The First Crack in the AI R&D Wall

AI just wrote a faster GPU kernel than humans have. That’s not a party trick — it’s the opening move in a recursive loop that ends somewhere most people aren’t ready to think about.

Fables, a small AI lab, built what it calls the first genuine and fastest megakernel ever written — a GPU execution primitive at the most performance-critical layer in the entire AI stack. Jack Clark’s Import AI flagged this as “the start of an RSI loop,” which is the most important three words in his newsletter this week. RSI stands for recursive self-improvement: the moment AI systems can meaningfully accelerate their own hardware-level execution, the improvement curve stops being linear. GPU kernels are not glamorous. They are also the single most performance-critical layer in the entire AI stack. Whoever writes the best kernels runs the fastest models at the lowest cost.

Here’s the thing everyone is missing: the significance isn’t that the kernel is good. It’s that it was written at all. For years, the steelman for “AI is a tool, not an agent” was that AI couldn’t contribute to its own substrate — that the recursion would always require humans at the chokepoints. CUDA kernel optimization was exhibit A. It requires deep systems knowledge, an intuition for hardware pipeline behavior, and the ability to debug failures that produce no useful error messages. The humans who do it well are extraordinarily rare and extraordinarily expensive. Fables just kicked that stool out.

Everyone says the real AI risk is misalignment — some future system that wants things humans don’t. The opposite is closer to true in the near term: the risk is a perfectly aligned system that wants exactly what Anthropic or OpenAI or whoever controls it wants, compounding faster than any governance structure can track. If AI can write better kernels than humans, it can help design better chips than humans, which run better models than humans, which write even better kernels. The loop doesn’t require consciousness or malevolence. It just requires being slightly better at each step. Compound interest is the most powerful force in the universe too.

The business implications are more immediate than the existential ones. NVIDIA’s moat is partially human — the CUDA ecosystem is sticky because the people who know it deeply are rare and embedded in NVIDIA’s customer base. If AI commoditizes kernel authorship, the switching cost to AMD or custom silicon drops. Every hyperscaler’s silicon team just got more valuable. Every NVIDIA competitor’s timeline just compressed. Tenstorrent, Groq, Cerebras — they’ve been fighting uphill against a 15-year CUDA talent advantage. AI-assisted kernel writing is the ramp they needed. NVIDIA is not going away; they’re too far ahead. But “too far ahead to catch” just became “far enough ahead to be interesting.”

The deeper stakes are about who controls the recursion. Right now, a handful of labs have the compute, the data, and the researchers to sit at this chokepoint. If AI R&D automation compounds inside those labs faster than the outside world can observe, the gap between frontier and everyone else doesn’t widen gradually — it gaps. Regulators aren’t equipped to measure this. Open-source alternatives don’t have the hardware budget to keep pace. The entities best positioned to monitor AI-accelerated AI development are the ones doing it, which is a conflict of interest the policy world has not yet seriously grappled with.

One megakernel is not a revolution. But every nonlinear transition looks incremental until it doesn’t.

The question isn’t whether the RSI loop is real. It’s whether anyone outside three zip codes in San Francisco is paying attention before it matters.