DeepSeek Panic, US vs China, OpenAI $40B?, and Doge Delivers with Travis Kalanick and David Sacks

The episode opens by promoting a separately released conversation with Ray Dalio about his new book on how countries go broke. The book is described as covering the historical context of debt cycles in different countries. It ends with a recommendation that the US bring its net deficit to roughly 3% of GDP, counting all expenses including interest. The speaker presents this as a recommendation to the administration and calls it timely given the change in administration.

The guest is then introduced as the cofounder and CEO of Cloud Kitchens and the cofounder and former CEO of Uber. He is said to be returning from a media hiatus spent working on Cloud Kitchens. The opening banter is mostly a running joke about one host's status as an early Uber investor.

Cloud Kitchens is framed around the question of what food will look like in the future. The pitch starts a hundred years out: very high-quality, very low-cost, highly convenient food made and delivered by machines, tailored to each person's dietary and taste preferences, and cheap enough to approach or beat the cost of going to the grocery store. The company is then described as working back from that horizon to 20 or 10 years out, through three businesses: real estate, software and robotics.

you do to the kitchen what Uber did to the car.

The business is contrasted with Uber's mix of bits and atoms. It is described as having "5 times more atoms per bit" and as heavy industrial work, compared to some of Elon Musk's companies, where the technology is interesting but the physical side has to be ground out.

Asked whether people will still cook, the answer is that cooking will survive as a hobby. The analogy given: liking horses without riding one to work. Cooking is called soulful and very human. The example offered is a parent who gets home late and can feed the kids a nutritious meal without cooking and without going to McDonald's. The claim is that food will be high quality, convenient and low cost all at once, and hyper-personalized the way internet content is.

The company is said to be past the lab stage with a machine called the Bowl Builder, which makes bowls across different cuisines. The speaker notes they are not working with Chipotle, Cava or Sweetgreen specifically and uses them only as analogies. The company built test brands modeled on those concepts and ran a real restaurant while building the machine, to prove the machine works. Customers are now touring it, and a rollout with 5 customers is planned for April.

The workflow described: restaurant operators come into one of the company's delivery-only kitchens (the company says it has tens of thousands worldwide) and prep food in the morning, then leave. Customers order online through DoorDash, Uber Eats and similar apps and build their own bowl. The machine dispenses each ingredient, hot or cold, plus sauce, then lids the bowl and puts it in a bag with utensils. The bag is sealed and comes out on a conveyor. A machine moves the bag to a locker at the front of the facility. The courier holds a phone app up to a camera, and the correct locker opens.

The operational point is that the on-demand grind of restaurant work goes away. Prep happens separately from when orders come in, and the machine handles final assembly, or plating.

One host asks whether diners could share their physiology with Cloud Kitchens so their food is always optimized for them. The answer redirects: Cloud Kitchens serves restaurants, not diners directly. Diners would share preferences with Uber Eats, DoorDash, Sweetgreen or another brand, while Cloud Kitchens runs behind the scenes as infrastructure. The speaker likens this to AWS or NVIDIA, but for food. The company's customer promise is quoted as serving "those who serve others."

The hosts sketch a version in which a brand asks for an encrypted hash of a diner's dietary needs or lipid panel, or for access to Apple Health, and Cloud Kitchens' back end handles the customization. The guest-side description adds that each bowl's label lists the grams of every ingredient plus a photo taken before lidding, and this can be sent to the customer while a courier is bringing the order.

On the MAHA movement and the food supply, the answer predicts supply chains getting "super wired up." Automation would move upstream from assembly to prep, then to distributors such as Sysco and US Foods, then to mechanization on farms. Traceability is the main example: knowing what kind of wheat went into a dish, whether it was really organic, even which field it came from. The speaker endorses MAHA enthusiastically and mentions buying merch from RFK Jr.'s website.

One participant describes an earlier bowl-building system from around 2015. Food was prepped in a commissary, loaded in bulk into canisters, and placed in a dispenser with 30 slots. The canisters moved down the device and opened to assemble bowls of rice, beans and other ingredients. The whole process was automated. The first automated store was being built when the participant took a medical leave of absence, and the company never put the machine into production, although it had working demos.

The same participant says they had a term sheet with Chipotle about nine years earlier to put the system into Chipotle stores, and were in early talks with Sweetgreen, which has since built its own system. The argument is that much of quick-service dining can be reduced to a bowl-based system and automated. This is called necessary when there is a labor shortage or labor price inflation. The early 20th-century New York automats are cited as a precedent: a commissary behind a wall, a coin, a turned knob, and a plate of food.

Concrete figures from that period: an 800-square-foot restaurant took in 3 million a year in revenue with a handful of staff and served about 800 people an hour during the lunch rush. A host clarifies that this was one location.

The automat is called classic "artificial artificial intelligence," in the style of the Mechanical Turk. The discussion then raises a specific limit on automation in quick-service restaurants. Existing locations are laid out for humans working in exact processes, with every square inch planned. Replacing a chain's front line means demolishing it and installing a new machine.

The listed costs: large capital spending, a store closed for two to three months, economics that start to fail, and human staff still needed in the location. Cloud Kitchens' answer is the delivery-only model, where kitchens and machines are designed for each other from the start, so these retrofit problems don't arise.

The newly appointed AI and crypto czar, addressed as Sacks, describes his first week in the administration as amazing and says it is hard to believe only a week has passed. The office described is in the old Executive Office Building, also known as the Eisenhower Building, with a pass to walk to the West Wing inside the gated White House complex. The Navy Mess is mentioned as a White House commissary that is just reopening, along with an invitation to bring other hosts to lunch there.

The main topic is the reaction to DeepSeek, a Chinese AI startup whose R1 language model is described as roughly on par with leading Western models such as OpenAI's o1. DeepSeek claimed it built the model for $6 million on only 2,000 GPUs. The host cautions about trusting claims coming out of China, contrasting them with OpenAI's reported $80 to $100 million cost to train GPT-4 and a claimed billion dollars to train GPT-5. The host states that $6 million is "about 7% of the cost of GPT 4." US export restrictions on NVIDIA H100s to China lead to a debate over whether DeepSeek actually has them.

Market reaction: NVIDIA had the worst day in stock market history by total market cap lost, down 17%, or $600 billion. TSMC, ARM and Broadcom also fell. The host also raises accusations that DeepSeek distilled OpenAI's model, calling them "kind of rich" given that OpenAI itself has been accused of taking other people's content.

The response draws on conversations with many of the top people in AI. It notes that a model release rarely becomes a global news story or wipes out a trillion dollars of market cap in a day. Two features are named as the reason: DeepSeek is Chinese, which brings in US-China competition, and it open-sourced R1, which brings in the open versus closed source debate. Without either one, the argument goes, the story probably would not have been as big.

Audience motives are offered as fuel. Much of TikTok's audience is international, and some of it liked the idea of the US getting a comeuppance in the AI race. People who favor open source or dislike OpenAI cheered a model that gives away OpenAI-like capability at one twentieth the cost.

The analysis separates base LLMs from reasoning models. Base models include GPT-4o and DeepSeek's V3, launched a month earlier, and are likened to a smart PhD who gives a quick answer. Reasoning models are built with reinforcement learning, a separate process from pretraining. They are likened to a PhD who goes off and does the work, breaking a complex problem into smaller ones and solving them step by step through chain of thought. The coming generation of agents is said to rest on this idea.