#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 restrictions. They cover XAI’s Grok Voice Think Fast 1.0, which claims significant leads in real-time voice benchmarks, and Anthropic’s new integrations with creative tools. The episode also details DeepSeek V4, highlighting its massive scale and novel hybrid attention architecture designed for efficient one-million-token contexts.
Topics discussed
Introduction and episode overview
Elon Musk's OpenAI lawsuit and opening statements
Anthropic's Claude 3.7 Sonnet and reasoning capabilities
Model alignment, deception, and sandbagging evals
GPT-4.5 release, benchmarks, and training data
xAI Grok 3 release and voice model benchmarks
DeepSeek V4 architecture and long context optimization
Tencent's Hi3 model and domestic AI infrastructure
Clawmark: A new benchmark for real-world agentic tasks
Google's $40B investment in Anthropic and AWS Graviton
Meta's blocked acquisition of Chinese AI startup Manna
OpenAI-Microsoft deal: AGI clause removal and revenue
Legal updates: Musk trial and Pentagon AI contract dispute
Google Gemini on-prem appliance for secure enterprise use
David Silver's new AI startup and reinforcement learning
AI sabotage risks and document editing degradation
Sparse autoencoders and feature attribution in LLMs
US government memorandum on AI security coordination
Teen boys dating AI chatbots and social impacts
CISA vulnerability disclosure and government AI security
Bit-flip attacks on neural networks and conclusion
Listen ad-free on Castria