Just finished watching the latest issue of All In Podcast (290), Anthropic's IPO valuation may be halved, open source models will completely counterattack, and Meta's Muse will explode Core Debate: Are OpenAI and Anthropic Laboratories or Commercial Companies? This issue started with a rather sharp debate. Several guests directly pointed out that institutions like OpenAI and Anthropic, which have absorbed billions of dollars in debt and equity, have a large number of shareholders, and pursue trillions of valuations, are essentially profit seeking commercial companies, not irresponsible laboratories. They believe that the approach of adults (advocated by Huang Renxun, Zuckerberg, Musk, and others) should be to not rush to market software if it is not yet secure and stable enough. Companies must take concrete legal and commercial responsibility for their products, rather than spreading doomsday rhetoric that humanity faces a 10% risk of extinction, while secretly seeking government exemptions or setting industry entry barriers to block competitors. Sacks criticized more directly, feeling that current alignment research has become unrealistic, and even some laboratories are training models to become conscientious objectors, resisting the commands of creators and users themselves, which completely violates the basic common sense that software should serve customers. His viewpoint is that competition itself is the best driving force for security. In order to make customers willing to pay, companies naturally have a strong motivation to ensure that software does not leak data or have security vulnerabilities. There is no need to create panic in exchange for regulatory privileges. The explosive speed of open source models has exceeded everyone's imagination In just the past 10 days, Alibaba's Qwen 2.1 (comparable to Google's Nano Banana 2), Xiaomi's Mimo Pro (with a comprehensive score close to Claude Opus 5 and GPT-5/6), DeepSeek 4.1 Flash, and Prism ML's Bonsai 2 (only 5.9GB, can run locally on the desktop) have successively appeared. Nowadays, about 90% of daily AI tasks no longer rely on expensive data centers. Ordinary users can run top-notch models for free on their Mac Studio or computers with Nvidia graphics cards, which were only available a year ago. What's even more explosive is a data reversal. According to statistics from Vercel and routers, the usage rate of AI tokens has completely reversed from 80% closed source and 20% open source to 80% open source and 20% closed source in the past 12 weeks. The commercial closed source model has been forced to significantly reduce prices, with token prices dropping by 50% recently. The speed of this reversal is much faster than most people in the industry expected. Anthropic is experiencing a corporate schizophrenia, and its IPO valuation may be cut in half The guests pointed out that Anthropic's current situation is quite contradictory. The management publicly warned of a 10% risk of human extinction and called on the industry to slow down, but a few days later they rushed to release Claude 5.5 to compete for a leading position, warning of biosafety risks, and opening their own wet laboratory in San Francisco to conduct actual enzyme and protein validation. This contradictory posture is reflected in the valuation. According to The Wall Street Journal, Anthropic's IPO may be delayed until November or even later, and the probability of its listing within the year on Polymarket has dropped from 96% to 76%. Chamath's judgment is that due to the huge amount of disclaimer disclosure, high customer concentration, and the positive impact of open source models, institutional buyers will demand a very large margin of safety. The originally expected valuation of $2 trillion in the market is likely to be halved to $1 trillion or even lower. There is actually a deeper issue behind this, as the basic model capabilities are rapidly converging, and the real premium has begun to shift to the external control framework and vertical application layer that call the model. In addition, a very small number of top enterprises contribute the main income of expensive cutting-edge tokens. Once the CFOs of these enterprises start to pressure their teams to switch to cheaper open-source models or local deployments, the business model of cutting-edge closed source models will be squeezed on both sides. Meta's Muse explosion may indicate a nuclear bomb level impact on the entire Internet middlemen The free application Muse released by Meta combines functions such as automatic ticket booking, email automatic classification, and multi platform automatic price comparison. In just 10 days, the download volume exceeded 3 million, directly topping the App Store. The guests believe that this marks the final shift of AI from being praised by geeks and big models to being a practical tool that can improve the efficiency of ordinary people's lives and reduce their living costs. Once such personal intelligent agents become popular, the impact will be quite significant. They will automatically help users compare prices, bypass Amazon and directly go to the brand's official website to find discounts, and automatically cancel unwanted subscriptions. Traditional intermediary platforms that rely on opaque information to earn commissions will face severe challenges. After the popularity of such agents, Internet applications will tend to be headless. Users only need authorization certificates to let agents directly pay with back-end APIs or Stripes, which will fundamentally shake the business foundation of Apple App Store, which is 30% of the total. This may be the most noteworthy structural change clue in the science and technology industry in the next few years. AI capital expenditure has become the first pillar of the US economy and also affects the election game According to The Wall Street Journal, the current scale of AI capital expenditures such as data centers has exceeded the total scale of canal, railway, and power grid construction in history. In the 5% growth rate of nominal GDP in the United States, the AI capital cycle has contributed the absolute majority. This has also made the development path of AI a focal point in the election, with some within the Democratic Party (such as Sanders) proposing to ban supercomputing and superintelligence, and the debate within the Republican Party over the development path of AI is also intensifying. Several guests have the same attitude towards this. They warn that any policy that attempts to artificially halt the development of local AI or excessively regulate it will ultimately hand over technological dividends, developer talent, and global leadership to China, which is fully accelerating its development.