Kimi 3: AI bench marks are just marketing | David Gerard
Summary
Isaac and guest David Gerard critique AI benchmarks as marketing tools rather than scientific measures, arguing that models like Kimi 3 are tuned to pass them. They discuss how the AI industry remains unprofitable and heavily subsidized, with vendors burning cash for hype. The episode explores the high cost of running local models and suggests that commoditization is unlikely until prices rise to sustainable levels.
Topics discussed
Sponsors: California Closets, Acast, and IG
Kimi 3 release and the unreliability of AI benchmarks
Model parameters, training data, and diminishing returns
Marketing hype vs. actual value and enterprise adoption
Cost of local models and the DeepSeek precedent
NVIDIA's dominance and China's chip restrictions
Microsoft's potential shift and the Western AI premium
Protectionism, market commoditization, and future outlook
Efficiency claims, copyright issues, and profit sustainability
Outro and final sponsor messages
Listen ad-free on Castria