Why Businesses Can’t Budget for AI

Luke Vargas introduces the episode with a roundup of major headlines, including the Pentagon's plan to livestream the execution of Nidal Hasan, the Fort Hood shooter, and SpaceX's move to enter the wireless carrier market by acquiring spectrum licenses, which caused shares of US telecoms to dip. Vargas also mentions the challenge businesses face in estimating AI token costs, promising a detailed exploration of this issue later in the show.

The episode reports that Navi Pillay, a former UN Human Rights Chief and current judge on the International Court of Justice, has been awarded the Nobel Peace Prize. The Norwegian Nobel Committee praised her for her lifelong commitment to international justice, including her work against crimes related to apartheid in South Africa, the Rwandan genocide, and war crimes in Palestinian territories. The award underscores the significance of international cooperation and justice according to the committee.

Defense Secretary Pete Hegseth announced that the Pentagon plans to livestream the firing squad execution of Nidal Hasan, the former army officer who admitted to killing 13 people in the 2009 Fort Hood shooting. President Trump approved the execution, marking the first military execution in over 65 years, independent of recent issues with lethal injection drugs. The execution is scheduled for December 3 at Fort Hood. Hegseth emphasized the decision's gravity, asserting that it should serve as a serious warning for such crimes.

An incident in New York involved an immigration officer shooting an injured man during a targeted operation to arrest an immigrant. The man, who was in a car with a five-year-old—who was unharmed—was hospitalized and survived. Mayor Zoran Mamdani criticized the incident, saying he had been negotiating with President Trump and Secretary Mullen for the man's release, but these efforts failed. The DHS provided limited details, and the White House did not respond to requests for comment.

The episode discusses overnight stock movements reacting to recent tech news, focusing on AI chip companies like Micron, which saw their stocks fall after reports suggested that OpenAI’s revenue was below expectations. The market remains unsettled by uncertainties about revenue guidance and the valuation of AI startups, exemplified by the controversial IPO plans of firm Firmness in Australia. The episode also covers SpaceX’s strategic move into the wireless industry through licensing deals, which initially caused telecom shares to drop despite analysts suggesting no immediate threat because of the high costs and time required for SpaceX to establish a significant presence.

Stephanie Stamm discusses the difficulty businesses face in predicting AI costs, focusing on the challenges of estimating token usage. She explains that rapid industry evolution, model selection variability, and unpredictable costs make accurate budgeting nearly impossible. A study from Stanford, CMU, UC Berkeley, and Microsoft shows that cheaper models can sometimes cost more tokens depending on the task complexity, highlighting the inefficiencies in current approaches.

Stephanie describes how she visualized AI models’ decision processes, illustrating the back-and-forth decision-making that CPU tokens are based on. Using a complex prompt run through two different Google models, she shows how the more capable Gemini 3.1 Pro efficiently solved the task at a lower cost, while the less capable Gemini 3 Flash used far more tokens and failed in some cases. The example underscores the importance of selecting the right model for the right task to avoid unnecessary costs.

Experts advise that companies should train employees on using appropriate models for specific tasks, set limits on token usage, and continuously adapt to evolving AI technology. Stephanie emphasizes that the field is moving so quickly that accurate forecasting remains elusive, and businesses must remain flexible. The rapid pace of change makes it difficult to create long-term budgets, and companies are encouraged to experiment and adapt as new models and methods emerge.

Catherine Bindley reports that AI’s influence extends beyond business into everyday speech, especially among teenagers and professionals. She highlights a trend called 'prompt speak,' characterized by contrast-framing language like 'it's not this, it's that,' which AI models favor. This way of speaking is increasingly seen in meetings and group chats, with some colleagues even calling out colleagues for overusing it. The phenomenon raises concerns about diluting natural language and making conversations more robotic, though it might help in efficiently communicating with AI tools.