NVIDIA AVO System Runs Autonomously for Seven Days and Outperforms Human-Written Code
Evolutionary code optimisation on a Blackwell B200 GPU beats cuDNN and FlashAttention-4 with zero human intervention
The AVO system follows the same evolutionary template as Google's AlphaEvolve from 2025, where candidate code solutions are mutated, tested against a fitness function, and iteratively improved through selection pressure. However, AVO advances the approach by promoting the AI agent from a simple candidate generator to a full variation operator that proposes, repairs, self-critiques, and verifies solutions before submitting them.
The researchers describe their philosophy as "blind coding," with lead author Bing Xu stating bluntly that "human cognitive ability is the bottleneck" in software engineering. The system consulted the full lineage of previous solutions and hardware documentation during its autonomous run.
This follows Google's AlphaEvolve achieving similar breakthroughs in matrix multiplication and Ramsey number bounds last year, suggesting evolutionary code search is becoming a repeatable paradigm rather than a one-off demonstration.
Analysis
Why This Matters
When AI systems can autonomously produce code that outperforms code written by the best human engineers at one of the world's top hardware companies, it signals a fundamental shift in how performance-critical software will be developed.
Background
Google's AlphaEvolve (2025) demonstrated the concept. NVIDIA's AVO (2026) pushes it further with more sophisticated agent architectures and longer autonomous runs.
Key Perspectives
Optimists see this as a force multiplier for engineers. Pessimists hear "human cognitive ability is the bottleneck" and see the writing on the wall for certain categories of engineering work.
What to Watch
Whether this approach generalises beyond kernel optimisation to broader software engineering tasks, and how quickly it moves from research to production tooling.