"What are you doing for the last 300 days of work?"

First, the tweet:

This is all the context we get.

Kevin Roose is a NYT tech columnist, host of a tech podcast, and author of several books. He is situated in San Francisco so he is positioned to credibly have actually been in a frontier AI lab. Everything else is speculation or inference on my part.

We don’t know if this alleged quotation came from OpenAI, Google, Anthropic, or maybe even a smaller AI lab (note: he did not say “frontier AI lab” which may or may not be salient.)

However, if we take his claim at face value, there is some signal here. What could it mean?

The first thing to understand is that Silicon Valley insiders are often delusional. Not in the psychiatric sense, but in the pragmatic sense. This is also partly true of any specialist. Ask a developer what “automating everything” constitutes in their world, and they’ll talk about CI/CD pipelines. Ask a masonry specialist what “automating everything” looks like in his world, and you’ll get a very different answer.

I bring this up because I have personally experienced this. Even within the world of technology, everyone only has their particular viewpoint, and people generally underestimate how complex the rest of the world is.

Therefore, my first primary inference about this quotation, if it is true, is that the speaker was probably referring to “saturating agentic capabilities.” And that’s not nothing. “Agentic capabilities” is tech jargon for “can do anything on a computer that a human can.” This is what I’ve long called KVM jobs, meaning “keyboard-video-mouse.”

The trends are pretty unambiguous.

Computer-using-agents (which were briefly called CUA, but simplified to just “agents”) have been ramping up quickly.

Just this year, AI models have surpassed humans on basic computer use as agents. But that’s not the whole story

Over the last 12 months, we’ve crossed a threshold, and you can see that we went from “basically useless” at the beginning of 2025 to “better than humans” in early 2026, and the models are still improving.

So when a frontier AI researcher thinks “we’re only 300 days from automating all work” they are probably looking at data like this.

However, there are many barriers between “what is possible in the lab” and “what actually gets deployed, integrated, and used.”

This underscores the problem with benchmarks. They are usually inadequate, and the benchmarks themselves become a moving target.

Jeff Bezos famously said “if the data disagrees with the anecdote, go with the anecdote.” In this case, the data would suggest that we’re close to saturating long-video understanding, and you would not be entirely wrong for believing “this also moves us closer to saturating KVM jobs.”

However… the C-suite hasn’t spoken yet

In my model of “lab—to—workforce” deployment, there are three basic steps. Maybe “major phases” is a better term.

The first phase is “hypothetically saturated in the lab.” There are lots of things that are possible in controlled environments that do not survive contact with reality. We’re constantly bombarded with science news like “scientists develop vaccine that cures 853 strains of cancer!”

(And then it never translates to anything in the real world)

AI and tech is no different from biotech and pharmaceuticals in that respect.

So phase 1 is “technically capable.”

But that leaves phase 2. Which is “executives buy in.”

Anyone who’s worked for an enterprise, a Fortune 500 company, or even a mid-sized shop that takes itself seriously knows that the C-suite is very cautious. Business executives are, first and foremost, risk managers.

Disruptive tech often translates to costly decisions. Furthermore, executives are constantly being sold shiny new whiz-bang gizmos from vendors. They are rightly skeptical. My wife’s small 30 person company was being pitched Gemini Enterprise by Google and it was quite the flop.

Your CEO is always going to be skeptical of any specific product even if they already know that AI is the way of the future.

Likewise, your CTO may even be the biggest skeptic. Why? Because they are disaffected and disappointed. Some of them have their heads buried in the sand, but that’s a different conversation. CTOs (in my experience) are often among the heaviest AI users, and they know first-hand that it is MASSIVELY transformative.

But it’s one thing to transform one job, when you’re tech savvy, and have the freedom to use it all day, every day. It’s another thing entirely, to scale that up to hundreds or thousands of employees.

Then, CFOs are an entirely different breed. A $50 per month seat for Claude or ChatGPT may very well pay for itself. But the CFO will ask “Okay sure, but is that the absolute best use of that money?” And from a technologist’s perspective, the answer is “fucking duh.” But that’s not how reality works.

The final phase, which arguably could be phase 2, is “product integration.” I put this at phase 3 because sometimes the integration is done by the consumer, and sometimes it’s done by another vendor.

Let me give you an example.

My wife uses tools like Canva and Slack at her work. Neither Canva nor Slack is an AI-native company. But they are deploying AI and agentic capabilities into their platforms. In this case, the vendor makes AI adoption all but automatic. My wife loves these tools. She’s quickly becoming a marketing automation guru. But she ALSO works at a company where the CEO is already all-in on AI and has made AI adoption and integration a first-class business priority.

However, in many other cases, “product integration” is about the end users becoming familiar with the products directly. Treating AI literacy as a first-class competency has yet to materialize in most businesses.

Why?

Skepticism, fear, HR, and legal.

In my consultation experience, HR and Legal often use “rules for thee, but not for me” when it comes to AI. They issue severe warnings about how much trouble you can get in for using AI, while giving themselves free passes.

As one executive told me years ago “What Legal wants, Legal gets.” When the lawyers decide that AI is fine and safe for them to use, they usually get their way. Ditto for HR departments. In fact, there’s a trend where some companies that are not tech-centric are rehousing IT under HR, which makes no damn sense to me, until you realize that technology is just considered an enabling business expense for most companies… (and this is another reason that Silicon Valley tech bros are delusional, they just have no idea how complex most business is)

The robots are coming!

KVM jobs are one thing, but what about the robots? Production of industrial and domestic humanoid robots is ramping up. They are real, and it is happening.

But…

They are not that bright yet. Most home use robots use teleoperators as fallbacks. This is pretty cool, and that will also serve as training data. Robots can also learn by mimicking humans. They can “one shot” learn by having you demonstrate a task once. This pattern will, ultimately, be remembered as the dark ages of humanoid robotics.

Hyundai is planning on deploying 25,000 Atlas humanoid robots at its plants. You don’t sign a contract that huge unless you expect real ROI.

Like agentic models, robot models are also rapidly reaching saturation. But the data shows something else entirely. Here’s a bunch of graphs showing where we’re at.

First, LIBERO is about transfer learning in skills so that robots can hit the ground running.

This first benchmark looks positively saturated, but like LLM benchmarks, it proved highly inadequate to actually pass the threshold of “actually useful in a factory.”

This data shows that the previous saturation was mostly over-fitting the data or rote memorization.

As with agentic computer use, we’re finding that these problems are always harder and more multidimensional than they first appear to be.

At the present time, in late May 2026, only two benchmarks are meaningfully saturated, but other basic functions are still in their toddlerhood.

However, recall that it only took OpenAI twelve months to go from 40% on computer-using benchmarks to fully saturated. Likewise, Jim Fan of Nvidia declared that robotic foundation models would be complete by 2027 or 2028, so we’re right on track.

Could all this be ready (in the lab) in 300 days?

Yes, I think that humanoid robotics and computer using agents could be fully or mostly ready for primetime… at least in the lab. Ten months in “AI years” is equivalent to decades in other technology domains. And every “wall” that we’ve hit has been either entirely illusory or routed around within months.

Diffusion still takes time.

It will take months, if not years, to fully deploy AI agents. There will be laggards in the industry, and rewiring large organizations is painfully slow and expensive. Plus, individual workers need to adapt.

But, is this “the last 300 days of work”?

Absolutely not. Will the next 300 days move us closer to a post-labor future? Yes, absolutely.