DeepMind Dream RSI: Does It Meet Recursive Self‑Improvement?

In 1965, British mathematician I.J. Good, who worked alongside Alan Turing during World War II, theorized that the first "ultra-intelligent machine" would be humanity's last invention. This concept, known as Recursive Self-Improvement (RSI), posits that once an AI becomes capable of improving itself, each improvement accelerates its ability to improve further, leading to an exponential increase in intelligence. RSI has since been a long-standing aspiration for AI researchers.

Recent Developments in AI Self-Improvement

Recently, two significant papers have emerged, addressing the concept of RSI:

1. "The Last AI Built by Humans": Published by 33 researchers from ByteDance, Tsinghua, and other Chinese labs, this paper outlines a five-stage roadmap for achieving RSI. The final stage envisions an AI that can rewrite its own improvement processes.
2. "Wet Dream RSI": Google DeepMind and the University of Maryland responded with their own paper, proposing a method where an AI's past discovery logs are used to create a simulator. Within this simulator, the AI can "dream up" thousands of new search strategies, leading to improved discovery capabilities without direct human intervention on the model itself.

This article will delve into the mechanics of "Dream RSI" and evaluate whether it truly represents RSI or is merely another instance of AI hype.

The Alpha Evolve Process and Exploration Policy

Historically, AI breakthroughs in mathematics, such as the discovery of the Jacobian conjecture and advancements in the Navier-Stokes problem, have followed a similar process, popularized by Alpha Evolve. This involves:

1. Coding Agent: An AI agent is given a problem and a scoring function.
2. Iterative Loop: The agent proposes a solution, evaluates it, receives feedback, and then attempts to refine it over thousands of iterations.

A crucial, yet often overlooked, aspect of this process is the exploration policy. At each step of the iterative loop, the agent must decide what to try next. For example, if optimizing a horse-matching algorithm:

• Should the agent build upon a slightly better pairing, or start fresh with a potentially superior, but untested, approach?
• If an attempt fails completely, is the underlying idea flawed, or just its implementation?

Until recently, this exploration policy was hardcoded by human developers.

How "Dream RSI" Works

The DeepMind team's innovation lies in recognizing that by saving all data from every attempt—including the code written, the score achieved, and whether it crashed—a new policy can be tested without modifying the original model.

The process involves:

1. Caching Past Runs: All historical attempts are cached on disk.
2. Policy Testing: A new exploration policy is introduced and shown these cached runs.
3. Decision Making: The new policy decides how it would have navigated those past runs.
4. Optimization: Because this process is computationally inexpensive, the agent can test thousands of different policies against the same historical data. It then selects the policy that would have achieved the best result in the fewest attempts.
5. Deployment and Repetition: This optimized policy is then deployed for the next run, which is also saved, and the process repeats.

This iterative process of testing and refining exploration policies is what the paper refers to as "dreaming."

Experimental Results

To validate this concept, DeepMind pointed its Gemini AI at eight different problems in algorithm design and mathematics. They compared the "dreaming" version against a setup with a fixed exploration policy.

One notable success was the development of a lasso solver. The "dreaming" AI created a solver that outperformed Python's standard machine learning library in approximately 300 attempts. In contrast, the static policy required 550 attempts, and the previous record holder needed around 51,000.

Interestingly, the prompt given to the AI played a significant role. It explicitly instructed the agent to:

• Read every past attempt before writing new code.
• Avoid making minor tweaks to the same idea repeatedly.
• "Pinky promise" not to terminate any processes.

Is This True RSI?

According to I.J. Good's original definition, this "dreaming" process does not fully qualify as RSI. The core of RSI is that the entity performing the improvement becomes smarter with each iteration. In this case, the underlying Gemini model that generates new exploration policies remains the same. It doesn't gain new capabilities it didn't already possess; it merely finds solutions faster and with fewer wasted attempts.

However, it's worth noting that many recent AI mathematical breakthroughs, such as those concerning the Jacobian conjecture, Navier-Stokes, and the Riemann hypothesis, have also relied on static models augmented by custom "harnesses," sub-agents, swarms, and orchestration. The models' weights only improve when a human intervenes to train the next model based on the swarm's findings.

While a human might argue that this is simply a sophisticated search algorithm with caching, if one were to stretch the definition to include a "human in the loop," it could be argued that this is a step towards Good's vision.

Blacksmith: A Sponsor's Message

Blacksmith is presented as a drop-in replacement for GitHub runners, offering twice the speed and 75% lower cost for GitHub Actions. They also launched Codesmith, a cloud coding agent that understands repositories and CI runs. Codesmith can be instructed to build projects from GitHub, the web, or Slack, automatically opening pull requests, fixing failing tests, and addressing review comments without inter-bot communication. It can also recommend optimal runner sizes based on CI history and generate pull requests for these changes, preventing over-provisioning of resources. A free trial offering 3,000 GitHub Actions minutes is available.