The Ad Hoc AI Licensing Regime

The episode opens with context around the Fable 5/Mythos situation, explaining that Senator Mark Warner's comments about the NSA were misinterpreted—the NSA hadn't been breached, but rather conducted red-teaming exercises revealing greater-than-expected capabilities. This sparked a week of speculation and market volatility, with prediction markets initially spiking above 60% likelihood of Fable 5's return when reports suggested the White House was engaging better with Anthropic co-founder Tom Brown than with CEO Dario Amadei.

The substantive outcome, however, is the emergence of what the host calls an "ad hoc, informal, unaccountable, and seemingly not particularly technically competent licensing regime." OpenAI's Sam Altman revealed that GPT-5.6 would launch in limited preview rather than open access, with the U.S. government approving access customer by customer. The host and multiple quoted experts (Neil Chilson, J.V. Mausowitz, Andrew Curran) express strong criticism of this approach, arguing it is arbitrary, non-transparent, makes no one happy, and doesn't actually slow development—only public release timelines—thereby widening the gap between what labs have internally and what the public can access.

The episode then shifts to positive developments in the broader AI ecosystem. Smaller organizations and startups are experimenting with Z.AI's GLM 5.2 model with positive results. Google reported that Gemma 4 reached 200 million downloads, indicating demand for lower-cost alternatives. Anthropic's Claude Tag integration in Slack demonstrates how UI/UX patterns and multi-player AI implementations are becoming as important as model advancement itself, with 65% of code at Anthropic now originating from Slack conversations. Will Brown from Prime Intellect notes a significant shift toward enterprises wanting to secure compute and post-train their own models in-house, particularly on open-source alternatives like GLM 5.2, suggesting organizations are beginning to understand how open source can win.

A KPMG survey found that CEO-led AI efforts were 3x more likely to produce ROI, suggesting that organizations with strong leadership engagement pursue more complex, integrated, or post-trained approaches rather than defaulting to closed-source state-of-the-art models.

Market signals remained mixed: early-week bearish sentiment about an AI bubble reversed when Micron delivered strong earnings and reinforced that supply chain shortages for AI infrastructure would persist. OpenAI is reportedly planning to delay its IPO until next year.

The episode concludes with post-recording updates: the U.S. lifted its block on Mythos for approximately 100 selected partners (major companies and government agencies), prompting concerns about gatekeeping and sparking passionate responses about open-source model development as a counterbalance. GPT-5.6 launches today in limited preview as government-requested, with Sol (a new efficient model at GPT-5.5 price) and Terra (5.5-level performance at half price) also announced. Sam Altman acknowledged this isn't optimal but expressed hope for faster general availability and a more transparent process. The host closes with Rune's perspective from OpenAI that early government engagement, while procedurally imperfect, is inevitable and better than too late, though concerns remain about non-Americans being left behind.