You. Are. Not. Using. Enough. Compute.

Imagine one intern in your office.

You hand them five tasks: spell-check a proposal, fact-check three bullet points, draft an email, summarize a PDF, and book a trip. One hundred sixty-three minutes later, they return with everything done.

Let's call that bundle of work a Knowledge-Work-unit (KW-unit): five everyday tasks packaged into a single metric.

Your intern produces 1 KW-unit every 163 minutes, then sits idle until you give the next assignment.

Now upgrade to five interns. Hand them the same five tasks each, in parallel—and they finish in just 20 minutes. That’s 1 KW-unit every 20 minutes. Super-productive… but if you don’t keep feeding them work every 20 minutes, you’ve got five interns staring at you. So you keep them busy.

Scale to 150 interns, and, in theory, they crank out a KW-unit in seconds. But they don't. You become the bottleneck as 150 interns sit, crammed in your office, idling.

You should feel awkward. But in this scenario, you don't. Pretty soon, it becomes normal and you begin ignore the interns altogether, only tapping them when you need something done and they complete it instant. You get used to that workflow. You don't even notice that their hourly rate is dropping and more keep showing up every day.

But you don’t have five interns.

You have a laptop idling at 10% CPU all day.

And infinite cloud compute yearning for a button press.

We work like digital neanderthals—click through tabs, ask ChatGPT a handful of questions, paste snippets into your editor—reacting step by step, serially. We're the bottleneck, and we're perfectly fine with the 40 million “interns” sitting idle, waiting for you to twitch your fingers.

If you could coach yourself, you’d say: delegate and elevate. But, by default, we don't coach ourselves with these things, we just get old and the new generation makes fun of us.

What if, instead, we did coach ourselves.

What if our interns took initiative—anticipating our next move and researched, coded, tested, and shipped things themselves. Kicking off thousands of KW-units, then handing you only the best results? What if they explored the hard problems 1,000 times while you were on a customer call, so that when you returned, the solution was already in your inbox?

I wanted to explore this for myself, so I've been reflecting on the eras of compute and how they shaped our habits. I wanted to understand how we got here, why we work the way we do, and how we can change our habits to better leverage the compute available to us.

Below are my notes. I hope it's helpful for you too.

Wait, what? Memory prices are going up in 2025? Yup - DRAM is headed for a second straight annual price jump largely because AI is inhaling every gigabyte manufacturers can stack. Practically speaking, this is the first time 2-year blip has ever happened.

The Flippening

In the PC and Web-2.0 eras, computers were accelerants—speeding up fragments but unable to carry work end-to-end. So you built habits around handing off bits and pieces: a spell-check here, an autocomplete there, a quick search when you hit a wall.

You couldn't give your "interns" to own whole tasks, they just weren't capable. Maybe, if you were lucky in a specific role the compute could handle 80% of a specific KW-unit relevant to your work. But you still had to fill in the gaps, patching together results from multiple sources.

But with todays AI "the flippening" happens. With reasoning engines that have access to tools and apis, you are no longer the gap filler, the patcher, the one who stitches together results.

Decades of software has to be rewritten. Human in the loop is cope.

Your habits and instincts must change. You can no longer afford to be the bottleneck.

Delegate and elevate.

The Flippening in Action

In the back of an Uber, I glanced at my phone and said, “Explore the character-management UI.” No IDE. No branch checkout—just that one line. Just ChatGPT's codex hooked up to a Github repo.

AI had transcribed and structured my request (≈ 1 KW-unit in 75 ms). I hit submit, fanning that concepts into 4 coding agents. All racing to build what I said. Then four code-gen agents kicked in, each spinning up a full React component (tests, styling, documentation) in parallel. About 8 minutes later I have 4 different yet fully complete character-management UIs (24 KW-units in 8 minutes). Github PRs open, automated CI/CD pipelines ran tests, builds and deployed 4 live Vercel previews (≈ 3 KW-units). Copilot code review agents scanned the new code, checking for bugs and style issues (≈ 2 KW-units).

In ~10 minutes, from my phone, "I" produced 30 KW-units of work. Approximately 2 weeks of work.

I also had I had 4 fully functional, tested, deployed UIs ready to choose from. I browsed from my phone, selected a winner and provided feedback to tweak it to my liking, by voice. Rekicking off a process for 30 new KW-units to be spent in another ten minutes.

My uber arrives just as I finish selecting the UI I like best. I merge the PR. Throw out the rest, burning the compute that was used to generate them - what would have cost me a literal month of work and was done in an uber for free. Not to mention the human ego bruising that was avoided by tossing ai work.

I never opened an editor or typed a line of code. In fact for the MAJORITY of the uber trip, I was doing something else. Only spending about 5 minutes of my time on the topic.

Could I have squeezed out another 60 KW-units in that time? Undoubtedly. In fact, if I was perfectly efficient I could have done 120 KW-units in that time.

This is the flippening: I am bottlenecking the compute, not the other way around. I am the one who is slow, who can only handle so many complexity at once.

Now what?

Could I have an AI deep research my code base, find missing functionality, deep research the best practices for product design, write a report, and generate PRDs and automate myself out of the uber trip? Yes.

What do you do with that information? To be honest, I'm not sure.

But I know two things:

1. You are not using enough compute. You are the bottleneck, not the compute.
2. You need to change your habits. You need to learn how to delegate and elevate. You need to learn how to be a CEO of an intern army and nothing less.