DeepSeek Researcher Reveals Path to Test-Time Parametric Continual Learning

DeepSeek researcher Shengding Hu outlined a concrete technical direction: pursuing test-time parametric continual learning rather than RSI or harness-level evolution, with the post-V3 goal of 'infinite context' via KV compression to 890 bytes per token and near-linear cost at 1M context.

2026-09-13 ~ 2026-09-13 · 2 related posts

• DeepSeek researcher outlines direct path to test-time parametric continual learning — teortaxesTex · 2026-09-13
• DeepSeek targets 'infinite context' and test-time parametric continual learning — teortaxesTex · 2026-09-13