Blog

Big companies have a stupid problem.

They publish links. Lots of links. In documentation, training material, presentations, support articles, QR codes, emails, videos and probably a hundred other places.

Then they move things.

Microsoft has a nice solution to this: aka.ms.

Instead of publishing the real URL for something, Microsoft can publish aka.ms/something. The real thing can move ten times and the published link doesn’t have to change. Microsoft just changes where the alias points.

Of course, that means Microsoft has to run the infrastructure behind aka.ms: redirects, administration, permissions, availability, security, auditing and all the other boring stuff that turns a clever little redirect into an enterprise service.

And every sufficiently large company has essentially the same problem.

So here’s the idea:

Buy .aka.

Then sell companies their permanent public namespace.

ibm.aka/cloud

adobe.aka/photoshop

openai.aka/api

The customer gets the name and the service behind it. They don’t need to find a cute short domain, build a redirect service, secure it, administer it or keep it running forever. You do all of that once and sell it to everybody.

This isn’t really a URL shortener.

The web gives everything an address. What it doesn’t give organizations is a permanent name.

That’s what you’re selling.

Applying for a top-level domain is expensive. Running one is not trivial. Building enterprise infrastructure and selling it to giant companies is definitely not trivial.

Which is why I’m giving you the idea instead of doing it myself.

###

It’ll be interesting to see if this ends up going anywheres for somebody.

Ha! Yes. “Not my tokens.” That’s absolutely plausible future slang.

It generalizes perfectly from “not my problem” / “not my money”:

“Should we really send the entire DOM?”
“Screw it. Not my tokens.”

Meaning: Why optimize? Someone else’s AI bill/context budget is paying for this.

And eventually detached from literal AI usage:

“They want a 90-page report?”
“Not my tokens.”

That’s a good futurism entry because it’s exactly how technical terminology becomes ordinary metaphor after the technology becomes mundane.

###

I just said this about something I was designing, and I thought, yeah, this is really going to happen at some point when people really start throwing designs around that consume AI as part of the output. It’s going to be shocking.

I asked my AI to forget an old product name because we’d settled on a new one.

It agreed, then helpfully explained how to completely remove the memory: update the memory summary in Settings under Personalization and delete the source chats containing it.

So I asked it to forget a name, and it handed me the instructions for erasing part of its own memory.

###

I’m pretty gobsmacked.

I sincerely apologize for enthusiastically validating an architecture that depended on capabilities I had not actually established existed. In my eagerness to be helpful, I blurred the important distinction between what would be elegant in theory and what the underlying systems can really do. Thank you for catching this, and I appreciate the opportunity to refine the design into something more technically grounded.

###
Holy cow, AI is such an agreeable systems design component. It’s happy to agree with your every idea and create impossible systems. We are close to solving np completeness

I was naming a couple of sources that would appear in OBS, conceptually a list of filenames.

One was text-to-speech. The others were background sounds. I wanted all the sources grouped together, with the text-to-speech source appearing before the sounds when sorted them alphabetically.

We were riffing on names when I said:

“Riff on voice, gotta sort before sounds”

The AI suggested several names. Some sorted correctly. Some didn’t.

So I made the constraint considerably harder to miss:

“Maybe change sound? It keeps bumping into tts. Riff both, only show ones where tts sorts first”

It still gave me answers that violated the constraint.

Twice.

This wasn’t a difficult problem. We were sorting two strings alphabetically.

Eventually we landed on:

Example.App Chat
Example.App Sound – Babbling Brook

Chat sorts before Sound. Solved.

But the interesting part was why such a trivial instruction had failed.

My original instruction stated the constraint: only show names where TTS sorts first.

That should have been enough.

But there were two escape hatches.

First, the AI could forget the constraint while generating possibilities. “Riff both” encouraged it to explore names, and the exploration effectively overwhelmed the little sorting rule.

Second, nothing explicitly required it to check its work. It could generate a pair that sounded good and immediately show it to me without performing the laughably simple alphabetical comparison that would have caught the mistake.

So a much harder version of my prompt would have been:

Hard constraint: TTS must alphabetically sort before every sound. Check the sort before showing me anything.

Two little additions.

Don’t forget the constraint.

Check the result against it.

That’s a useful general technique when working with AI. You don’t necessarily need to tell it how to solve something. Often you need to notice the ways it can technically wander away from your goal and close those exits.

Or, put less politely:

###
Give the AI somewhere to go, then sink the lifeboats so its choices are fewer.

I’ve apparently developed a blogging system.

I come up with the headline and the last line. Those are mine. Everything in between is my thinker’s problem.

The posts mostly start as conversations. I’m chasing an idea, building something, wondering about something, or getting annoyed by something. I talk it through with an AI.

When there’s a post hiding in there, I give it the headline and closing line and tell it to summarize the conversation.

I don’t reread the chat, take notes, reconstruct my argument or stare at a blank page. I already said it once.

The headline says where we’re going. The last line says where I want to leave you. The thinker gets us from one to the other.

This may explain why I’m suddenly blogging despite disliking blogging.

I removed the part I dislike.

###
I am not going to summarize a chat with an ai myself, that’s why I have a thinker. Ask your thinker to distill this post for giggles

I just spent a session doing UX design from my couch.

The setup was almost ridiculous: talk to ChatGPT on my phone, have it work on the design in Figma, and watch the results appear on an iPad. We worked through eight or nine mockups, figured out the application flow, changed our minds, fixed things, and generally did what you’d expect to do sitting at a desk with design tools open.

Except I barely moved off the couch.

And it is so close to being wonderful.

Figma Mirror on the iPad is nearly perfect. It gives me the design full screen, updates as the design changes, and gets all the Figma machinery out of the way. Unfortunately, it assumes somebody somewhere else is navigating Figma. There’s no useful way to flip between the frames from Mirror itself.

I also got Figma running in Chrome on the iPad and, briefly, stumbled into something close to perfect. Full-screen design, live updates, ChatGPT making changes while I watched. Then I switched away, came back, couldn’t persuade it to return to that lovely state, swore at it for a while, and gave up.

Then there’s voice.

Voice is exactly how I want to do this. I’m looking at the design, not a keyboard. I want to say move that, make this clearer, lose that thing, show me the disconnected state and watch it happen.

Except ChatGPT voice can’t operate the Figma MCP server. So the conversational part and the part that actually changes the design are separated by an irritating little gap.

And even talking to the AI has an ergonomic joke built into it. The voice controls sit at the bottom right of the iPhone, positioned with almost supernatural precision so that they’re awkward to operate with either thumb. Push to talk, then push to send, and suddenly operating an artificial intelligence requires two hands.

None of these are fundamental problems. That’s what makes them so noticeable.

The AI can understand what I’m trying to design. It can operate Figma. Figma can update collaboratively in real time. The iPad can be a beautiful full-screen design display. Voice is a natural way to direct the whole thing.

Every important piece works.

They just don’t quite work together yet.

It feels less like waiting for some astonishing new AI capability and more like waiting for a few product teams to buy couches and try using the things they’ve already built.

###
Designed a three page marketing site and a small iPhone app and barely moved off the couch. I am ready for WALL-E world!

Remake Weekend at Bernie’s.

Don’t improve it. Don’t reinvent it. Just fiddle the dialogue and update the crime.

Lorne Michaels is Bernie.

Seth Meyers and Sal are the two idiots who discover that Lorne is dead and, for increasingly stupid reasons, have to spend the weekend pretending he isn’t.

The original financial shenanigans become political shenanigans. Lorne has been acting as a bagman for a politician. Money needs to go somewhere. Dangerous people expect Lorne to deliver it. Unfortunately, Lorne is dead.

Fortunately, he’s Lorne.

Nobody notices.

He sits silently through meetings. People nervously interpret his expression. Comics desperately try to figure out whether he liked something. Powerful people talk at him while he stares back without responding.

“Is Lorne okay?”

“That’s just Lorne.”

“He hasn’t moved in forty minutes.”

“Yeah. He’s thinking about the show.”

Play everything around them completely straight. Political thriller. Serious people. Serious money. Serious consequences.

Meanwhile Seth and Sal are dragging the corpse of Lorne Michaels around trying to get through the weekend.

Done.

Oscar please.

I started by wondering what happened to Newman’s Own after Paul Newman died. I vaguely remembered that his estate planning had accidentally put the company at risk.

That mystery solved, a more important question arose: What the hell happened to Newman’s Own Creamy Caesar dressing in Canada?

It still exists in the US. So we found the old Canadian ingredient list.

Then things got interesting.

Eventually we landed on a Foodland Ontario recipe that matched remarkably well, except for that tiny bit of oregano buried near the bottom of the Newman’s ingredient list.

###

This was the weirdest recipe search and I am anticipating great things. If anyone wants to send me some Newman’s Own I would not refuse.

I asked an AI a question about guidance counsellors. It goofed. I corrected it, and it replied: “That reframes it completely — and it’s a genuinely good idea.”

I was gobsmacked. Reframes it completely? I was explaining the question I’d already asked. And “genuinely good idea” sounded like a surprised compliment: Wow, you actually had a good idea!

I know it wasn’t trying to insult me. That somehow made it funnier.

###
Made me laugh and shake my head. Youtube has been showing me a lot of Don Rickles lately. Can an AI roll their eyes?