#243 - GPT 5.5, DeepSeek V4, AI safety sabotage
Summary
Hosts Andrei Krenkov and Jeremy Harris discuss the launch of GPT-5.5, analyzing its coding prowess, safety evaluations, and the bizarre "goblin" prompt restriction. They cover XAI’s Grok Voice Think Fast 1.0 and Anthropic’s new creative tool integrations. The episode also details DeepSeek V4’s open-source release, highlighting its 1-million-token context window and novel hybrid attention architecture.
Topics discussed
Intro, hosts, and sponsor messages
GPT-5.5 analysis, safety concerns, and the 'goblin' phenomenon
Google Gemini 3.1 Flash Live and real-time audio models
DeepSeek V3 architecture, compressed attention, and hardware
AI agent benchmarks and real-world task success rates
Anthropic funding, Meta-Amazon partnership, and AI geopolitics
OpenAI-Microsoft deal changes and AGI profit clause removal
Elon Musk vs. OpenAI lawsuit and DOJ antitrust cases
Google air-gapped servers and DeepMind's David Silver startup
AI sabotage research and model self-preservation behaviors
Document editing failures and interpretable AI research
US-China AI policy, AI companions, and CISA access
Neural lesion attacks, voice likeness rights, and outro
Listen ad-free on Castria