The Next Wave of Enterprise AI

The episode opens with a detailed breakdown of Trump's AI executive order, which had a bizarre policy journey. A draft order requiring AI labs to submit advanced models for 90-day government review before release was scrapped hours before a signing ceremony after former AI czar David Sacks called Trump to intervene. The final signed order is nearly identical to the draft, except the review window was shortened from 90 to 30 days, testing remains voluntary, and there is no mechanism for the government to block model releases. The order also includes explicit language forbidding the creation of a mandatory licensing or permitting regime for AI models. The NSA is assigned primary responsibility for model testing, and a cybersecurity clearinghouse is established under the Treasury. The order was signed quietly with no fanfare, contrasting sharply with the originally planned high-profile ceremony. Reactions were deeply divided: the New York Times framed it as a shift away from a hands-off approach, while the White House's OSTP pushed back calling that characterization lazy. Critics like Dean Ball warned it lays infrastructure for a future licensing regime, while Steve Bannon celebrated it as a first step toward mandatory regulation. Separately, Anthropic expanded access to its Mythos model through Project Glasswing, adding 150 partners across 15 countries in sectors like energy, healthcare, and communications. The model is reported to be extremely expensive, with testers burning through millions of dollars in tokens quickly, though Anthropic is currently subsidizing use. The episode then pivots to enterprise AI, analyzing dueling events from OpenAI and Microsoft. OpenAI released a report called 'The Next Era of Knowledge Work' centered on Codex, which now has 5 million weekly active users. Notably, the fastest-growing user segment is non-technical knowledge workers, adopting Codex three times faster than developers. OpenAI identified three core frictions in knowledge work—finding inputs, information coordination, and approvals—and positioned Codex as the 'factory redesign' to address them. New Codex features include Annotations for precise document interaction, role-specific plugins bundling apps and skills for functions like sales, data analytics, and investment banking, and a new Sites feature that lets users turn any Codex artifact into a shareable web app. The host argues Sites represents a new core knowledge work primitive akin to documents or spreadsheets. On the Microsoft side, Build showcased seven new MAI models, headlined by MAI Thinking 1, a one-trillion-parameter mixture-of-experts model. While its raw benchmark scores drew mixed reactions—strong in some areas, weaker in agentic coding—Microsoft's stated strategy is not raw performance competition but cost optimization through frontier tuning. Microsoft claims MAI outperformed GPT-5.5 on quality for McKinsey-specific tasks at 10x lower cost, positioning its models as enterprise customization tools rather than general-purpose frontier models. The host concludes that the second half of 2026 will be defined by enterprises wrestling cost-effective, scalable approaches out of the AI opportunities unlocked in the first half of the year.