How AI Is Changing What Work Means

The episode opens with Meta’s Muse reaching 10 million U.S. downloads in 40 days, a pace achieved by only about a dozen apps. The other apps were almost entirely major mobile games such as Pokémon GO, Fortnite, and PUBG; the only two non-gaming examples cited were Disney+ and Meta’s Threads. ChatGPT, Claude, and Gemini did not grow as quickly as mobile apps because they launched first on the web and added mobile applications later. Meta devoted 39% of its October advertising spend to Muse, up from 23% in September, making it the 21st-largest U.S. brand by impressions for the month. Sensor Tower reported seven-day retention of 30% and 14-day retention of 23%, compared with approximately 25% for ChatGPT at both points and roughly 5% for the average iOS app after 30 days. Daily active users reached 2.5 million, up 15% week over week. The host argues that comparisons with Instagram, Facebook, and Google are unfair because those are unusually “sticky” products, but says the important unanswered question is how Muse will fit into users’ lives over time. A post by Signal on X framed the problem as a third category of retention. Desire-driven products such as TikTok, Instagram, and games provide an immediately rewarding habit, while dependency-driven products such as email, calendars, and banking become infrastructure. Personal agents currently sit between those categories: they are not reliably entertaining or indispensable, and one failure can be enough to make users abandon a utility-oriented task.

Muse also faced a security problem in its sharing feature. Shared conversations, searches, and images were placed behind URLs that were supposed to be difficult to guess, but a design choice made the links easy to discover. Researchers from security firm Clareock told Reuters and Ars Technica that they enumerated 6,800 Muse sharing links and opened several hundred. They found information that appeared unintended for public access, including a small business owner’s loan application, a real-estate investor’s internal performance report, and a national marketing company’s internal revenue review. Meta temporarily removed the feature and said it was working on a fix. The host then turns to AI agents in cars. Gemini in Android Auto had previously handled voice controls, maps, audio, text messages, and web searches, but it was not connected to a user’s broader personal context and reportedly often failed at basic tasks. The new version can access Google-based services and, for example, determine whether a driver will miss a flight by checking email details and traffic conditions. Google also says Gemini can give directions to a coffee shop based only on a photograph of a friend drinking espresso there the previous week. The host is unsure whether these integrations will become indispensable or whether people will simply use the phone already sitting on the car console. Still, he expects companies to keep experimenting: Tesla has integrated Grok into its vehicles, and he assumes other automakers will follow.

An updated draft of Anthropic’s financial disclosure, obtained by The Wall Street Journal, triggered concern over an $80 billion net loss in the first half of the year and as much as $750 billion in compute spending over the coming years. The host explains that approximately $72 billion of the loss came from the accounting treatment of stock grants and related instruments, leaving an actual operating loss of $8.7 billion for the first half. Financial podcaster Jack Farley said the accounting loss would only be worrisome if it were economically real; with convertible notes, warrants, and a dramatically higher market capitalization, the paper loss from being short calls is not a fundamental operating loss. The larger commitment is the compute spending. About $450 billion was reportedly firm and uncancellable, payable over seven to ten years, while the larger $750 billion figure may include softer commitments to “neo-clouds” that still have to build facilities. Anthropic reportedly booked $16.3 billion in revenue during the first half, despite previously claiming a $65 billion annualized run rate by July. It later said it added another $16.8 billion in unaudited third-quarter revenue, representing 45% quarter-over-quarter growth. The host says the disclosure mostly confirms what observers already knew: Anthropic has issued substantial stock compensation and expects to spend more than half a trillion dollars on compute. Like a potential SpaceX IPO, the offering will be judged primarily on beliefs about the next decade rather than recent quarters. Investors have reportedly pushed back on a rumored $2 trillion valuation and said they would be more comfortable around $1.5 trillion. OpenAI is simultaneously trying to close a reported $30 billion fundraising round at a $1.4 trillion valuation before the Anthropic IPO, intensifying the competition for capital.

The episode’s central theme begins with OpenAI’s acquisition of Orchestra Bio, a four-year-old AI biotech startup that last raised $20 million. Orchestra develops agents for drug-discovery workflows, focusing less on inventing new compounds than on operational tasks such as monitoring compound manufacturing and tracking clinical trials. Its team will join work on GPT Rosalind, OpenAI’s life-sciences model, which began as a research preview in April and is now a commercial product. OpenAI also offers Rosalind Workbench, a collection of agents that connect the model to scientific instruments and analyze results. OpenAI describes its goal as helping scientists generate ideas, design experiments, interpret findings, and coordinate the work required to advance treatments. The company emphasizes partnerships with existing pharmaceutical organizations, including Novo Nordisk and Moderna, rather than building a complete biotech company itself. In the episode’s framing, OpenAI is developing tools for the industry. Anthropic is pursuing a more direct model. Reuters reported that it quietly created its own biology laboratory, a claim later confirmed by its head of life sciences, Eric Cudera Abrams. He said real laboratory work remains the final test for biology and described Anthropic’s approach as similar to a conventional biotech company: some work happens in its own facilities and some with external partners. The lab reportedly has a limited remit, but its existence raised broader questions about whether frontier AI companies will eventually perform the work their customers currently do rather than merely supply intelligence to them.

Citrini Research, better known for financial and macroeconomic analysis, began the experiment after looking for investment opportunities outside crowded AI infrastructure markets. The firm saw life sciences as a neglected sector: funding had suffered after the pandemic boom, returns had been poor for roughly half a decade, and AI’s possible impact on drug discovery had not yet been reflected in valuations. Because the team lacked biology expertise and did not believe expert interviews would resolve the disagreement between “AI has not created an important drug” and “AI will cure all diseases within five years,” it decided to run an experiment rather than simply ask for opinions. Working with Anthropic, Citrini began with a 35-page prompt previously tested by Anthropic, modified it, and used Claude to design proteins targeting scorpion neurotoxins. Three hundred subagents generated 318,000 candidates, from which 211 promising options were selected for testing. An independent laboratory found that 17 proteins bound to four different scorpion toxins. The proteins had not yet been shown to neutralize the venom, so the result was an initial binding result rather than a demonstrated antivenom. Citrini said the project cost under a couple hundred thousand dollars. The target was chosen because scorpion stings affect about 1.2 million people each year and cause roughly 3,200 deaths, disproportionately among children in places with limited medical infrastructure. Anthropic argued that developing treatments for such neglected diseases is usually unattractive because the treatments are expensive to produce and unlikely to recover development costs. AI could change that calculation by lowering the cost of early discovery. Critics objected that protein binding had already been demonstrated in the literature and that neutralization was the meaningful test; Citrini acknowledged the limitation but said it did not want to delay publication for another three months.

Citrini’s experiment also showed that speeding up digital research does not eliminate the physical steps needed to make a treatment real. The firm said its virtual biologists compressed roughly a year of work into a couple of days, but DNA still had to be synthesized, proteins produced and purified, toxins obtained in usable form, and experimental controls established. The figures cited in the episode were 64 hours for dry-lab design, 39 days for results, and an estimated four to seven years for release. Former OpenAI researcher World of Pooh placed the first two charts together and said they “hurt my soul,” emphasizing the distance between computational invention and an approved product. The host draws three conclusions. First, AI labs may eventually compete with their customers, but this will not happen all at once across entire industries. Even adding an internal wet lab leaves a vast network of people, institutions, testing processes, and regulatory steps between an AI-generated idea and something usable in the world. Second, medical breakthroughs may not automatically repair public opinion about AI. The host expects AI to contribute through many incremental discoveries, especially for neglected conditions, rather than one miracle cure. He also notes that public distrust of Big Pharma suggests that even curing cancer would not guarantee a favorable reaction. The third conclusion is about the meaning of work. Citrini began with its ordinary investment process—finding underrecognized opportunities—and then used AI to perform hands-on biological research that would previously have been inaccessible to financial analysts. The firm remains a financial research company, but the host expects more “orthogonal” work of this kind, in which AI enables people to explore adjacent domains and changes not only how efficiently they perform their core jobs, but what work they are able to attempt.