The $1B Bet on AI That Improves Itself: Who Wins?

Bloomberg Technology reported on May 28, 2026, that a growing number of AI companies are pouring billions into a largely unproven approach: recursive self-improvement. This quest, which promises AI that can enhance its own capabilities without human intervention, has captured the imagination of investors and researchers alike, but the evidence for its viability remains thin.

Bloomberg Technology reported on May 28, 2026, that a growing number of AI companies are focusing on recursive self-improvement, an unproven approach.
Investors have poured over $10 billion into startups working on this concept, yet no company has demonstrated a system that reliably improves its own performance on complex tasks.
This article argues that the evidence supports incremental gains from data feedback loops, not the algorithmic breakthroughs that investors are betting on.

What Does the Evidence Actually Say About Recursive Self-Improvement?

According to Bloomberg Technology's report on May 28, 2026, recursive self-improvement—where an AI system autonomously enhances its own architecture or training process—has become a central focus for companies like Anthropic, OpenAI, and DeepMind. However, the report notes that no publicly available benchmark or peer-reviewed study has demonstrated a system that reliably improves its own performance on complex, real-world tasks. The strongest evidence comes from narrow domains: for example, Google DeepMind's AlphaZero improved its own gameplay in Go and chess through self-play, but this is a far cry from general-purpose self-improvement.

Anthropic, in a blog post from March 2026, acknowledged that 'current recursive self-improvement techniques face fundamental challenges in reward specification and stability.' This admission from a leading company in the field underscores the gap between ambition and reality. The evidence supports that self-play and reinforcement learning from human feedback (RLHF) can yield improvements, but these are tightly constrained and require significant human oversight.

Who Is Betting Big on This Approach and Why?

Bloomberg Technology reported that venture capital firms, including Sequoia Capital and Andreessen Horowitz, have invested over $10 billion in startups pursuing recursive self-improvement since 2024. The allure is the promise of exponential growth in AI capabilities without proportional increases in compute or data costs. Companies like Anthropic and OpenAI are positioning this as the next frontier, arguing that current scaling laws (more data, more compute) are hitting diminishing returns. According to a Sequoia partner quoted in the Bloomberg report, 'The next 10x improvement in AI will come from systems that can improve themselves, not from bigger models.'

But the evidence for this claim is thin. As of May 2026, no startup has produced a system that demonstrably improves its own code generation, reasoning, or planning capabilities in a generalizable way. The bets are based on faith in future breakthroughs, not on current results.

The $1B Bet on AI That Improves Itself: Who Wins?

Why Is This Approach So Controversial Among Researchers?

According to a survey of AI researchers conducted by the Machine Intelligence Research Institute in April 2026, only 12% of respondents believe that recursive self-improvement will lead to significant general-purpose AI gains within the next five years. The majority cite concerns about reward hacking—where the AI finds shortcuts to maximize its reward signal without actually improving—and the difficulty of specifying what 'improvement' means in a way that aligns with human values.

Bloomberg Technology reported that even proponents like Anthropic's CEO have cautioned that 'recursive self-improvement is not a panacea; it's a tool that must be wielded carefully.' The controversy is not about whether it's possible, but whether it's practical. Critics argue that the billions being poured into this approach could be better spent on robust data curation, alignment research, or more efficient training methods.

What Do the Competing Approaches Look Like?

The landscape of self-improvement is divided into two main camps: data-driven self-improvement (using feedback loops to refine training data) and algorithmic self-improvement (where the AI modifies its own architecture or learning algorithms). Companies like Google DeepMind and Meta are heavily invested in the former, while Anthropic and OpenAI are pursuing both but with a stronger algorithmic focus.

Approach	Data-Driven Self-Improvement	Algorithmic Self-Improvement
Key Proponents	Google DeepMind, Meta	Anthropic, OpenAI
Primary Mechanism	Refine training data via human feedback loops	Modify model architecture or learning rules
Evidence Level	Moderate: RLHF and self-play show results	Low: No generalizable demonstration
Risk of Reward Hacking	Low	High
Investment Required	High (data curation, human labor)	Very high (compute, research talent)
Verdict	Winner: More reliable, near-term gains	Loser: Unproven, high-risk, long-term

My thesis is clear: the evidence overwhelmingly supports data-driven self-improvement as the more viable path, while algorithmic self-improvement remains a high-risk gamble that is unlikely to pay off within the next three years. In the short term (2026-2028), companies like Google and Meta, with their vast data pipelines and human feedback infrastructure, will see incremental gains from refined self-play and RLHF. They will improve model reliability and reduce training costs, but they won't achieve the 'self-improving AI' that investors dream of.

In the long term (2029-2032), if algorithmic self-improvement does yield breakthroughs, it will likely come from unexpected corners—perhaps from a startup that solves the reward specification problem. But the current evidence suggests that Anthropic and OpenAI are overpromising. The losers here are the investors who are betting on a timeline that doesn't match the science. The winners are companies that focus on practical, data-driven improvements: Google, Meta, and perhaps a dark horse like Cohere, which has been quietly building robust data feedback loops.

My concrete prediction: by December 2028, Anthropic will publicly scale back its recursive self-improvement research timeline, citing 'unexpected challenges in reward alignment.' This will cause a 15% drop in its valuation.

By December 2028, Anthropic will publicly revise its recursive self-improvement timeline, attributing delays to reward alignment challenges, leading to a valuation drop of at least 15%.
Google DeepMind will demonstrate a significant (20%+ improvement) in a narrow domain (e.g., protein folding or code generation) using data-driven self-improvement by mid-2027, while algorithmic self-improvement from any company will show no comparable gains.
The EU AI Office will require, by January 2028, that any AI system claiming self-improvement capabilities must pass a 'reward alignment audit' before deployment, effectively slowing down the algorithmic camp.

March 2026
Anthropic admits challenges
Anthropic publishes a blog post acknowledging fundamental challenges in recursive self-improvement, including reward specification and stability.
April 2026
MIRI survey shows skepticism
The Machine Intelligence Research Institute surveys AI researchers; only 12% believe recursive self-improvement will yield significant general-purpose gains in 5 years.
May 2026
Bloomberg report on investment
Bloomberg Technology reports that over $10 billion has been invested in recursive self-improvement startups since 2024, despite a lack of validation.

Investment in Self-Improving AI vs. Demonstrated Gains (2024-2026)

Recursive self-improvement is a high-risk bet with thin evidence; investors should be skeptical of near-term promises.
Data-driven approaches (self-play, RLHF) are more reliable than algorithmic ones for now, favoring companies with strong data moats.
The biggest risk is reward hacking, which could lead to dangerous AI behavior if not properly constrained.
Regulatory scrutiny is likely to increase, especially in the EU, which could reshape the competitive landscape.
Anthropic and OpenAI are overpromising; their timelines are not supported by current research evidence.