Josh JohnsonNotes from JJ
The World Jensen Helped Build
A leather jacket, an engineering legacy, and the machinery that keeps our ideas moving.

Coming to you a little early this Sunday. After being late last week, I figured I owed you guys one.
This week has me thinking about the machinery behind the magic: the hardware Jensen helped build, the models I can afford to leave working, and the computer that keeps going after I walk away.
The Future Had to Run on Something

Jensen Huang has resigned as NVIDIA's CEO, president, and a board member. NVIDIA announced the departure on November 12, following a controversy that includes reported affairs and a board investigation into a suspected foreign intelligence effort. The investigation is ongoing.
What I want to acknowledge here is the engineering legacy. To me, Jensen is one of the greatest founders of our time. He helped build the machinery that made a whole generation of other people's ambitions possible.
NVIDIA helped pioneer the modern GPU for gaming: hardware built to do lots of calculations at once so games could draw richer worlds. That same parallel computing muscle found work in cryptocurrency mining and became foundational to modern AI. Gaming, crypto, AI—different uses for the ability to do an enormous amount of work simultaneously. CUDA helped developers put that hardware to work far beyond graphics.
If you want to see why that matters, watch the MythBusters Mona Lisa paintball demo from NVIDIA's NVISION 2008 event. Adam Savage and Jamie Hyneman contrast a machine painting one shot at a time with an array firing together to produce the Mona Lisa almost instantly. It's a wonderfully ridiculous way to explain parallel processing: some jobs get a lot faster when you can work on many pieces at once.
We talk about the software and the personalities selling it, but underneath the demo, the math has to happen. Jensen and his teams helped turn the machinery for better games into a platform people could build entire companies around. That has changed the world. And eventually it reached guys like me putting agents to work from a desk at home.
That took engineers, teams, and decades of conviction. Jensen didn't do it alone, but his leadership helped make it real. I can take the circumstances of his departure seriously and still have enormous respect for what he helped build. I do.
Thank you, sir.
Back to Building Budget Swarms with Gemini

Last week I said I wasn't in a rush to try Fable. I finally got around to it. It's good. As expected. I gave it a UI revamp, and a single agent made quick work of it. That's exactly the kind of job that makes me understand why people pay for the heavy stuff.
I'm still price conscious. Fable 5.2's standard API pricing is $8 per million input tokens and $40 per million output tokens. A useful result can absolutely justify that. It doesn't mean I want every part of every project running at that price.
I used to think of Pro as the DeepMind heavy hitter and Flash as its distilled, quicker little brother. Argon, Carbon, and Hydrogen feel more like different levels of what I'd once have called Pro—especially for work that needs to keep going. RIP Flash. Still available; just no longer the little brother in my head.
My new favorite setup for a large software build or rebuild is Argon managing a swarm of Hydrogen workers. Break the work into pieces, let the workers take their assignments, and keep it moving while I do something else. Google's Hydrogen and Orchestrator announcement gets into this division between deeper planning and more economical execution.
The warning from my own runs: low, low cost; slow, slow work. Great when I can leave it running all afternoon. Less great when I need it in ten minutes.
I've gone pretty far past caring who owns the top benchmark row. I care whether the work keeps moving without asking me to babysit every step, whether the result holds up, and what it cost to get there. Lately, this setup has been doing that for me.
Always-On Agents: Give the Job a Computer and a Schedule

I think even people who are pretty good at AI still get stuck on this part. They know how to write a prompt. They understand that a few prompts can become a workflow. But ask what they have running every Tuesday without them, and the answer is often: nothing yet.
I remember earlier in my career (potentially dating myself here) using macros and Windows Task Scheduler to make things happen while I was away. The scheduling part hasn't fundamentally changed. What's changed is what the worker can do once the alarm goes off: read something messy, make a judgment, use a website, and follow a more complicated process.
Here's the piece we skip when we make it sound magical: there has to be a computer available to do the work. When an engineer says “host” or “server,” that's the basic idea. A computer somewhere runs the job. A “cluster” is a group of computers working together. You don't need a cluster to get started.
A prompt tells it what to do. A schedule tells it when. A computer gives it somewhere to happen.
| 01 · SomewhereA computer available when the job runs. | 02 · SometimeA schedule that starts the workflow. | 03 · SomethingA useful result, delivered where you want it. |
A laptop can do it if it's powered, connected, and awake. The trouble is that we close the lid and take it places. That's why I keep talking about mini PCs: a little box that can stay put and keep working. Hosted compute is another option. For example, Cloudflare Workers has scheduled triggers for supported jobs. A hosted service runs on somebody else's computers; your laptop can be off. Pick the setup that supports the actual task—a web check and a desktop app don't need the same things.
Start with something you'd be glad to stop remembering. Check Rolex Explorer listings against your criteria. Back up your repositories weekly. Watch a flight price and email you when it drops. The job needs the context, access, and somewhere to put the result. It also needs a way to tell you when it fails.
Here's a prompt worth trying:
Help me set up a daily agent that checks these sources against my saved criteria and emails me only when there's a match. First, recommend where it should run and show me the plan. Include the schedule and time zone, required access, a spending limit, and a failure alert. Test it once before enabling the schedule. Keep purchases and submissions for my approval.
That's a real starting point. Ask for the recurring outcome, then have the agent help build the setup that can deliver it. You don't need to learn everybody's infrastructure vocabulary first.
The Takeaway
- The future needs machinery. I don't want to forget the people who built it.
- Buy the result you need. Sometimes that's one expensive agent; sometimes it's a slower team.
- Give the recurring job a home. A great prompt can't run on a sleeping computer.
— Josh