---
title: 'In the News: September 17, 2026'
description: 'A Google/DeepMind paper reports measured gains from a self-improving coding-agent exploration layer, and a new roadmap defines what "genuine" RSI requires.'
canonical_url: 'https://darkfactory.dev/news/2026-09-17-evening'
markdown_url: 'https://darkfactory.dev/news/2026-09-17-evening.md'
collection: news
date_published: '2026-09-17T19:20:00-04:00'
date_modified: '2026-09-17T19:20:00-04:00'
---

# In the News: September 17, 2026


Dream-RSI reports efficiency gains from rewriting a coding agent's exploration policy while leaving the underlying model unchanged. A separate 35-author survey proposes five autonomy levels for recursive self-improvement and does not present a system that reaches the top level. A 50-comment Hacker News discussion with one Dream-RSI author centers on whether that work qualifies as RSI.

## 1. Dream-RSI rewrites a coding agent's exploration policy

**[Dream-RSI: Recursive Self-Improvement through Evolving Worlds](https://arxiv.org/abs/2609.14858)** · Tong Zheng, Xidong Wu, Zheng Zhang and 14 co-authors, Google, Google DeepMind, University of Maryland and University of Virginia · arXiv, posted September 14, 2026

In a large search space, a fixed exploration strategy stops adapting as the space grows, while tuning the strategy online consumes a full discovery run for every policy test. Dream-RSI treats a coding agent's history, the record of what it tried and what happened, as a replay simulator. A candidate exploration policy can be tested by re-walking that recorded tree in a different order instead of re-running the agent, which the authors call "dreaming." The method leaves the underlying coding agent unchanged and rewrites the orchestration layer that decides what to explore next. In algorithm engineering on Gemini 3.1 Pro, the paper's code repository reports 1.22x faster downstream runtime and 1.74x less discovery compute than Recursive Fixed Exploration, plus 162x fewer calls than SimpleTES. In GPU kernel engineering, 4 of 4 kernels improved at 2.09x higher performance for equal budget. In mathematical optimization, 2 of 3 tasks matched or beat the baseline. The [Hacker News discussion](https://news.ycombinator.com/item?id=49726955) (207 points, 50 comments a day after posting) split over whether "RSI" is the right word for it. Commenter bob1029 called the exploration-policy optimization "not recursively intelligent in any way," and mohamedmohey argued "if this is RSI then all RL is RSI." Co-author xidong_wu replied in the thread to confirm the mechanism: "This paper optimizes a controller/policy which will be used to agent itself in the next round." Full reproduction code and the discovered programs are not released yet; the repository lists them as still being prepared.

**Why it matters:** Dream-RSI offers a concrete harness pattern: record exploration decisions, replay them offline to evaluate policy changes, and deploy the best-performing policy. The method reduces the cost of iterating on how an agent chooses its next action.

## 2. A 35-author roadmap defines five levels of recursive self-improvement

**[The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement](https://arxiv.org/abs/2609.11873)** · Yi Duan, Fan Wu and 33 co-authors, corresponding author Xuanhe Zhou · arXiv, posted September 10, 2026, revised September 15

The survey builds a five-level autonomy framework, running from a system that executes human-specified updates to one that revises the mechanisms governing its own future improvement, and uses it to classify existing work. The Darwin Gödel Machine, which evolves coding agents and raised its score on a SWE-bench subset from 20 percent to 50 percent, still leaves its archive-maintenance and parent-selection rules outside the system's own control, so the paper credits it with better outputs but not with a self-revised improvement process. Gödel Agent, which rewrites both its task policy and its own improvement logic, ended 14 percent of its 100 recorded MGSM optimization trials below where it started. The authors use that result to argue that persistence alone is not evidence of progress. The paper also cites first-party operational figures: Anthropic reports agentic workloads use roughly four times the tokens of ordinary chat, rising to about fifteen times for multi-agent systems, and OpenAI reports its internal coding-inference compute grew 100-fold over six months of GPT-5.6 development. On Hacker News, in the same thread discussing Dream-RSI, several commenters made the same point this paper makes explicitly: repeated use of "RSI" and "AGI" without a shared definition lets very different claims travel under one banner.

**Why it matters:** The framework separates systems that produce better outputs from systems that revise the mechanism producing those outputs. Applied to Dream-RSI, that distinction separates a better exploration controller from a system that governs every part of its future improvement.