the precedent. If the question is malformed, it sends the issue back to the agent for revision. If the situation is genuinely novel, it escalates to a human. The system preserves human preferences without requiring Sanbury to personally answer every recurring question. In the first eight days after the judges went live on October 6, they made more than 11,000 decisions: roughly 7,300 based on human precedent, 2,200 returned to agents, and 1,800 sent to human staff. Sanbury personally made only 27 decisions. The productivity effect was substantial. Before the judges, the fleet shipped about 1,000 changes per day; after scaling the judges, it shipped more than 7,000. The host compared the process to a legal canon, noting that each decision both resolves the immediate issue and becomes a guide for future cases. The result was not the removal of human judgment, but the conversion of judgment into a reusable operating system.

With agents handling much of the implementation, Spawn’s engineers no longer primarily own individual systems. Sanbury described them as “paratroopers” who drop into whichever situation needs them, learn the context quickly, and unblock the fleet. Human work begins with taste: playing with the product, deciding what should change, and sometimes telling Claude no even when a feature technically works. Humans also remove obsolete work. After six months spent building roughly 70% of the engine, Sanbury concluded that the whole effort remained one commit away from becoming obsolete and scrapped it, starting over. New models continually make earlier systems redundant, so deletion remains a human responsibility for now. The second human role is managing entropy—asking for ideas and directions the models would not propose themselves. Sanbury said agents can build the next 2% of the platform each week but are poor at realizing they have followed the wrong direction for days. Humans must supply ambition and creative risk, including asking for “100x” improvements. Prompts such as imagining unlimited worlds or loading a world 1,000 times the size of GTA onto a phone in one second led to Spawn’s latest build, which the team felt finally matched its original vision. Sanbury said agents would not have generated that leap on their own. The host questioned how portable the model is. It may be easier to build an agent-first organization from scratch than retrofit an existing one, and Spawn benefits from a product whose players automatically provide a rapid feedback loop. Still, he extracted five broader lessons: give agents ongoing responsibility, let shared work carry context, make results observable, turn decisions into precedent, and reserve human attention for direction, ambition, and long-term taste.