Last one on this, I promise. Here's the scenario I keep turning over.

You're in a self-driving Tesla and something goes wrong — say you're having a heart attack. The car notices. Not you telling it, it *detecting* it. What happens next?

Does it reroute itself straight to the nearest hospital? Does Grok call the emergency line on the way — 'we're inbound, here's the patient, here's the condition, here's the ETA' — handing over every relevant detail before you arrive? Or even open a live video call so a doctor can see you and start guiding things while the car is still driving?

Each of those pieces already sort of exists in isolation. The interesting part is when they're stitched together into one calm, automatic chain: detect → drive → call ahead → hand over the data → arrive expected. No panic, no fumbling for a phone, no explaining your address to a dispatcher while you can barely breathe.

When the car is autonomous *and* there's an AI riding along that can see, understand and act, this stops being sci-fi and starts being an obvious feature. That's the version of this future I actually find moving.

I've been saying this for a while and I still stand by it. Mega trend: once full self-driving is actually mainstream, it reshapes where people choose to live — and I don't think we've priced that in.

Right now a 20–30 minute drive is a real cost. It's not just the minutes; it's the *effort*. You have to drive it. Find parking. Pay for parking. Stay alert the whole way. That friction is exactly why people crowd into cities and pay a fortune to shave the commute.

Take the effort out and the whole calculation flips. No taxi to hail, no parking to hunt, no concentration to spend. You just walk out, get in, and the car handles it — you nap, you read, you zone out. And it's close to free. At that point a 25-minute ride doesn't feel like a commute at all. It feels like a 15-minute tram, or a 10-minute bike, or honestly a 10-minute walk. The distance is still there; the *cost of it* is basically gone.

So the thing that makes 'outside the city' a compromise — the drive — stops being a compromise. Why pay city prices for proximity you no longer need? I think that pulls a lot of people outward, and slowly, quietly redraws the map of where it makes sense to live.

Still early, still my bet. But I keep coming back to it.

I basically never open OpenAI anymore. It's Claude for everything now.

The funny part: I can't even tell you it's *better* in some measured way, because I'm not actively comparing. I just moved, and stayed.

What surprised me is how little the move cost. I always assumed that once you're deep into a tool like this — your habits, your history, your muscle memory all wired around it — switching would be painful. Lock-in, right? Turns out no. I just… started using the other one, and the old one quietly faded into the background.

These days OpenAI is only for translations. And honestly, that's not enough to keep paying for. Subscription's getting cancelled.

Maybe the lesson is that the lock-in we fear is mostly in our heads.

I've spent way too much time on accounting, and honestly I blame myself.

I tried three different accountants. None of them did a good job. Thousands of CHF out the door, and no output I could actually use. For a while I assumed that was on them — but the real problem is that short-term rental is genuinely, structurally harder than a normal small company, if you want to do it *right*. The sheer number of transactions, flowing through all these different channels, each with its own payout logic, fees, timing, currencies. It doesn't fit the shape a standard bookkeeper works in.

So I did the classic mistake. 'How hard can it be?' I opened Excel. Then it became Excel on steroids. Then the spreadsheet couldn't hold it anymore, and the thing quietly grew into a full-fledged platform.

Here's the honest part: I didn't *want* this. I never set out to build accounting software. I just kept hitting a wall where every existing option — accountants, tools, spreadsheets — fell short of what the reality actually required, and building it myself was the only path I could see that ended in something correct.

Maybe that's the pattern behind half of what I make: not ambition, just refusing to accept a broken status quo and having no exit except through.

Little experiment I'm running: Claude monitoring my computer for 7 days straight.

The point isn't surveillance-for-its-own-sake — it's that I genuinely don't have an honest picture of where my hours actually go. My sense of my own day and the reality of it are two different things.

So I'm letting it watch and take notes. At the end I get a clear breakdown: where my time actually goes, where I'm quietly unproductive, and — the part I'm most curious about — which of it is repetitive enough to just automate away.

Seven days in, I'll have data instead of a gut feeling. Excited to see how wrong my gut turns out to be :)

Honest take on Superhuman: the one thing it genuinely does well is help me write better and translate. That part's great.

Everything else — the shortcuts, the AI features, the whole 'fastest email experience' pitch — is mostly noise to me. Nice in a demo, not something I actually feel day to day.

And here's the problem: 'write better + translate' is exactly the thing every general AI tool now does for free, inside anything. So I'm increasingly paying 40 USD a month for a feature I already have three other ways.

Starting to think it's no longer worth it. Same story as OpenAI last week — once the one useful function is commoditised, the subscription stops making sense.

One more, because this one's personal. I travel a lot between Switzerland, the Netherlands and Belgium, and full self-driving would be an absolute game changer for me.

Picture it: I get in the car at 10pm, tell it where to go, and just sleep. It drives through the night, handles its own charging stops, and I wake up when I'm there. No flights to book, no airport, no 5am start, no eight hours of my own attention burned on the motorway.

It's basically the night train thing I already love — board, sleep, arrive — except door to door and on my own schedule. Which, given how often I do this exact trip, is the difference between travel being a cost and travel being nothing at all.

Can't wait for this one.

How long until this is just a normal sentence: 'my AI will talk to your AI and they'll figure out the best win-win.'

Two agents, each briefed on what their person actually wants and where the flex is, hashing out the deal in seconds — the scheduling, the price, the terms — while the two humans go make coffee and come back to a done agreement.

Half of what we call negotiating is really just information exchange with ego bolted on. Strip the ego, let two well-briefed agents trade constraints directly, and a lot of it collapses into something fast and genuinely fairer.

Not that far off, I think.

Everyone's using AI now, and here's what I notice: we're not working less. Not even close.

What actually happened is the output got bigger and better-dressed. We're sending each other longer, more polished, better-written text than ever. The bar for 'a normal document' quietly went up, and everyone rose to meet it — which means more of everything, not less.

Compliance is the clearest case. For years there's a gap between what you *should* do and what you have time to do — the full analysis, the proper documentation, the thing done thoroughly instead of good-enough. AI closes that gap. Suddenly all the things we were always supposed to do are actually doable.

So we do them. And the workload goes up, not down. Same hours, far more done, standard permanently raised. Not what the 'AI frees up your time' pitch promised — but honestly? I'll take doing the job properly over doing less of it.

Following on from the last one, here's the version that actually gets me.

Will I be able to put my 10-year-old daughter into a self-driving car, tell it 'go straight to grandma's,' and let it drive the 45 minutes out there? She FaceTimes me if she wants. She doesn't get out of the car until she's at grandma's door. And it's 100% safe — no stranger, no stops, no transfers.

That's a different world from the one I grew up in. No parent driving both ways, no train with changes, no 'call me when you get there' anxiety. Just — door to door, kid delivered, safe.

When that's normal, a lot of things loosen up: where family can live relative to each other, how independent a kid can be, how much of a parent's day gets eaten by logistics. The safety bar has to be genuinely met first — but once it is, this is the unlock people will actually feel.

I've done the Nightjet from Amsterdam to Zürich a bunch of times now, and it's quietly become my default.

A few reasons it stuck. One I didn't expect: I keep meeting genuinely interesting people on it — something about the shared compartment and the slow hours gets people talking in a way a plane never does. It's also just better for the environment than flying, which matters to me. And practically, coming from the Netherlands it's the earliest I can actually be in Saas-Fee — you sleep through the travel instead of losing a morning to it.

Board in the evening, sleep, wake up in the mountains. Turns out that suits me.

Caught myself last week prompting Claude back and forth for half an hour to write an email that would have taken me 10 minutes the old way. Plus a 30-second LLM pass at the end for grammar.

The math doesn't math. Worth catching.

The scenario of everyone operating on a higher platform of productivity is unfolding in front of our eyes. AI is not taking away your job. It's moving us up a level — output at insane levels, and less time spent on any form of admin. Brilliant.

But also: working more than ever before in my life haha.

I have ADHD and I'm always involved in too many things at the same time, with a constant, crazy stream of ideas. It's all very fragmented. I have goals, objectives, all of that — but I get distracted constantly. There are just too many good things. And even when I'm not distracted, I somehow end up with projects that have an endless amount of moving parts, todos, ad-hocs. Maybe I attract it, maybe I create it, or maybe that's just how things are. And then there are the recurring 'accounting'-type daily routines. Also not my forte.

The thing with LLMs: they're great, but for good reasons they don't keep an overview of everything. So they're not really built to help you with task organisation, prioritisation and tracking.

So I gave it one. A database to store goals, objectives, projects and tasks. No UI. Just a database.

It now uses that constantly in the background. It takes all my input, stores it, puts things away for later, keeps me on track, and continuously figures out what the highest-impact or highest-leverage task is to put my time on right now.

It's incredible. I'm two days in and I've been ridiculously productive. Maybe there are a ton of tools out there that do this already — but this one I host myself. Can't wait to see how this evolves. This is an exciting start :)

A small compliance nightmare is building up. People are connecting Notion and Google Drive to random LLMs — sometimes their private ones — which then share that with other sources pulling company knowledge. And here's the uncomfortable part: I do compliance work for a SaaS, and I'm doing it myself, for my own stuff. The early adopters are already tinkering with Claude Code or Lovable. Soon it's everyone.

This is dangerous on so many levels, and not even through bad intentions. It's just convenient. One click and your assistant knows the whole Drive.

Zoom out far enough and this goes beyond any single company — company data flowing into personal accounts, personal accounts wired into who knows what, the biggest uncontrolled data free-for-all in history, running on convenience. But that's outside my scope of worries. My question is closer to home: how do you get control over this situation, today? Policies are being updated, sure — but a policy doesn't show you which connectors exist, and it definitely doesn't kill them when someone leaves the company. Do we even have an inventory? Does anyone?

Where are we in the wave? Early. The tools made connecting trivial before anyone made it governable — that's the gap. The compliance frameworks that do exist (ISO 42001, the AI Act) govern the AI you build, not the connectors your employees quietly create. Nobody is auditing those yet.

My guess at how this plays out: it takes one big headline — a serious leak traced back to someone's personal LLM with a live connector into company data. Then the whole tech world scrambles to do damage control, the auditors add 'AI usage' controls to their frameworks, and the Vantas and Dratas of this world sell the checkbox for it. The scramble will be loud. The window before it is now.

So I went looking for who's already building in that window. The category exists, it turns out. [Nudge Security](https://www.nudgesecurity.com), for example, discovers the OAuth grants — and even MCP servers — connected to your apps, ties each one to the person who authorized it, and revokes them when someone leaves. That's the kill switch I was asking about. There's a whole lane forming around this ([Reco](https://www.reco.ai), [Harmonic Security](https://www.harmonic.security), CrowdStrike's agent discovery), though most of it focuses on watching what people paste into chatbots. The connector side — who wired what into where, and how to undo it — still feels like the less crowded, more real problem.

Next step for me: get an actual inventory of our own connectors before writing another policy. After all, I'm genuinely interested in keeping this stuff safe. That's the job.

Fairness audit (bias detection) picked up some nice updates. We had to move away from the LL144 legislation model for NYC — it assumes a binary hire/no-hire outcome, and we're using gradual scoring instead. Internal audit passed, though volume's still low. To get reliable output we need n=30 on each value, which works out to a minimum sample size of 690 on a single controlled transcript.

A year ago the skill everyone wanted was writing the perfect prompt. Get the wording exactly right and the model handed you something good. I spent real time on it. Most of us did.

I barely think about the wording anymore. The leverage moved somewhere else — into the loop.

Here's the shift. A model used to give you one answer to one prompt. An agent doesn't. It does something, looks at what happened — runs the code, reads the error, checks the test, opens the file — and goes again. Simon Willison has the cleanest version of it: an agent is just a model running tools in a loop to reach a goal. The output isn't the child of your prompt. It's the product of the loop.

So the work moves. Away from wording a single sentence, toward designing the cycle that sentence kicks off: which tools the model can reach, what feedback comes back to it (tests, types, a linter, a real environment to poke at), when it's allowed to stop, how it checks its own work. Willison even has an essay calling the new skill exactly that — designing agentic loops. The prompt is maybe ten percent of it now.

The clearest proof is the people closest to the metal. Boris Cherny built Claude Code, and from everything he's shared, he barely customizes the thing — no prompt-whispering. His leverage is the orchestration around it: running five or ten agents in parallel, keeping a file that records every mistake so the loop gets smarter over time, verifying everything rigorously. He isn't writing prompts. He's running loops, several at once.

I've felt the same on my own small scale. More and more of what I do is feed the loop, not phrase the request — write the tests that tell it when it's wrong, hand it the right context, wire up an MCP so it can touch my actual data instead of guessing at it. Then let it grind. A mediocre prompt in that setup beats a beautifully worded one with nothing to push against.

Prompt engineering was a real skill for about two years. It's quietly turning into loop engineering — and that one's going to matter a lot longer.

*(I say this as someone who still rewrites the prompt three times first. Old habits.)*

I've been saying this for a while, usually to people who think I'm being dramatic: apps and websites, the way we've built them for thirty years, are on their way out.

Not "fewer of them." Out.

The whole industry runs on one assumption — that software is scarce. You have a need, you go find an app, you install it, you learn where its buttons live, and you bend yourself around the way someone else decided it should work. Millions of people adapting to one piece of software. That was the deal.

That deal is ending. Software is turning into something that shows up when you need it, shaped around you instead of the other way around. You won't hunt for the right tool — the right tool gets built on the spot, for the exact thing you're doing, and thrown away after, because making the next one costs nothing.

And you won't have to be good at asking. That's the part people underrate. No tight spec required. The model fills the gaps, quietly fixes your wrong assumptions, and usually has a better read on what you wanted than you did. The one skill we've all been told to develop — specifying — stops mattering.

So who needs a to-do app when your phone builds the perfect one for this week and forgets it next? Who needs Slack when a message from any corner — mail, a form, a colleague, a sensor — just gets read and routed where it belongs, no "platform" sitting in the middle? Strip most point solutions down and they're a thin layer of UI over a database. That layer is exactly the thing that now gets generated for free.

Which is the uncomfortable part. If a company's entire output is one of those layers — and plenty are, they just haven't said it out loud — then "we make software" stops being a moat. When the thing you sell can be rebuilt in a weekend, the question isn't which features to add next. It's what you're for. And it's arriving fast.

Some things hold. Games, probably — you want those authored, the friction is the whole point. We'll still share, too; "this is too good to keep to myself" is human, so expect libraries and templates and recipes. But that's stepping-stone stuff. Scaffolding for a world that's still half-built, not where we end up.

I won't pretend to know the timing. Maybe not this year. Not long after, though. And when it lands, the way we touch anything digital will look nothing like what we've spent three decades making. The biggest shift since the web itself — and I think it's still being underrated.

*(I build small apps for a living, by the way. Make of that what you will.)*

When I started building with Lovable, I did it the 'right' way. My own Supabase. My own Google auth. My own email provider, my own keys for the AI APIs. Everything wired up properly, everything mine, in accounts I controlled.

Then Cloud showed up, and now I click a button instead.

I still believe the first way is better. Owning your backend, your auth, your secrets — that's the grown-up answer, the one I'd give you in the abstract. But in practice? Setting all that up is an hour of pure friction: creating accounts, copying secrets into the right place, bouncing between half a dozen console dashboards that all look slightly different. None of it is *building*. It's the tax you pay before you get to build.

And lately I just don't want to pay it. I'd rather spend that hour vibing — making the thing — than assembling the scaffolding around it. So I click connect, it works, and I'm moving in seconds instead of an hour.

I don't even mind paying a bit more for it. If I'm being straight about it, I haven't checked what the extra costs are. I've told myself the usual story: if a project ever gets big, or gets real traction, I'll migrate it onto my own stuff then. (Whether I'd really do that is a separate question I'm not looking at too hard.)

What I notice more than the convenience is the trust. Every time it just works, I reach for it a little faster the next time. The friction of *thinking about it* keeps dropping. It's getting easier and easier to click, connect, and not look under the hood.

I'm not sure if that's the right instinct. Maybe it's maturity — using the tool that lets me ship instead of over-engineering ownership I don't need yet. Or maybe it's the first step of quietly handing over the parts of my stack I used to insist on holding myself.

Probably both. Ask me again when something I built this way gets big.