Fable writes first GPU megakernel, and experts see an AI self-improvement loop

On July 6, 2026, AI company Fable published a result that Anthropic co-founder and Import AI newsletter author Jack Clark called "the start of a recursive self-improvement loop": its model independently created the first true GPU megakernel ever built, and, according to its authors, the fastest that has ever existed.

What is a GPU megakernel and why is it hard?

A GPU kernel is a program that runs simultaneously across thousands of graphics processor cores. Writing efficient kernels is considered one of the hardest tasks in infrastructure programming: an engineer must understand in detail the architecture of a specific GPU, memory access patterns and numerical methods. That is why most AI frameworks, including PyTorch, JAX and TensorFlow, run on top of kernels that NVIDIA specialists and small teams such as the FlashAttention developers have hand-tuned for years.

A megakernel is the next level of complexity: it fuses several operations into one, eliminating intermediate memory accesses. This can reduce latency and raise throughput when training large models. Before Fable's work, writing megakernels remained the preserve of a few rare experts.

• July 6, 2026: date the result was published in the Import AI #464 newsletter
• The first "true" megakernel written by an AI model
• Claimed to be the fastest megakernel as of the publication date
• Assessment by Jack Clark, co-founder of Anthropic
• The term "RSI loop" applied to a concrete engineering achievement for the first time

Why is it called an "RSI loop"?

Recursive self-improvement (RSI) describes a situation in which AI starts accelerating its own development: a smarter model optimizes GPU computation → faster GPUs make it possible to train an even more powerful model → that model writes even better kernels. A classic positive feedback loop.

For years, RSI remained conceptual: philosophers and AI safety theorists discussed it, but no one pointed to concrete examples. Clark, one of the industry's most cautious commentators, did not soften his wording and directly called Fable's work the "start" of such a loop. For a seasoned observer, this is an unusually excited assessment.

"The start of an RSI loop" is how

Clark characterized the result in Import AI #464 on July 6, 2026.

It is important to grasp the scale: this is not about AI "slightly speeding up" code writing. Megakernels are the infrastructure foundation on which the next generations of AI systems are built. If models can reliably produce such code, one of the fundamental bottlenecks goes away: the shortage of engineers able to optimize computation at the hardware level.

What this means

Fable's achievement signals that automating AI development is moving from abstraction to practice. If independent benchmarks confirm the claimed performance of the megakernel, the event will go down as a turning point: the moment AI began seriously automating its own infrastructure. The industry should watch whether other labs, such as Google DeepMind, Meta FAIR and Mistral, adopt this approach or try to get ahead of Fable.

*Meta has been designated an extremist organization and is banned in Russia.