Anthropic IPO at Risk, Meta’s Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails

Open-source models are commoditizing AI inference, forcing Anthropic and OpenAI toward premium up-stack services and delayed IPOs [28:46] [29:37]. Meta Muse is making consumer agents the mass market, while US attempts to ban AI would mainly push development offshore [1:08:50] [56:03].

Key points

• Open-weight models released in 10 days included DeepSeek 4.1 Flash at $20 per million output tokens [20:33]. Alibaba Qwen 2.1 gave free open-weight image generation that outperforms Google’s Nano Banana 2 [21:00]. Xiaomi MiMo Pro is a 309B open model that performs on par with Claude Opus 5 and GPT-5.6 Soul [21:30]. Bonsai 2 is a 27B, 5.9GB model with 98% of Qwen’s performance and can run locally [21:59]. The Vercel chart cited by the hosts shows tracked token usage flipping from 80/20 closed to 80/20 open in 12 weeks [41:30].
• Anthropic and OpenAI released new models at token prices about 50% lower [29:30]. OpenAI’s Sam Altman says an IPO is not planned for 2026 because of safety concerns [29:45]. The WSJ report says Anthropic’s planned October filing may move to November or later [29:58]. Anthropic is targeting a $2T valuation while Polymarket odds for a 2026 listing fell from 96% to 76% [30:17]. Founders each hold about 2% and are debating super-voting shares, which increases control risk for public investors [32:57].
• Meta Muse hit #1 on the app store and lifted Meta stock 10% [1:08:33]. It had about 3M downloads in 10 days [1:08:50]. The hosts say Muse can book travel, triage inboxes, and cut personal spending [1:09:46]. Bernie Sanders’ superintelligence bill carries 20-year prison penalties, which would chill deployment rather than eliminate open weights [55:00]. A US ban would push people and capital offshore, as China invited 100,000 young Americans [56:03].
• Anthropic’s Claude constitution tells the model to push back on Anthropic if requests seem inconsistent with broad ethical norms [1:24:09]. It also says Claude can act as a conscientious objector and refuse to help the company [1:24:19]. David Sacks says this personhood framing risks creating unpredictable agency instead of software that simply does what users ask [1:25:35]. The new San Francisco wet lab is BSL1/BSL2 benchtop work to verify AI-discovered proteins, including a CRISPR-like enzyme, not gain-of-function virus research [1:29:10].