#255 - Gemini 3.7, Jalapeño, Qwen 3.8, Drones
Summary
Hosts Andre Kerenkov and Jeremy Gladstone discuss Google’s Gemini 3.7 Flash and xAI’s Grok 4.6, analyzing strategic shifts toward mid-tier efficiency. They cover Anthropic’s invisible text watermarks and cybersecurity expansions, alongside OpenAI’s safety updates and custom Jalapeno chip benchmarks. The episode also touches on Meta’s new desktop app and the broader implications for AI infrastructure and frontier model competition.
Topics discussed
Intro, sponsors, and listener comments on Qwen 3.8B
Google's strategy: Gemini updates, TPU focus, and leadership losses
xAI acquires Cursor and releases Grok 4.6 coding model
Synthetic text detection via watermarking and distribution shifts
Cloud security updates and $35M open source security credits
OpenAI's private safety processing and teen emotional attachment limits
Meta's Muse Code and ad campaign management tools
Custom AI chips: OpenAI, Anthropic, and the Nvidia margin war
OpenAI executive departures and IPO timing concerns
Anthropic's hardware push, revenue model, and economic impact
Thomson Reuters fine-tunes open source models for fiduciary AI
Qwen 3.8B performance benchmarks and open source trends
AI-guided drones in Ukraine and autonomous weaponization risks
OpenAI security pause and industry skepticism on slowing down
New lawsuit accuses Rock of generating CSAM
Research: Hyperparameter tuning for small-scale AI experiments
Security flaw: Extracting encrypted chain-of-thought from APIs
Massive activations in hybrid linear attention models
LinkedIn AI saturation and closing remarks
Listen ad-free on Castria