DeepSeek Scales RL For Million-Token Context And Agentic AI
——0——
There are two parts of DeepSeek V4 project. - How do we make inference very cheap, even at extreme sequence lengths *and* hundreds of turns? - how do we make it infinitely parallelizable? If you don't see how this ends up in very strong agents, you've got some reading to do.
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex> but being smart across a long task by reading a codebase, planning changes, keeping track of what you're doing across thousands of steps that's a different problem. nobody solved it by making things cheaper. However, that's the plan and that's what DeepSeek is doing right now.
8:27 PM · May 27, 2026 · 3.5K Views
8:34 PM · May 27, 2026 · 7.4K Views