#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, which claims significant leads in real-time voice benchmarks, and Anthropic’s new creative tool integrations. The episode also details DeepSeek V4, a massive open-source model featuring novel hybrid attention architectures for efficient million-token context windows.
Topics discussed
Intro, hosts, and sponsor ads
GPT-5.5 analysis, reasoning, and deception tests
xAI Grok 3 release and performance benchmarks
DeepSeek V4, Tencent HiFree, and open source models
New Clawmark benchmark for agentic AI
Anthropic funding, AWS Graviton, and Meta acquisition
OpenAI-Microsoft deal changes and Elon Musk trial
DOJ vs Anthropic and Google Gemini private deployments
New AI startup funding led by David Silver
AI safety research: sabotage and document editing errors
Temporal sparse autoencoders and interpretability
Policy updates, AI relationships, and neural lesion attacks
Listen ad-free on Castria