Fable creates first GPU megakernel — experts discuss AI self-improvement loop

Fable Wrote the First GPU Megakernel — Experts Talk About 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 beginning of recursive self-improvement loop": its model independently created the first truly real megakernel for GPU in history — and, according to the authors, the fastest one ever created.

What is a GPU Megakernel and Why is it Difficult

A GPU kernel is a program executed simultaneously on thousands of cores of a graphics processor. Writing efficient kernels is considered one of the most complex tasks in infrastructure programming: an engineer must understand in detail the architecture of a specific GPU, memory access patterns, and numerical methods. That's why most AI frameworks — PyTorch, JAX, TensorFlow — work on top of kernels that NVIDIA specialists and small teams like FlashAttention developers have refined for years.

A megakernel is the next level of complexity: it combines multiple operations into one, eliminating intermediate memory accesses. This reduces latency and increases throughput when training large models. Before Fable's work, writing megakernels remained the domain of rare experts.

• July 6, 2026 — date of result publication in Import AI newsletter #464
• First "real" megakernel written by an AI model
• Claimed as the fastest megakernel at publication date
• Author of assessment — Jack Clark, co-founder of Anthropic
• Term "RSI loop" first applied to a concrete engineering achievement

Why is it Called "RSI Loop"?

Recursive Self-Improvement describes a situation where AI accelerates its own development: a smarter model optimizes GPU computing → faster GPU allows training an even more powerful model → that one writes even better kernels. Classic positive feedback loop.

For many years RSI remained conceptual — it was discussed by philosophers and AI safety theorists, but not referenced with concrete examples. Clark, one of the most cautious commentators in the industry, did not soften his wording and directly called Fable's work "the beginning" of such a loop. For an experienced observer, this is an unusually excited assessment.

"The start of an RSI loop" — how

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

It's important to understand the scale: this is not about AI "slightly speeding up" code writing. Megakernels are the infrastructure foundation on which next-generation AI systems are built. If models can consistently create such code, it removes one of the principal bottlenecks: the shortage of engineers capable of optimizing computing at the hardware level.

What This Means

Fable's achievement is a signal that automation of AI development is moving from abstraction to practice. If independent benchmarks confirm the claimed megakernel performance, this event will enter history as a reference point: the moment when AI began seriously automating its own infrastructure. The industry should watch whether other labs — Google DeepMind, Meta FAIR, Mistral — pick up this approach or try to surpass Fable.

*Meta is recognized as an extremist organization and is banned in Russia.