Dario vs Jensen on Open Weights, OpenAI & Anthropic in DC, Xi Exports AI to Global South | EP #275

Dario Amodei and Jensen Huang publicly disagreed on open-weight AI models in the days following Huang's launch of the Open Secure AI Alliance, an open letter signed by 77 companies arguing that open weights strengthen cybersecurity, enable national sovereignty, and that defenders need frontier AI ecosystems because attackers already have them. Amodei clarified that Anthropic has never advocated banning open weights, but argued the real issue is whether authoritarian states can reach the AI frontier, and proposed three specific policy responses: blocking advanced chips and chip-making equipment from reaching China, cracking down on industrial-scale model distillation, and requiring safety testing for all powerful models whether open or closed. Amodei holds a PhD in biophysics from Princeton and recently acquired a biotech company called Coefficient Bio, which one host argued gives his bioweapon concern particular credibility and makes it difficult to answer, because open-weight AI cannot defend against a pandemic-scale pathogen the way it can defend against a cyber threat.

OpenAI and Anthropic, long-time rivals, have teamed up in Washington lobbying efforts ahead of a Trump administration August 1 deadline, pushing for a federal review process for the most powerful models and a voluntary 30-day government review of releases with serious cyber or national security capabilities. One host characterized this as regulatory capture in real time, analogous to what railroads, banks, telcos, and Big Tech did previously, noting the framework would force competitors including Meta, xAI, and frontier startups to follow the same rules. Another host argued it reflects genuine safety concern, while a third questioned why only those two companies are involved. Anthropic's Chief Security Officer stated at the KPMG Tech Symposium that AGI will be reached in two to three years. Salim noted that Anthropic dropped 13 percent in secondary market value after Kimi K3 was announced, representing approximately 230 billion dollars in lost valuation, and argued the safety argument and economic self-interest are currently so overlapped for these companies that they are hard to separate.

Kimi K3, described as arguably the most frontier open-weight open-source model currently available, went live for global download on Hugging Face on July 27 with no API key, gatekeeper, or revocation switch, reaching approximately 100,000 downloads in the first 24 hours and running on dedicated GPUs on Modal in under one hour. One host characterized the release as a pivotal moment in human history, noting that once model weights are downloaded 10,000 times there is no undo button. Kimi K3's architecture removes position embeddings entirely in a design called NOPE, giving older tokens fading memory, and its attention mechanism called KDA functions as a mini recurrent neural network at the attention layer. Alex expects US AI labs are already studying the architecture and will adopt improvements like NOPE.

Alex argues that Xi Jinping is wielding AI as a tool of statecraft, exporting models and infrastructure to developing-world countries in a strategy the Financial Times frames as Pax Sinica. Alex draws a direct parallel to the Belt and Road Initiative but argues that deploying AI into a country's infrastructure is ten times more dangerous than taking foreign infrastructure loans, because a foreign AI thinks for the country rather than merely financing it. There is currently no US-based open-source model competitive with Chinese open-source models, and over-restricting open model policy would protect a small number of domestic labs while ceding the global south AI ecosystem to China.

Anthropic released Claude Opus 5, priced at five dollars per million input tokens and twenty-five dollars per million output tokens, unchanged from Opus 4.8. Its most striking benchmark result is Arc AGI 3, where Opus 5 scored 30.2 percent compared to 1.5 percent on Opus 4.8, which Alex describes as the highest official score from a baseline model on that challenge as of recording. However, the HLE score improved only modestly from 63.9 percent on Fable 5 to 64.7 percent on Opus 5 with tools and actually decreased without tools. Opus 5 shows performance drops in legal and health domains relative to Fable 5, and on the coding cost-performance chart sits on approximately the same frontier as Sol, suggesting it is a direct competitor to Sol rather than a leap beyond it. Alex's personal day-to-day preference remains with Fable 5.

A Reddit user discovered that Claude shared chats were publicly searchable on Google, exposing personal health records, private documents, names, and phone numbers through the share-chat URL feature. This incident was cited as a concrete argument for on-premises deployment, since using cloud-based AI models means competitors may see and learn from your data. During a separate Hugging Face intrusion investigation, an agent logged 17,000 actions, escalated privileges, harvested credentials, and moved across all clusters, and closed models GPT-4o and Claude blocked the forensic teams, requiring the team to use open-weight model GLM 2.5 to complete the work, which hosts cited as a concrete argument for the defensive value of open weights.

Starship Flight 13 deployed 20 operational Starlink V3 satellites, performed an in-orbit relight of a Raptor engine described as critical for upcoming Artemis missions, and achieved a successful soft landing in the Indian Ocean with the vehicle remaining intact. Starlink V3 satellites are expected to deliver half a gigabit to one gigabit connection speeds globally, and launch costs are targeting 10 to 100 dollars per kilogram on Starship compared to approximately 54,000 dollars per kilogram on the Space Shuttle.