LWiAI Podcast #238 - GPT 5.4 mini, OpenAI Pivot, Mamba 3, Attention Residuals
Summary
Hosts Andrei Karpathy and Jeremy Harris discuss OpenAI’s new GPT-5.4 Mini and Nano models, analyzing their pricing and efficiency. They cover Mistral’s consolidated open-source Small 4 model and the emerging "operating system" race for AI agents, featuring Meta’s Manus and NVIDIA’s NemoClaw. The episode also touches on NVIDIA’s DLSS 5 graphics tech, OpenAI’s delayed adult mode, and the company’s strategic pivot toward enterprise productivity.
Topics discussed
Introduction and episode preview
OpenAI GPT-4.5 Nano: Pricing and performance
Mistral models and reasoning capabilities
AI agents on local computers and OpenClaw
NVIDIA's strategy: Hardware, CUDA, and agents
Generative AI in video game graphics
AI roleplay, erotic content, and safety concerns
OpenAI's strategic shift and product focus
NVIDIA Grok chips and inference hardware
Mistral vs Cohere and European AI markets
ByteDance access to NVIDIA chips in China
Meta's AI delays and internal reorganization
AI safety: Stenography and reasoning theater
Defenses against fine-tuning misalignment
Anthropic's AI safety evaluation framework
US export controls and AI chip policy
Model architecture: Residual streams and layers
Mamba 3: State space models and efficiency
Listen ad-free on Castria