Developer Builds Python SDK to Measure True Cost of AI Self-Improvement

An independent developer has created an open-source Python SDK called dream-rsi-sdk, built around the recently published Dream-RSI concept, which improves AI agent behavior by replaying exploration history and rewriting policies offline without changing model weights. The project's central goal is to account for the full cost of self-improvement, including training and validation, rather than reporting only deployment savings. An early experiment using a cloud-based model showed the full process consumed 100 model requests compared to a 24-request baseline, exposing how deployment-only metrics can be misleading. A follow-up local experiment using Ternary-Bonsai-27B-Q2_g64 via llama.cpp yielded a more honest result: the learned policy path cost 248 total operations versus 256 for the fixed baseline, a modest 3.125% all-in reduction with quality unchanged. The developer highlights an amortization insight — the upfront 24-operation preparation cost breaks even after roughly 48 tasks, making the approach increasingly worthwhile at scale.