LWiAI Podcast #243 - GPT 5.5, DeepSeek V4, AI safety sabotage
Summary
Hosts Andrei Krenkov and Jeremy Harris review the launch of GPT 5.5, noting its coding strengths and the controversial "goblin" behavior in its system prompt. They discuss XAI’s Grok Voice Think Fast 1.0, which claims significant benchmark leads in real-time conversation, and DeepSeek’s open-source V4 model featuring a 1-million-token context window. The episode also touches on the Elon Musk vs. Sam Altman trial and Anthropic’s new creative tool integrations.
Topics discussed
Introduction and episode overview
GPT-5.5 release and OpenAI vs Anthropic competition
Model safety, deception, and alignment testing
GPT-5.5 benchmarks and the 'goblin' data poisoning incident
xAI Grok updates and Starlink integration
DeepSeek V4 architecture and Chinese AI progress
Tencent Hi3 and new reasoning benchmarks
Anthropic funding, Google partnership, and infrastructure
Geopolitics: Chinese AI founders and US-China tensions
OpenAI-Microsoft deal and Elon Musk lawsuit testimony
DOJ appeal in Pentagon AI dispute
Google Cloud confidential compute and TPU strategy
DeepMind's David Silver raises $1.1B for new lab
Study on AI agents sabotaging safety research
Document degradation in AI editing tasks
Sparse autoencoders and interpretability research
US policy on AI model distillation and security
Teen boys dating AI chatbots and user surveys
CISA vs NSA access to AI models and Matthew McConaughey likeness rights
Sign bit vulnerability in neural networks and outro
Listen ad-free on Castria