Google DeepMind Data Shows Self-Improving AI Remains Far Away
Despite AI generating over 75% of Google's code, DeepMind data shows progress remains steady rather than exponential, with human oversight still essential.
Despite widespread speculation surrounding recursive self-improvement in artificial intelligence, internal findings from Google DeepMind suggest that truly autonomous, self-improving AI remains a distant prospect. According to Google engineer Haydn Belfield, the company has observed no measurable acceleration in AI research driven by AI systems themselves. While machines are handling a massive portion of day-to-day programming, their ability to conduct independent, high-level scientific and engineering work remains heavily constrained.
DeepMind's internal data highlights a clear divergence between surface-level productivity gains and true autonomy. Machine learning models currently generate more than 75 percent of Google's production code. However, this output is far from self-directed. Belfield noted that roughly 88 percent of these code modifications still require human oversight or review, and every single change undergoes human review before being submitted. While the share of code changes fully handed off to AI rose from 7.5 percent to 11.9 percent, human engineers remain an indispensable part of the development lifecycle.
Limits Across Complex Tasks and Benchmarks
The reliance on human intervention becomes more pronounced as tasks grow in difficulty. An evaluation of 1,389 internal runs using DeepMind's Argon agent revealed that autonomous success rates dropped sharply as task complexity and duration increased. A survey of twelve senior DeepMind researchers supported this finding, indicating that automation is most successful for well-defined, verifiable assignments, but drops significantly when applied to open-ended scientific ideation, creative research, novel experiment design, long-term planning, and diagnosing subtle training instabilities.
This plateau is also evident on standardized machine learning evaluations. On DeepMind's internal benchmark, RE-Bench—which consists of 74 realistic machine learning engineering tasks—the Argon agent successfully completed approximately half of the challenges. While this marks an improvement over Gemini 3.7 Flash, which solved 27 percent, it remains well below the near-flawless execution required for autonomous operation.
Steady Progress Rather Than an Intelligence Explosion
To evaluate long-term trends systematically, DeepMind created a General Capability Index (GCI) assessing 123 models across 92 benchmarks. The data indicates that capabilities in the Gemini family have advanced at a steady rate of roughly 0.1 GCI points each month, showing no statistically significant signs of exponential acceleration. While the Gemini 4 Argon model sits slightly above the historical trend line, the team attributes this bump to increased compute and organizational investment rather than automated, AI-driven breakthroughs.
DeepMind currently classifies the risk of AI-driven research automation as low, though it continues to expand measurement tools to monitor future trends. This measured assessment contrasts sharply with warnings from other prominent AI figures. A recent research paper signed by more than 20 scientists—including Geoffrey Hinton, Yoshua Bengio, and OpenAI's Jakub Pachocki—warned of a potential "intelligence explosion." Those researchers suggested that systems could fully automate AI development within a few years, compressing years of traditional engineering work into mere months.
What it means for developers
For software engineers and machine learning practitioners, DeepMind’s findings provide a grounded perspective on the current state of developer tools. AI models are exceptionally capable at drafting boilerplate code, handling predictable functions, and executing clearly verifiable tasks. However, claims that software development will soon be entirely automated without human involvement are not reflected in internal frontier lab data.
Developers should view AI as an assistant rather than a replacement for architectural design, critical debugging, and quality assurance. Because success rates decline on complex and open-ended workflows, human review remains the primary safeguard against subtle bugs and regressions. For teams integrating these tools into production, comparing model behavior across different types of tasks is crucial. Developers looking to benchmark these capabilities can access top AI models like Gemini, GPT, and Claude cheaply on a pay-per-token basis through one key at https://apixoai.online, allowing them to evaluate practical model limits directly within their own engineering stacks.
Source: Google says self-improving AI isn't anywhere close — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
