Deployment lag is the only safety margin

The capability question is already settled for about a third of the economy. What is still protecting those jobs is how slowly anyone actually adopts anything.

Zac Wickstrom · HashCash Labs · October 2026

Most arguments about AI and work are arguments about capability. Can the model do the job or not. I think that's the wrong variable. LLMs are now permanently more knowledgeable than any human you will ever meet. Capability is not the question. So why haven't they taken over the world already? Why are companies still hiring? Many enterprise companies have been stuck in what's been called "AI onboarding purgatory" for years now. Capability has come detached from deployment, and by a very wide margin. That gap is all the adjustment time we have left as a civilization.

Deployment lag is the time between the first real demonstration that a machine can do a task, and the point where machines are doing a meaningful share of that task for money. Demo to payroll.

For example: a model passed the bar exam years ago. The number of lawyers didn't move. Something is sitting in between. That something is the only thing preventing economic apocalypse in October 2026.

Most Americans don't have a laptop job

The exposure question isn't about intelligence; it's about how much of a job is purely digital. It's "do you have a laptop job?"

So we sorted the 22 BLS occupational groups into three tiers. Shares are from BLS OEWS, May 2025.

And by "exposed", I don't mean the job vanishes. The people who stay are the ones who handle the weird cases. My first pass assumed 90% of exposed work goes. I now think 60 to 70% is more realistic, because the exceptions need a person.

The tiers are judgment calls on whole groups, and the within-group splits are ours. The total is a reasoned range, not a forecast. It also assumes nothing beyond technology that already exists.

For scale: US unemployment peaked near 25% in the Great Depression. But, of course, those jobs came back...

So why hasn't it happened yet

Because the lag isn't one number. It's different by an order of magnitude across those three tiers, and the reasons have nothing to do with how smart the model is.

The Three Barriers

1. Physical. A capable humanoid is a manufacturing problem, not a software problem. Replacing physical labor means building machines at car-factory volume. That's a decade-scale industrial buildout even with the design done and the money committed.
2. Legal. A model that can diagnose isn't a model that gets to diagnose. Someone has to carry the insurance. Every licensed trade has a legal person on the hook. Putting a machine in that chain is a legislative project in fifty states, not a product launch.
3. Political. The slowest brake — and the one nobody models. No politician gets votes for reducing the number of government jobs. Public sector work is protected not because machines can't do it, but because "creating jobs" is what has driven politics for a century.

So the number worth watching isn't Terminal-Bench 4.0 scores. It's how quickly the physical, legal, and political barriers fall.

For the physical, we now have robots playing Chopin. For the legal, NHTSA has proposed letting fully driverless cars ship without brake pedals, and says steering wheels could be next. And for the political — well, unless you were at USAID, you might be safe for a while. Fingers crossed.

I used to think robots were the slow variable

Hands were the hard part. Backflips are a solved control problem. Fast two-handed work with fine force control wasn't, and that's what a plumber needs. It's also what a pianist needs.

And now there's a robot playing Chopin on stage at Beijing's World Robot Conference. As a pianist, I can't overstate how much this blows my mind. Chopin isn't a party trick. It's exactly the kind of fast, fine, two-handed control that was supposed to be decades out.

Tesla's newest Optimus hand has 22 degrees of freedom, plus three more in the forearm, and it's being built to feel what it's holding. Tesla's higher-volume Optimus factory in Texas is slated for 2027, and I actually believe they'll hit that date. The $30,000 robot butler will take a couple more years, but it'll come the way the Model 3 came: probably by 2029.

Once the hardware is that dexterous, the rest is software. An over-the-air update could wipe out a trade overnight, probably in the early 2030s. Plumber.exe.

The emergent stabilizer

This is the part I find most underrated, and it runs opposite to the usual panic.

Screen work goes first, because it has no brakes. That puts a lot of capable people into the market looking for work. Which makes human labor cheaper. Which raises the bar a robot has to clear to be worth buying.

So the first wave actively slows the second wave. If labor gets cheap enough, the person who could afford a robot hires a human instead. That's a real negative feedback loop sitting inside a process everyone models as a clean exponential.

It doesn't help anyone. It isn't good news. It just means the two curves don't compound the way people assume.

It also predicts something specific and ugly. The damage concentrates on new entrants. A displaced senior analyst has a network and a track record. A twenty-two-year-old with a fresh degree has neither, and the rung they were going to climb was made of screen work. They majored in computer science. So I think youth underemployment is the leading indicator, well before any aggregate unemployment number moves. It won't look like mass layoffs. It'll look like a hiring freeze that never ends. And, of course, we're already seeing that.

Why any of this matters for safety

Almost every practical safety argument is an argument about buying time. Time to build evaluations. Time to understand what these things are doing. Time for institutions to catch up.

Deployment lag is that time. And it isn't a policy. It's an accident of supply chains, insurance law and who gets blamed.

I just want to emphasize: 33–50% of all jobs that exist can be done by LLMs today. Deployment lag is the only reason they're still here.

Hot take: policy aimed at capability is aimed at the wrong variable. An "AI slowdown" is never going to happen. It's a multinational prisoner's dilemma. Compute thresholds and training-run rules govern what gets built. Such policies, and the conversations about implementing them, do almost nothing about the gap between what's built at frontier labs and what's actually deployed into the economy. And that gap is the wall clock on our escape room.

I could be wrong

I almost want to say something like "a claim that cannot fail isn't worth making," but if that were true, every proven theorem would be such a claim. And I don't really believe that. But of course, this blog post is a blog post, not a theorem. I could be wrong. So here's what would break this one.

One caution on that second test: the unemployment rate only counts people who are still looking for work. Young people who give up don't show up in it at all, so a calm headline number can hide them. The numbers to watch are labor-force participation and the share of young adults neither working nor in school.

The measurement I'm working on

Nobody is measuring this directly, which is strange given how much rides on it.

Take fifty task classes across the three tiers. For each one record two dates. First credible public demo of the capability. First year machines did a measurable share of that task in production. Publish the spread.

What comes out is a distribution, and I think the variance would be interesting to see. Partially because it'll likely render the law of accelerating returns as a scatter plot, and I'll briefly feel like Ray Kurzweil when I look at it. If the lag is six months for text and fifteen years for licensed physical work, then "when does AI take the jobs" has no answer. But personally, I suspect there won't be any meaningful work for humans by the mid-2030s. For once, government bureaucracy might be doing us a favor.

Sources

• BLS: Major occupational groups as a percentage of total employment, May 2025
• TRT World: "Cyborg Chopin" at the World Robot Conference
• The Decoder: Optimus' new 22-degree-of-freedom hand
• Transport Topics: NHTSA proposes dropping the brake-pedal requirement for driverless vehicles
• Electrek: Optimus production timeline, including the 2027 Texas factory