#255 - Gemini 3.7, Jalapeño, Qwen 3.8, Drones
Summary
Hosts Andrei Kurenkov and Jeremy discuss recent AI developments, including Google’s Gemini 3.7 Flash and SpaceX AI’s Grok 4.6, analyzing their strategic implications for frontier modeling. They cover Anthropic’s invisible watermarking and cybersecurity expansions, alongside OpenAI’s new safety features and custom Jalapeno chip benchmarks. The episode also touches on Meta’s desktop app and the broader industry shift toward vertical integration.
Topics discussed
Intro, sponsors, and episode preview
Google Gemini updates and hardware strategy
SpaceX AI Grok 4.6 and Cursor acquisition
AI text watermarking and detection
Mythos 5 security tool and human vulnerabilities
OpenAI safety features and teen mode
Meta's agent app and business model risks
OpenAI's Jalapeno custom chip and energy needs
OpenAI executive departures and reorg
Anthropic's growth and compute economics
Thomson Reuters launches in-house AI model
Small dense models and distillation trends
Autonomous drones and AI in modern warfare
OpenAI pauses RL fine-tuning for safety
Scaling laws for small models
Chain of thought extraction and quantization
AI slop and web traffic statistics
Listen ad-free on Castria