Ep 386: Was 2025 a Great or Terrible Year for AI? (w/ Ed Zitron)
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this episode, Cal Newport is joined by AI critic and Better Offline podcast host Ed Zitron to analyze whether 2025 was a great or terrible year for AI. They systematically review the biggest AI stories month by month, from DeepSeek's January disruption to OpenAI's December "Code Red" crisis. (03:16) The conversation reveals troubling financial realities behind AI companies, with massive costs exceeding revenues and a shift from ambitious promises to desperate attempts at monetization.
- Main themes: Financial sustainability of AI companies, the disconnect between marketing hype and technical reality, the evolution of AI from revolutionary promises to practical limitations
Speakers
Cal Newport
Computer science professor at Georgetown University and author of "Slow Productivity" and "Digital Minimalism." Newport is known for his research on technology's impact on society and deep work practices. He hosts the Deep Questions podcast and writes extensively about technology criticism and productivity philosophy.
Ed Zitron
Host of the Better Offline podcast and writer of the "Where's Your Ed At" Substack newsletter. Zitron has emerged as one of the most informed AI industry commentators, known for conducting thorough investigative reporting including analyzing earnings reports, talking to sources within tech companies, and following financial data rather than just reporting on company narratives.
Key Takeaways
AI Companies Are Operating at Massive Financial Losses
Ed Zitron's investigative reporting reveals that major AI companies are spending far more on compute costs than they're generating in revenue. (92:00) For example, OpenAI spent $8.67 billion on inference costs through September while generating only around $4.5 billion in revenue during the same period. This represents a fundamental business model problem where costs increase with usage, making profitability nearly impossible at scale.
The Industry Shifted from Scaling to Post-Training Due to Diminishing Returns
By 2025, the traditional approach of making AI models better by simply making them larger hit a wall. (39:40) OpenAI's Project Orion, their attempt to scale up from GPT-4, failed to deliver expected improvements despite massive investment. This forced the industry to pivot toward "reasoning models" and test-time compute, which essentially means using more processing power to run existing models rather than training fundamentally better ones.
AI Agent Promises Were Abandoned as Technically Unfeasible
Despite 2025 being declared "the year of AI agents," OpenAI ended the year by deemphasizing agent development and declaring a "Code Red" to focus on making ChatGPT better. (100:00) The reality is that current AI systems cannot reliably perform multi-step autonomous tasks in real-world environments. What was marketed as revolutionary workplace automation turned out to be expensive, unreliable prototypes.
Desperate Revenue Strategies Reveal Industry Desperation
The launch of Sora video generation app and OpenAI's deals with companies like Disney represent desperate attempts to find new revenue streams. (84:00) These moves signal that core AI products aren't generating sufficient revenue to justify their development costs. The shift toward consumer entertainment applications shows the industry moving away from transformative B2B promises toward more traditional content monetization models.
Media Coverage Often Amplifies Marketing Rather than Technical Reality
Throughout 2025, technology journalism frequently repeated company marketing claims without adequate technical scrutiny. (73:00) Ed Zitron's experience of having technical stories about AI limitations rejected by multiple reporters highlights how industry access and advertising relationships can compromise journalistic independence. This creates an information asymmetry where the public receives optimistic projections rather than realistic assessments.
Statistics & Facts
- OpenAI spent $8.67 billion on inference costs through September 2025 while generating only around $4.5 billion in revenue during the same period, according to Ed Zitron's investigation. (92:00)
- Anthropic spent $2.66 billion on AWS in three quarters of 2025, with similar spending on Google Cloud, totaling over $5 billion in compute costs while generating approximately $5 billion in revenue. (83:00)
- The average household in the 1980s had their television on for 7-8 hours per day, measured by Nielsen audio meters, demonstrating how pre-internet media consumption paralleled today's smartphone usage patterns. (128:00)
