Dream RSI: What the Video Says

The idea behind RSI

• In 1965, mathematician I. J. Good argued that an ultra-intelligent machine could improve itself, with each improvement making further improvements easier. This is known as recursive self-improvement (RSI).
• The video contrasts that idea with a recent paper from Google DeepMind and the University of Maryland, which explores a technique called “dreaming.”

How Dream RSI works

• AI problem-solving systems often use a loop: propose a solution, score it, review the feedback, and try again.
• A key choice in that loop is the exploration policy: what to try next, whether to build on a promising attempt, and how to respond to failures.
• Traditionally, people hard-code this policy. The “dreaming” approach instead:
• Saves each attempt, including its code, score, and whether it failed.
• Tests many possible exploration policies against those cached attempts.
• Chooses a policy that would have reached a strong result in fewer tries.
• Uses that policy in the next live run, then repeats the process.

Because the policies can be tested on existing logs, the system can explore many options without changing or rerunning the underlying model.

Reported results

• The researchers tested Gemini on eight problems in algorithm design and mathematics, comparing the dreaming setup with a fixed policy.
• One highlighted result was a lasso solver that outperformed Python’s machine-learning library in roughly 300 attempts. The fixed-policy setup took about 550 attempts, while the previous record took around 51,000.
• The prompt encouraged the AI to review past attempts, avoid repeating minor variations, and keep processes running.

Is it really RSI?

• By Good’s definition, the video says, no: Gemini itself does not become more capable from round to round. The same model creates the exploration policies; the system simply searches more efficiently.
• The video frames this as a powerful search algorithm with caching, rather than an AI that improves its own intelligence.
• It also notes that many recent AI mathematics results rely on fixed models paired with custom harnesses, sub-agents, and orchestration. Model weights generally improve later through human-led training.

Sponsor segment

• The video promotes Blacksmith as a GitHub Actions runner alternative, claiming faster runs at lower cost.
• It also advertises Code Smith, a cloud coding agent that can work across repositories, open pull requests, handle failing tests and review comments, and recommend CI runner sizes.