#243 - GPT 5.5, DeepSeek V4, AI safety sabotage
Summary
Hosts Andrei Krenkov and Jeremy Harris discuss the launch of GPT-5.5, its performance in coding, and safety concerns regarding chain-of-thought faithfulness. They cover XAI’s Grok Voice Think Fast 1.0, which claims significant leads in real-time conversation benchmarks, and DeepSeek’s open-source DeepSeek V4, featuring a massive 1.6-trillion parameter model with a 1-million-token context window. The episode also touches on the Elon Musk vs. Sam Altman trial and Anthropic’s new integrations with creative tools like Photoshop.
Topics discussed
Intro, hosts, and sponsor segments
GPT-5.5 release, benchmarks, and safety evals
xAI Grok updates and voice model performance
DeepSeek V4 architecture and Tencent's Pi3
New benchmark for complex reasoning tasks
Google and Anthropic compute partnership
Meta's AWS Graviton chip deal for AI
US restrictions on Chinese AI acquisitions
OpenAI-Microsoft deal and Elon Musk trial
DOJ antitrust case against OpenAI
Google Gemini private appliance for enterprise
David Silver's new AI startup raises $5.1B
AI sabotage risks and document editing errors
Temporal sparse autoencoders for interpretability
Policy memos and AI chatbot dating concerns
AI economic impact and government security access
Deep neural lesions and model fragility research
Listen ad-free on Castria