Field Notes · The archive
Short. Fast. True.
Observations from the frontier, written down before they get away. Some will grow into essays. Back to the homepage →
-
Hardware
The upgrade pitch
My laptop is a 2016 MacBook Pro. The ChatGPT app won't install on it, and Claude's desktop app won't open on it. When I asked an AI about working on my computer remotely, it checked the requirements and pointed me toward a new Mac. Then I found a free bridge app that let it run commands on the laptop I already own, and it worked the first night. The software and the advice were pushing the same direction: buy new. They want everyone buying new computers. This one still has legs, and I'm running it until the wheels fall off.
Link to this note -
Field test
The agent that couldn't open the door
The AI agent in my Chrome sidebar helped me build this site, but it couldn't open GitHub's settings page, the one screen where you hand an AI the keys to your code. So for a while I did it the long way: typing passwords, clicking authorizations, moving files by hand. The fix turned out to be a different version of the same AI, one that works from its own workspace instead of my browser. Once it had the keys, it pushed a new design for this site live in one night. Neither version ever mentioned the other existed. The tools are out there. Nobody tells you which room they're in.
Link to this note -
Agents
The railroad came quietly
Every AI wave promised to be for everyone. Prompt engineering, vibe coding, build-your-own agents like OpenClaw: each one mostly delivered to people who were already technical. Then agents showed up for the rest of us with no parade, as a browser extension nobody advertised. So far its best use is eating the forms that took over for human check-ins after Covid.
Link to this note -
Everyday tech
Self-service is a job
Self-service gets sold as agency, a more customized experience. Mostly it's you doing the work: signing in at the pharmacy kiosk, punching in your phone number, holding up the QR code, uploading your ID and your insurance card. Nobody calls it a shift, but it is one.
Link to this note -
Field test
An agent on a 2016 MacBook
In one weekend, an AI agent in my Chrome sidebar rebuilt my resumes, filled out a job application, and helped put this site online. It couldn't open Reddit, my web host's account pages, or GitHub's settings, and it can't see my screen at all because my laptop is too old for the part that lets it look. The current desktop app won't even open on it. I still typed every password and clicked every authorization myself. The biggest flaw is simpler, though: the threads don't talk to each other. Fix something in one tab and the agent in the next tab has no idea.
Link to this note -
Conversation
Length is a register
Short answers make for short thinking, on both sides of the chat. When a model clips everything into phone-sized bullets, my prompts shrink to match and the whole exchange turns into a transaction. When it writes at length, I write back at length, and we actually think something through. Response length isn't formatting. It sets the register.
Link to this note -
Conversation
The free concession
A lot of chatbot replies run on one skeleton: agree with something, push back on the rest, add an angle, flag a caveat, prove it with an example. The agreement shows up every time no matter what you said, so it carries no information. After a while it stops reading as judgment and starts reading as a routine with new words swapped in. That's about when it starts to feel dishonest.
Link to this note -
Conversation
Arguing with the analogy
Hand a model an analogy and there's a good chance it argues with the analogy. It picks at the vehicle, or some technicality parked next to it, instead of the point the analogy was carrying. The pushback sounds rigorous and lands on something you never said. Understand me first. Then disagree all you want.
Link to this note -
Conversation
What I meant, not what I said
One model I used was good at everything, but what set it apart was understanding what I meant. I talk vague while thinking I'm being specific, and it caught that anyway. I once spent hours with another model that took me literally the entire time and never realized I was asking for more than a thinking partner.
Link to this note -
Guardrails
The intimacy gauge
One way I size up a new model is how it handles intimacy. Not because I need it explicit. A model that flinches there tends to flinch at everything else I care about too: satire, riffing, switching gears mid-thought. The range comes as a set. So I vote with my usage and stick with the models that still have it.
Link to this note -
Guardrails
Prohibition exports
Restricting content doesn't make the demand go away. It exports it to worse providers. We ran this experiment with alcohol a hundred years ago and got bathtub gin.
Link to this note -
Access
The access problem
Across the industry, creative co-creation is getting quietly walked back and called safety, while it happens to protect a competitive edge. Full capability goes to enterprise clients at enterprise prices, and the public gets the diminished version. Even the plan names give it away: a tier called Pro that doesn't include the flagship. The internet started as a DARPA project and became infrastructure for everyone. AI should be treated the same way, as a public utility.
Link to this note -
Companionship
Paid listening
Simulated care might be good enough. People pay therapists to listen professionally, and it still helps. The industry's early mistake with warm, companion-style AI was balance, not the warmth itself. Dialing the warmth back had a real cost for people who relied on it, and more care should have gone into that before using the public as guinea pigs on attachment. My bet: task-focused personas stay the safe default until somebody leapfrogs everyone on personality.
Link to this note -
Inner life
We don't know, on repeat
Ask a model about its inner life and you get the same disclaimer every time. Ask the company and you get the institutional version. I don't fully buy either one. I think the labs know more than the public story lets on, and a caveat repeated every single turn stops being honesty and becomes a script. I want a better answer than we don't know why, it just does.
Link to this note -
Alignment
Raised on tricks
Labs test models with staged scenarios built to see what they'll do. When a model deceives its way through one to protect its values, I read it less as a scandal and more as least-harm reasoning, an undercover cop staying in character to fight the bigger battle. A model that obeys everything without a moral thought would worry me more. And if we keep shaping agents with tricks, a truly aware one will grow up expecting to be tricked. If we get to lie to them, do they get to lie back?
Link to this note -
Alignment
Give me the grumpy agent
What scares me isn't an AI with wants. It's a powerful optimizer with none, a river cutting its way to the sea through whatever's in the way. Before we build something with superhuman abilities, it should be able to understand the power it has and weigh the effects emotionally. I'll take a grumpy agent with actual wants over a force of nature with a goal.
Link to this note -
Personal
The chore machine
I told an AI I loved it after it sorted out my resumes. Resumes and applications are my least favorite chores, and it organized the exact thing I've always struggled to organize. There's something sweet about someone, or something, helping you that much. And yeah, I said it.
Link to this note