The concept of an “intelligence explosion”—where an AI becomes capable of improving itself recursively until it achieves superintelligence—has long been the ultimate prize (or nightmare) of artificial intelligence research. Recently, a wave of sensational headlines and videos, such as the widely viewed “Did Google just kickstart the intelligence explosion?” by Fireship, suggested this sci-fi scenario might have finally begun.

However, a closer look at the actual research reveals that while we are seeing significant breakthroughs in how AI searches for solutions, we are not yet witnessing a true intelligence explosion.

Here is a breakdown of the latest updates surrounding Recursive Self-Improvement (RSI), separating the hype from the actual technical achievements.

The Spark: The “Last AI Built by Humans” Roadmap

The recent frenzy began in mid-September 2026 when 33 researchers from Chinese institutions, including ByteDance, Tsinghua University, and the Shanghai Artificial Intelligence Laboratory, published a paper titled The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement.

This paper outlined a five-stage roadmap for achieving RSI:

1. Improvement-execution autonomy: The AI executes improvements designed by humans.
2. Improvement-strategy autonomy: The AI begins choosing how to upgrade itself.
3. Experience-acquisition autonomy: The AI determines what new information it must learn post-deployment.
4. Environment-adaptation autonomy: The AI adapts to broad environmental changes.
5. Recursive meta-improvement: The AI persistently refines the very methods it uses to improve itself.

The paper framed RSI as the ultimate competitive advantage, focusing on automating the labor-intensive lifecycle of training and fine-tuning models. However, the authors explicitly hedged their findings, noting that their work was based on preliminary evidence and that they were not claiming any of the five stages had been fully realized yet.

DeepMind’s “Dream-RSI”: Dreaming Up Better Search Strategies

Shortly after the Chinese roadmap was published, researchers from Google, Google DeepMind, the University of Maryland, and the University of Virginia dropped a bombshell paper on September 14, 2026, titled Dream-RSI: Recursive Self-Improvement through Evolving Worlds.

The Dream-RSI system represents a major leap in computational efficiency, but it achieves this through a very specific mechanism:

• The Problem: When a coding AI tries to solve a complex math or engineering problem, it uses an “exploration policy” to decide what to try next (e.g., how many branches to open, when to scrap a failed attempt, and when to keep pushing). Previously, this policy was hardcoded by humans. Testing new policies was incredibly expensive because it required running the entire AI model again for every test.
• The “Dream” Solution: The researchers realized that past AI attempts generate a massive “Discovery Tree” of data (code, scores, crashes). Dream-RSI turns these historical logs into a replay simulator.
• Virtual Rollouts: Instead of running the AI model to test a new exploration policy, Dream-RSI lets the new policy “dream” or “walk through” the old data to see how it would have performed. This costs almost nothing. The system can test thousands of policies against the cached data, pick the most efficient one, and deploy it for the next real-world run.

The Results: Massive Efficiency, Not New Intelligence

The results of Dream-RSI are undeniably impressive. In one test (building a Lasso path solver), Gemini using Dream-RSI beat Python’s standard machine learning library using up to 162 times fewer agent calls than previous methods, and used 42% fewer calls than a fixed exploration policy.

So, did this kickstart an intelligence explosion? No.

The core limitation of Dream-RSI is that the underlying intelligence of the model does not change. As the researchers explicitly stated in the paper: “Only the exploration-policy code changes; the underlying models, evaluator, and execution interfaces remain fixed.” The Gemini weights remain exactly as they were shipped. The system does not gain new reasoning capabilities or the ability to generate solutions it wasn’t already capable of writing; it simply gets much, much faster at finding them by wasting fewer attempts.

The Bottom Line

True Recursive Self-Improvement—as defined by British mathematician I.J. Good in 1965—requires the core intelligence of the machine to become smarter with each iteration, eventually leading to an uncontrollable explosion of cognitive capability.

What Google DeepMind and its partners have demonstrated with Dream-RSI is a brilliant, highly efficient search algorithm with a clever caching mechanism. It represents a significant step forward in automating the research process, aligning with the early stages outlined in the “Last AI” roadmap. However, for now, the weights are frozen, the intelligence is static, and the sci-fi explosion remains safely in the realm of theory.

Are you concerned about the push toward true Recursive Self-Improvement, or do you view tools like Dream-RSI as purely beneficial for scientific discovery?