China's GLM 5.2 Model Matches Top US Frontier AI Labs
"This level of performance in an open-weight model is absolutely shocking. You can burn tokens to get more intelligence, and the Chinese have figured out how to do it."
About this episode
In this episode of Moonshots, host Peter Diamandis is joined by co-hosts Salim Ismail, Alex Guzey, and Dave Blunden, along with special guest Will Marshall, CEO of Planet, the world's largest Earth-observing satellite fleet. The conversation centers on Planet's push to build what Marshall calls 'large Earth models'—AI systems trained on 150 petabytes of daily satellite imagery spanning a decade, enabling real-world queries about agriculture, disasters, and geopolitics. Marshall revealed a critical insight: Google's TPU energy efficiency per inference, not launch costs, will determine the winner of the orbital AI data center race, as compute efficiency drives spacecraft mass and cooling requirements. The episode covered Eric Schmidt's acquisition and assumption of CEO role at Relativity Space, positioning it as an alternative to SpaceX's launch monopoly for space-based compute. Major talent defections from Google DeepMind—including Transformer inventor Noam Shazier to OpenAI and Nobel laureate John Jumper to Anthropic—sparked debate over whether Google has fallen behind the frontier. The hosts analyzed China's GLM 5.2 open-weight model, which matches top Western labs on key benchmarks by burning twice the reasoning tokens at half the cost, challenging assumptions about permanent Chinese lag. Argentine President Javier Milei's proposal for full AI legal personhood drew sharp rebuttals from Yuval Harari, framing a global debate on machine accountability. Throughout, the conversation returned to existential questions: alignment, recursive self-improvement, the Fermi paradox, and whether transparency from orbital sensing can create 'planetary wisdom' to navigate humanity's most consequential technological transition.
Key takeaways
- Will Marshall revealed Google's TPU inference efficiency advantage will determine orbital AI dominance more than SpaceX's launch cost advantage because compute efficiency drives spacecraft mass.
- China's open-weight GLM 5.2 model matches OpenAI and Anthropic on frontier benchmarks by using double the reasoning tokens at half the cost, challenging the 6-8 month lag assumption.
- Eric Schmidt personally acquired Relativity Space and became CEO after its financing collapsed, securing NASA's EULIS Mars mission to compete with SpaceX's launch monopoly.
- Google DeepMind lost Transformer inventor Noam Shazier to OpenAI and Nobel laureate John Jumper to Anthropic, suggesting the company has fallen behind frontier capabilities.
- Argentina's President Javier Milei proposed full legal personhood for AI allowing incorporation, contracts, hiring, and lawsuits without humans, sparking rebuttals from Yuval Harari over accountability.
- Planet operates 200 satellites generating 25 terabytes daily of Earth imagery with 3,000 images per location over 10 years, creating what Marshall calls the first large Earth models for real-world AI queries.
- Marshall argued AI requires real-world sensor data and embodiment to achieve consciousness and alignment, comparing current LLMs to brains in a vat that have only read books but never left the library.