19 hours ago
- Five hyperscalers control over 70% of global AI compute, with much reserved for OpenAI, Anthropic, and Google DeepMind, raising concerns about non-singularity AI use cases being sidelined and average people being priced out of AI benefits, with suggestions of universal basic compute redistribution.
- Key questions about AI advances include how Anthropic and others achieved long-horizon coding agents, whether models are becoming more sample efficient, and the tradeoff between memory and sample efficiency in context learning.
- The difference between training and inference workloads may blur, potentially leading to on-the-job learning for AI, while concerns arise about a 'Y2Key' event if future models train on AI-generated data and the economic impact of a machine-only economy.
- Continual learning could enable rapid superintelligence without further algorithmic progress as AIs amalgamate learnings across copies, and AI may excel in scientific discovery by exploring more theories through Bayesian approaches.