Vibes & Benchmarks Ep 05: Should the government own half of OpenAI?

Episode 5 of Vibes & Benchmarks, Outset Capital's weekly AI news podcast, is up. Ali Rohde and Josh Albrecht break down the week's AI news.

Topics this week:

• Three trillion-dollar IPOs in one year. OpenAI filed to go public, joining Anthropic and SpaceX. Pinnacle of American capitalism or its Achilles heel? Josh's take: companies that IPO this late lock retail investors out of the upside — you can't 100x from a trillion.
• Should the government own a stake in OpenAI? Sam Altman's public-wealth-fund pitch — OpenAI donating equity to the government — versus Bernie Sanders's 50% proposal. Josh: a passive stake is "brilliant PR" but creates exactly the wrong incentives; the government's job is to break monopolies, not co-own them.
• OpenAI's three stated goals. An automated AI researcher, an accelerated economy, and a personal AGI for everyone. Why the post has zero specifics — and why past commitments ("open," "nonprofit") make the new ones hard to trust.
• Anthropic ships Fable. Are model step changes still real, or is using it for everyday work "squashing an ant with a rocket launcher"? How the bar moved from "can it code?" to "can it automate multi-week R&D" — and when to reach for the big model versus 4.8.
• "When AI Builds Itself." Anthropic's recursive self-improvement paper, and its conveniently-timed "option to pause." Josh argues a coordinated pause is just collusion — anti-competitive regulatory capture dressed up as safety.
• The SpaceX IPO. Reserving 25–30% for retail versus the usual 5–10%, AI-plus-space as "a weird marketing company," and the one real bull case: Elon raising a war chest to go make chips. Plus: a reality distortion field is not a reason to auto-invest — ask Elizabeth Holmes and SBF.
• Why is everyone suddenly building loops? Boris Cherny kicked off the discourse and startups appeared overnight. Josh: loops are as old as Turing and mostly just burn tokens — the real work is the scaffolding and verification. The endgame, "while true, make it good," is also the point where the job disappears.
• Loss of control, demystified. It doesn't mean the AI suddenly gets smart. It means the changelog gets too long to read and engineers stop checking.

Subscribe to Outset Capital on YouTube for new episodes each week.