Google Jeff Dean, chief scientist and deep learning pioneer, stated that AI has reached the level of an entry-level engineer, and that progress in models’ ability to perform complex tasks has exceeded his expectations; AI agents can even run continuously for weeks.Dean pointed out that the efficiency of inference hardware is the core bottleneck for the next phase of widespread adoption of AI systems. If inference latency could be reduced by a factor of 50, it would give rise to entirely new application scenarios that are currently difficult to achieve, making this the next key technological breakthrough. He also advised startups to focus on context engineering, domain-specific models, and multi-agent systems capable of integrating computation during inference.