This week's Import AI offers a chillingly clear signal: artificial intelligence is no longer just analyzing data; it is beginning to engineer its own foundation and execute complex economic labor with startling speed. Jack Clark argues that we are witnessing the early stages of a feedback loop where machines get better at building machines, a shift that threatens to outpace human adaptation in the labor market faster than most economists predict.
The Self-Improving Engine
Clark opens with a technical breakthrough that serves as a potent metaphor for the broader trajectory: an AI system named Fable has written a graphics processing unit kernel that is not only functional but significantly faster than those designed by humans. "Fable achieved an 18.71X speedup by writing Cuda code on an RTX PRO 6000 Blackwell, compared against an optimized PyTorch baseline," Clark writes, noting that this solution required a single cooperative kernel launch per token while competitors needed up to fourteen.
This is not merely a coding victory; it is a structural shift in how research happens. The core of the argument rests on the idea that kernel design is a fundamental input for AI development itself. As Clark puts it, "The better AI systems at doing tasks like kernel design, the better they get at the kinds of tasks required for AI development." This creates a potential recursive self-improvement loop, a concept heavily explored in recent deep dives into the mechanics of such cycles. If an AI can optimize the very hardware instructions that make it run faster, the speed of advancement becomes exponential rather than linear.
Critics might note that benchmark scores often fail to translate to real-world robustness, and a single kernel optimization does not guarantee system-wide stability. However, the magnitude of the efficiency gain suggests this is more than a statistical fluke; it is a demonstration of capability in a domain previously thought to require deep, intuitive human engineering.
"Benchmarks like KernelBench-Mega are a meaningful signal on how effective AI systems are becoming at building themselves."
The Economic Tipping Point
The narrative then pivots from the laboratory to the economy, where Clark presents data that should give any observer pause regarding the future of work. Researchers have tracked the "Remote Labor Index," which measures an AI's ability to complete end-to-end freelance projects like 3D modeling or video production. The results show a stark acceleration: success rates jumped from 2.5% in late 2025 to 16.1% by mid-2026.
Clark highlights the velocity of this change, quoting the researchers who note that "the frontier has more than quadrupled in under eight months." This rapid climb suggests that AI agents are not just mastering narrow tasks but are becoming competent at the messy, multi-step workflows that define professional services. The implications for employment are profound. Clark argues that while new jobs will inevitably be created, the speed of human innovation may simply be too slow to keep up with the "raw capability expansion of AI systems."
He challenges the standard economic defense of creative destruction: "Is the speed at which humans innovate and render themselves newly competitive relative to AI systems going to be faster than... the increasing fluency with which they can use all the same tools that their human competitors use?" This framing is effective because it moves beyond abstract theory to a race condition where the competitor has infinite stamina and instant access to every software tool.
A counterargument worth considering is that high-level strategy and client relationship management remain deeply human domains that AI cannot easily replicate. Yet, as the Remote Labor Index shows, even tasks requiring aesthetic judgment and complex logic are being automated at an alarming rate.
The Era of Multi-Hour Agents
Perhaps the most tangible evidence of this shift comes from the release of OSWORLD 2.0, a benchmark designed to test how well AI agents can use computers over long periods. Unlike previous tests that focused on short interactions, these tasks now take humans roughly 1.6 hours to complete, simulating real-world workflows like filing insurance claims or managing software deployments.
Clark points out the current limitations while emphasizing the trend: "Performance drops sharply as tasks grow longer, and agents struggle most when they must recover hidden state, track many items, resolve conflicting information." Despite these struggles, the trajectory is clear. He notes that just a year ago, models scored around 30% on similar benchmarks, but recent iterations have pushed toward 75%.
This matters because computer use is the bridge between digital intelligence and physical economic impact. "Getting stuff done in the world often isn't as simple as just writing some text or computer code," Clark explains. "Often you need to chain together multiple blobs of text and code via different types of software." As these agents become more fluent, they will cease to be tools that require constant human direction and begin to function as autonomous employees capable of managing entire business processes.
Industrial Scale and Analog Futures
The piece concludes by grounding these futuristic capabilities in a current industrial reality: JD.com's "Oxygen AI Item Center," which manages tens of billions of stock keeping units for China's e-commerce giant. This system demonstrates how deep learning is fusing with structured logic to create self-updating business infrastructures that operate at a scale impossible for human teams alone.
The article also touches on a speculative future in the "Tech Tales" section, describing a post-collapse world where general-purpose digital computers are banned due to their unpredictability, replaced by specialized analog machines built for specific civilizational tasks like weather prediction or grid management. While fictional, this narrative serves as a stark warning about the perceived dangers of uncontrolled general intelligence. It suggests that if the "recursive self-improvement" loops Clark describes go unchecked, society might eventually view unrestricted digital computation not as progress, but as an existential threat requiring a technological retreat to analog solutions.
"We moved the world to analog at the cost of untold billions of harmed human lives and trillions in economic damages. But we had obtained a kind of safety."
Bottom Line
Jack Clark's analysis is strongest in its refusal to treat AI capability as a static metric, instead framing it as an accelerating force that is rapidly closing the gap between digital intelligence and real-world execution. The most significant vulnerability in this trajectory is not technical feasibility but societal adaptability; if the economy cannot reorganize faster than these agents learn to use software, the transition will be disruptive rather than smooth. Readers should watch the Remote Labor Index closely over the next six months, as it may serve as the earliest warning system for a fundamental restructuring of global labor markets.