Import AI 464: Fable writes GPU kernels; AI automation; and analog computation

Two numbers worth sitting with

Most weeks in AI produce a lot of noise and very little signal. Every so often a week produces two numbers that, read together, are hard to shrug off. This was one of those weeks, and we think the people running real businesses should understand what the numbers actually say, past the headlines. We also think the honest reading of them is more encouraging than alarming, provided you act like a builder rather than a spectator.

Here they are.

First, an index that measures how much genuine paid remote work AI can complete on its own, the kind of jobs people hire freelancers for, moved from 2.5 percent in October to about 16 percent now. Second, an AI system produced the fastest known submission to a demanding public benchmark for GPU kernels, the low-level code that makes chips run AI models, beating the standard human baseline by more than eighteen times and beating every rival model in the same contest.

One number is about AI doing our work. The other is about AI improving the machinery that runs AI. That combination is why this week matters more than most, and why the opportunity in front of well-run companies is larger than the risk.

What the labor number really measures

The reason the first figure is worth attention is that it resists the usual benchmark trick. Plenty of tests reward a model for answering quiz questions or passing exams, tasks that look impressive and translate poorly into value. The Remote Labor Index instead uses work with an invoice attached: 3D design, animation, data analysis, building small web applications. Real deliverables that a paying client would inspect and reject if they were wrong.

Against that harder bar, AI cleared one in forty tasks last autumn. It now clears roughly one in six. The rate of change is the story. A capability that improves by a few points a year is a curiosity. A capability that grows six-fold in three quarters is a planning opportunity, because it means the set of tasks you can safely hand to a machine is expanding faster than most operating plans assume.

We want to be precise about what this does not mean. One in six is not five in six. Most of these jobs still fail. Completing a task once in a controlled test is a long way from delivering it reliably, unsupervised, at the volume and consistency a business requires. Anyone selling you inevitability is overselling. But anyone telling you this is a toy is not reading the slope, and reading the slope correctly is where the advantage lives.

Why the kernel result is the one to watch

The second number is quieter and, in our view, more encouraging.

Writing a fast GPU kernel is not glamorous work. It is the deep plumbing of modern computing, the code that decides how efficiently a chip turns electricity into results. It has long been the domain of a small number of highly specialized human engineers. An AI system just did it better than the reference humans and better than its peer models.

Sit with the implication, because it is a good one. The tools are getting better at building the tools. When a system can improve the very infrastructure that makes systems like it faster and cheaper to run, the cost of capability falls and keeps falling. For a business, cheaper and better infrastructure is not a threat. It is the tailwind that turns yesterday's impossible project into next quarter's line item. The first turn of that flywheel is now visible in a public benchmark, which is a very different situation than a year ago when it was a thesis on a whiteboard.

How we think this actually evolves

We are in the business of putting these systems to work, so we watch the trajectory closely rather than the headlines. Here is our fact-based read on where this goes.

AI keeps taking the floor. The routine, the repeatable, the high-volume work that nobody built a career hoping to do, moves steadily toward automation. This does not empty out companies. It reprices human attention. The people and firms that learn to direct these systems move up the stack, toward judgment, client relationships, taste, and the messy context a benchmark can never capture. The failure rate that looks like a weakness today is really a map of where humans still add the most value, and that map is the most useful planning document a leader can hold.

The winning posture is neither the optimist who assumes full automation next quarter and overspends, nor the skeptic who assumes nothing changes and wakes up outcompeted by a leaner rival who adopted early and learned the failure modes first. The winners treat this as an engineering reality to be measured. Which tasks in your operation are already in reach at acceptable quality. Which are close. Which are nowhere near, and why. That map is worth more than any forecast, and it changes every few months, which is precisely the point and precisely the advantage for teams that keep it current.

The honest, optimistic reading

At Skylab we make our living helping companies adopt these systems, so we have every commercial reason to hype this and every professional reason not to. Credibility is the only asset that compounds in this field, so we will say plainly what we believe.

The capability curve is real and it is steep. The failure rate is also real and still high. Both are true at once, and the good news is that this is exactly the condition in which prepared companies pull ahead. When a technology is powerful but uneven, the edge goes to whoever measures carefully, adopts deliberately, and compounds small wins while others argue about the endgame.

For most of computing history, the reasonable default was continuity: next year looks much like this year, only a bit faster. When the tools begin improving the tools, that default gives way to something better for the people who prepare, a period where a modest, disciplined head start compounds into a real lead.

The useful question was never whether the shift arrives on some particular date. It is what your business wants to be standing on when it does, and that is a decision you can make now, deliberately, while the numbers are still small enough to experiment with cheaply.

If you want help drawing that map for your own operation, honestly and without the hype, that is the work we do.