Stochastic的动态

With DeepSeek V3, we now have a fully open-source model that outperforms GPT-4 and Claude 3.5 Sonnet on multiple benchmarks. What is even more impressive is that this model was trained with only $5.5m on H800s! (note: DeepSeek V3 was trained on 2,048 H800s vs 16,000 H100s for LLaMA 3) 2025 will be wild. Here are my predictions for AI in 2025: - As scaling laws slow down and smaller open-source models become more performant, more AI solutions will be deployed privately / on-prem - How you leverage proprietary data will become more important than ever - Domain-specific reasoning enabled not by the largest models but by smaller models that are tuned on deep domain knowledge - Mainstream adoption of real-time voice interactions in enterprise settings - Many "GPT-wrapper" companies will fail to build upon their initial success targeting "low-end" applications (we saw this happening with GPT-3 wrapper companies) as they become commoditized - General AI assistants and agents will eventually be dominated by big tech as the models themselves become more performant - further driving out GPT-wrapper companies that are competing in the same space - More emphasis not on how to build around AI models but on how to build deeper integration into the models - More proactive AI agents, not just reactive ones - Real autonomous agents will emerge Is there anything that I missed?