AI

AI Setup Evolution

by

I’ve been writing a lot about AI and tech projects lately, which makes sense since that’s the stuff that ends up on this blog. It isn’t all that’s going on in my life, just to be clear. But my AI setup has quietly evolved in August and September in a way that’s letting me undo a bunch of stuff I only built because of the cloud AI’s limitations, so here we go again.

First, the accounts. I’ve been using Claude a lot more than any other AI system lately, but I’m still mixed between Claude and ChatGPT for coding work. Instead of paying for Claude Max or the equivalent top tier ChatGPT plan, which run about five times the price of the regular subscription, I pay for two mid tier plans instead, the Claude Pro and the ChatGPT Plus. Partly that’s so when I run out of tokens on one, I can switch to the other. But that’s a small reason. The bigger ones are that if one gives me a bad answer, I want to try the other, and sometimes I want a second opinion and will ask both the same question. I also don’t think I’d actually use five times the usage, so paying for the bigger plan on either one would just mean I’m not getting my money’s worth. There are times I max out both in the same stretch, so maybe I am doubling up after all. Who knows.

Interesting side note, I’ve been leaning on Codex more for coding lately, with a bit of chat here and there. I like their interactive audio, though I don’t use it as much as I probably hold onto the idea of using it. I’m building sites and things with Codex now, not because I think it’s better, I started because a friend recommended it after I hit a specific wall with Claude. I still use Claude for coding too, but I keep them separate on purpose. Coding work tends to run a project’s token budget down fast, and I don’t want that eating into the day to day chatting I do in Claude. The coding side is pretty interchangeable between the two since everything gets saved locally to wherever I’m working anyway.

Now, the actual problem I’ve been solving. With ChatGPT, and then Claude once I moved to it, the challenge was always context. Take my car tracking prompts as an example. It’s one of many, holiday planning is another, so is tracking the general inventory of stuff I own, but let’s stick with the car since we’re actually replacing it with another used one right now. I’d feed it a lot of detail about the car, our usage, our spending, all of that. Then every new chat, I wanted it to already know the context, how to respond to me for that narrow kind of conversation. So I had some generic prompts, plus a very specific auto inventory prompt, and I’d load all of that into a fresh chat every time.

Then there’s the data itself. Registration, MOT dates, how much we paid for the car, fuel spend over the last couple of months. All stuff AI is genuinely good at trending and drawing insight from, and useful for research too. But I needed to save it somewhere. My old process was to have the prompt itself describe how to save the data, which in most cases meant getting the AI to structure everything into a JSON file I could download and keep a version of.

The problem is I don’t want to keep an indefinite chat going, especially with Claude, though ChatGPT has some of the same issue. As a chat gets long, the AI starts forgetting the earlier stuff and compacts it, and it even warns you this is happening. I wanted everything solid enough that if I deleted that chat and started fresh, the new one would already have everything the old one had.

So the prompts were only half the system. The other half was the JSON, or a CSV for a pure inventory file, sitting there holding the actual data. I built one of these little systems for the car, one for holiday planning, one for holiday budgeting, a stack of them, each with its own context switch. It worked, but it meant a real proliferation of chats. Start a chat, load the data, double check which version of the file I was working from, and when I was done for the day, have the AI spit out a fresh JSON with a version number in the filename and save it to disk. I ended up with thirty or forty versions of some of these files sitting around. Not done with the topic forever, just done with it for that session, so I’d save and move on, and next time load the latest one and keep going.

That whole dance existed because neither Claude nor ChatGPT could write to a file in its own project folder, only read from one. I could load something into a project, but I couldn’t have it write back out to that same spot. That was the actual limitation driving all of it.

Then, much more recently, within the last month or so, Claude’s Cowork got a lot easier to use and a lot more functional. It’s mid September as I’m writing this, and just yesterday I read that they’d simplified the interactions with it even further, which is great timing for me. With Cowork, I’ve gone back through each of my little systems and rewritten the prompts. Now I can just tell a new chat to look at a directory on my computer, read the markdown file with the instructions, read the JSON, and write straight back to that same JSON. No more versioning gymnastics on my end, because I don’t need to care about it anymore. Time Machine on my Mac is already snapshotting everything hourly and backing it up elsewhere, so the versioning is handled at the file system level, which is where it belonged all along.

I still keep these as separate projects in Claude, because I want specific context tied to specific folders on my computer. My car information lives under an auto folder inside a broader life inventory folder, my computers have their own folder, my domain names have theirs, each with its own markdown for how I want it handled. In the past I’d have to manually load the data file and the instructions file into Claude every single time I started up. Now the project just has a tiny instruction pointing to a markdown router, a file in the root directory that lists out all the different contexts I’ve got and where to find the detailed files for each one. It knows to go look in those folders, and inside something like auto, it’ll find an instructions markdown, the data file, and whatever else belongs there. I can start one omni chat for anything life inventory related, or a specific car chat, and either way it’s not my job to keep the local files in sync anymore, because that’s just handled.

For a while I could already do a version of this on my own computer in a coding session, but not from the cloud, not from my phone on the couch, or my iPad while we’re figuring out where to go for dinner or nailing down holiday plans. Now with Cowork, I can start a session that runs as long as it needs to. I don’t care if it compacts and forgets things partway through, because all the data that matters is already saved to a file, and it’s got a remote session into my actual computer at home.

In another post I mentioned why I might end up getting a Mac Mini instead of relying on my MacBook Air, and this is exactly why. If I take the laptop with me, none of these sessions work, since there’s no computer at home to remote into. I’m fine with a dedicated home machine for this, because lately the only reason I’ve been taking a laptop anywhere has been to do AI stuff with it in the first place. It’s a bit circular. If I can run everything through a remote Cowork session back to a computer at home, I don’t need to bring a computer anywhere at all.

I’ve been going on holiday without a computer for a while now, aside from that one AI use case, to the point where I’ve packed the laptop a couple of times telling myself I might need it, and then never actually taken it out of the bag. My 11 inch iPad Pro has become my go to screen for everything, and this new Cowork setup makes that even more true than it already was.

So that’s where things have migrated to for now, and we’ll see how it keeps evolving. Ideally I’d love to get to a point where I don’t need a home computer in the loop at all and can just use remote storage instead. But I don’t use a mainstream service like Dropbox, OneDrive, or Google Drive, which is probably where that kind of direct read and write access would actually get supported first. Last I checked, Google at least would let you read from those places but not write back to them. When that changes, it’ll be worth revisiting, but for now I’m sticking with my own storage.

That’s the state of the AI setup as it stands. Check back in a couple of months to see how much of this I’ve already torn apart and rebuilt again.

Teaching Future Scott to Trust the Money Bot

by

Simplifying my finances is one of those things that’s been on the list forever and just never got looked at. It sat there, neglected, while everything else got attention instead.

Part of the reason is that I spread certain banking functions across different banks. If you follow personal finance blogs, you already know that’s not the recommended move. I try to follow the You Need A Budget methodology, which is the app I use to manage my budget, and the whole philosophy there is keep things as simple as possible. One account, one clear picture.

But my problem is I don’t want all of my money for the month just sitting in a current account. I don’t want that to be where irregular spending, or spending in general, comes out of. Some banks solve this with what people call pots or envelopes, virtual ways to segregate money inside one account. This bill comes from here, that bill comes from there, savings live somewhere else, all under one umbrella. I’m not going to say which bank I use for this, but not many banks in the US or the UK actually offer it, so if you go looking you’ll probably figure it out. The bank I use also has a developer API, which is both potentially dangerous and potentially powerful.

What I wanted was something that could handle the bills coming out of one pot (mortgage, electric, the recurring stuff), set aside money for the credit card bill so it’s always covered when it lands, round up spending a bit each week and stash the extra somewhere, and keep the actual monthly savings separate from all of that. Then whatever’s left sits in the current account for general spending.

The catch is that thing needs constant care and feeding. Bills are fixed, so you pad them a little and move the money over. Paycheck comes in, and you have to decide how much sits where, because I want money sitting somewhere earning interest until I actually need to spend it. It’s a lot of small administrative decisions that need to happen on a schedule, and none of it is hard, it’s just tedious.

So naturally, when you’re a tech person with an API, and a private server sitting at home that you’ve locked down tight, my answer was to build something.

I got access to the developer API and dictated all of it into my notes app first, walking through every use case. Then I fed that into Codex, since that’s what I’m building with these days, and told it what I wanted. It came back and said, yes, absolutely doable, confirming the one thing I already knew I’d need. Everything had to live in one bank, using pots instead of separate current accounts, because the API can only do things within that bank itself, not across open banking connections. That part wasn’t news, it’s why I’d spent the past month or so slowly relocating everything into one place beforehand, betting the API work would do what I thought it could. I was playing the long game on it.

I know at least one or two friends who’ve played with this same API before and had good things to say about it. Someone else at work has built something much bigger with it too, which I’ll get into another time once I’ve actually built my own version of that.

What I ended up with is a website with a UI that tracks the different use cases and, more importantly, the logic rules, because there have to be very specific rules in place so it never moves money it doesn’t have or shouldn’t touch. I built it with a few maturity levels. At the bottom, it’s purely informative, just telling me what’s going on and what it suggests. Above that, semi-manual, where it recommends a move and I go execute the transfer myself. That’s roughly where I am right now, moving from just getting notified to getting a recommendation I still have to act on.

After that, the plan is a warning stage, a day or two heads up before it makes a move on its own, and eventually full automation, where it just tells me after the fact that it moved money, on top of whatever notification I’d get from the bank itself anyway.

The whole point is that if I’m not around to do it, my wife doesn’t have to think about it, and it saves me from manually shuffling money around every week or month. The only thing I’ll still need to do is reconcile it in You Need A Budget, since it’ll pull the transactions in through open banking like anything else. Because the moves are predictable, I’ll know what I’m looking at when they show up.

The logic itself is actually pretty simple. It’s the fact that it touches money that makes me cautious, which is why there are four or five approval levels instead of just flipping it on. The plan is to ratchet it up gradually over a few months, and if it’s consistently accurate after five or six months, I’ll trust it with full automation.

That’s where the AI part gets more interesting, past the point of just building the thing. Once it’s built, I don’t really need AI for it anymore beyond patching bugs or tweaking the UI. The real question becomes whether what I built for manual human intervention is actually the smart way to do it. Is the amount of headroom I gave the bills pot right? Are the intervals right? What actually makes sense here? I’ve asked it to help me look back at a couple months of data eventually and suggest improvements, which feels like the more useful use of AI in this than the building itself.

There’s a part of my brain that’s irrationally nervous about handing this over, even though it can’t make trades and can’t send money anywhere outside my own accounts, and I’ll see everything it does. It’s managed risk, I’d say. But if it works the way I think it will, it saves real time every month, so it’s worth the caution.

Next up, I want to see if there’s a US equivalent I could set up something similar with. My early research says no, but we’ll see.

Censoring My Own News, And Couldn’t Be Happier

by

Following on from my post about doom scrolling through the news and feeling the way I do right after eating that second or third piece of cake M told me I shouldn’t have, I finally focused on trying to actually fix my news problem instead of just avoiding it. I asked ChatGPT the other day whether it could build me a rule structure for news that would, in essence, censor things I read. Things I don’t want, things I do want. Turns out it could.

I consume most of my news through RSS feeds in an app called Reeder. My first attempt was simple. I exported my OPML file, the list of everything I subscribe to, handed it to ChatGPT, told it exactly what I wanted and didn’t want to see, and let it build the logic. I ran it once or twice. It took ten to fifteen minutes to process each time, was very compute intensive, and just wasn’t doing what I wanted.

So I had a different thought. I already had this RSS feed file. Could I get AI to build me an actual tool to do the heavy lifting, with the AI only handling a smaller slice of the logic on top? I’ve had good luck with that pattern before, where the AI builds a local tool for the grunt work and only steps in for the parts that actually need it. Turns out yes, it could, and of course once I proposed it the AI agreed it was a much better approach than my first idea. Which, seriously, if it was that much better, why didn’t it just suggest that in the first place and save me the time? That’s a separate emotional issue I need to unpack in its own therapy session about AI use.

It suggested FreshRSS, an open source RSS reader I’d actually used before, back on an earlier version of my VPS. I’d stopped using it at the time because Reeder already syncs everything to my devices through iCloud, so self hosting the same feeds felt redundant. But I’m generally in favor of hosting my own stuff when I can, so we brought FreshRSS back with all my existing feeds, added another container alongside it to do the preprocessing and staging, and then a third layer that’s the actual AI step, pulling the processed data and turning it into a summary rather than a raw doom scrolling feed.

The first couple of summaries it gave me were pretty thin, partly because my source list was thin. For months I’d been whittling down my RSS feeds because Claude or ChatGPT kept telling me two sources were basically duplicating each other, so I’d drop one. Now that AI is doing the curation instead, it’s going the other direction. It’s telling me to add more sources so it can build a fuller, more complete picture on the topics I actually care about.

One example. I’m an American living in the UK, and I also work in financial services, so I want to track the FX rate between the US and UK. Instead of just pulling from general news sites, it started going straight to the Bank of England and the US Fed for specific announcements. Something I would never bother doing manually, but a computer doesn’t mind, and then it digests it down for me.

It’s been running for about a week now, and it’s genuinely doing what I asked. I find it pretty funny that if a news provider censored my news like this without asking, I’d be outraged. But since I’m the one asking for it, I couldn’t be happier. There are hints here and there of touchy topics, since some hinting is unavoidable given what I’m tracking, but nothing that pulls me in to go dig further. Which, again, is exactly the point.

I’ve since added a bunch of topics I’d forgotten about in the first round, and it’s giving me good paragraph length blurbs with source links if I actually want to go read more. It’s a real step change from doom scrolling and bookmarking articles for later that I never get back to. I’m not trying to be comprehensive here. I want a picture of what’s going on, fifteen to thirty minutes of reading, and then I stop and I’m not depressed about the state of the world. I know I’m building myself a bit of a bubble, but I’m trying to stay reasonably well rounded by excluding specific things rather than whole topics.

One thing it set up that I haven’t tested yet is a weekly digest of blogs and general interest reading, more of a how-to and hobby mix. Sort of a Sunday paper edition, which I thought was a nice touch, even though I’m not sure I actually want that on a Sunday. So that’ll need some tweaking. For now the plan is it generates once a week but I can also just run it on demand. Still early beta stage since I don’t have a few weeks of usage on it yet.

Behind the scenes, this is all running on a public VPS I keep separate from the one hosting my blog, set up specifically for stuff like this. It’s pretty locked down, but the AI agents can hit the API to pull the processed data out of the app when needed. Next up is probably tuning the sources further, and I’d like to get it to a point where it publishes the results to a private webpage or file I can load in a browser and use Safari’s read aloud option on, instead of me manually asking Siri to read sections. That’s not solved yet.

The other thing I didn’t fully appreciate until I started this project is how much of a time sink visiting news sites had become, the same way social media is. It truly is an addictive habit you don’t realize you have until you’re trying to stop it. I catch myself now with a free second reaching to check the BBC or CNN, and stopping myself, and I hadn’t really noticed how often I was doing that until I started noticing myself not doing it. My to do list has gotten better as a result, since I’m spending that time on things I’ve actually said I want to do instead. More on that another time.

AI For the Mundane Stuff…

by

I have been using AI for increasingly mundane things I always meant to do but never actually got around to. Most recently, a mystery Mac DMG file.

I had a Mac DMG file. About eight gigs. Password protected. I have no idea why I password protected it. I knew roughly what was in it, a big archive of installer packages going back years, but I could not get in because the password I thought it was was not working.

I have been saving installers since back when software did not just update itself over the internet and you actually needed to hold onto major versions. Over time the archive grew across different Mac architectures. PowerPC stuff, then x86 and Universal binaries, and now arm64. So there is a lot in there going back a long time.

I put Claude on the case. I gave it my best guesses at what the password might be, because I did not want it going through every possible combination until the end of time, and it set up a cracking tool locally on my MacBook Air and let it run with those hints. It got in.

So now I had all this data, plus a bunch of other old archives, and the problem was figuring out what was what. File dates only get you so far. Some of it was obvious, certain versions of Microsoft Office are a dead giveaway for the era and architecture. But a lot of it was not.

This matters because I have two PowerPC Macs I picked up vintage that I actually want to use, and I have been thinking about picking up an older Intel Mac for similar reasons. Half of it is just because it is cool to have. The other half is that occasionally I do want to run some old software. So getting the right installer on the right machine is not a trivial question.

A few prompts later, Claude had written a set of scripts that parsed through everything, sorted it by architecture, and organised it all into the right folders. A few back and forth sessions where it asked me questions and I confirmed things, and it had a pretty solid result.

Then I went a step further. I had it size and create read only DMG images for each architecture, correctly sized so the PowerPC stuff would fit on a DVD if I want to burn one, or on a USB key I can move to one of the older machines. Right file systems, right sizes, the lot. About 100 gigs of stuff in total so it took a while to churn through, but my actual involvement was maybe 20 to 30 minutes of back and forth.

I had wanted to do this for years. Total elapsed time was maybe three hours, almost all of it the computer doing the work.

Now it is all backed up, the images are ready, and whenever I get around to spinning up those old machines I can actually do it. Which might not be for a while. But at least now I can.

Collecting AI Models Like It’s a Hobby

by

It’s almost like I collect AI models and providers. Each one has its own use case, and I probably could consolidate. But in some situations it’s not even worth the effort to do it.

As of writing this in early May 2026, and I mention the date because I usually have a backlog of a couple of months before things go live, my main tool is Claude. Well worth the money and then some.

But in addition to that I’m paying for Venice.AI as my privacy AI. Very cheap per year, limited use case, but for what it does it’s worth every penny.

Then there’s Lumo from Proton. I’m about one week into an experiment of not using it to see if I actually miss it. The honest answer is probably not, because I’m now running local models on my new MacBook Air with 32 gigs of RAM, which I got specifically to be able to do this. I’ve been running Mistral 3 3B locally and early results suggest it might actually be more private and more capable than Lumo for what I need it for. So that £120 a year might be on its way out.

I also have Perplexity Pro, which I don’t actually pay for directly. It comes bundled with another subscription I have. It’s slightly better than the others for web search type queries so I use it for that. Would I pay for it separately? Probably not. But I have it so I use it.

And then there’s ChatGPT. I touched on this in the last post but I’m currently trialling whether running Claude Code and ChatGPT’s Codex side by side is actually cheaper than paying for the next Claude tier up. The overages on Claude Code have been real, and the maths might just work out. Still figuring that one out.

So yes, I have more models and services than I probably need. Lumo is the obvious cut. Beyond that I’m genuinely not sure. The ecosystem as a whole is delivering value right now, even if it’s a bit sprawling.

What I do know is that two years in, I’m using these tools in ways I couldn’t have imagined when I first paid for Copilot on a 30 day trial just to see what would happen. And the people I talk to who haven’t really dug in yet, I get it, it can feel overwhelming. But the gap between what you can do with these tools and what most people are doing with them is pretty wide right now. And that gap is only going to get wider.

Levelling Up to Claude Code

by

So I mentioned at the end of my last post about AI that switching to Claude unlocked a new level of what I was doing with it.

It started small. Someone at work mentioned they’d been using Claude Code to build actual apps, and they’re not a developer either. Just someone with ideas and enough curiosity to see what happens. That stuck with me.

I had a specific problem I wanted to solve. I have some VoIP numbers through my internet provider, actual UK mobile numbers I can receive texts on. Useful for giving out to LinkedIn contacts or anyone I don’t want having my main number. The problem was sending replies meant logging into a clunky website every time. Nobody wants to do that.

So I asked Claude about it. One thing led to another and it said it could help me build a web app for that. When I asked if we could run it in Docker it said of course. I already run Docker on a few systems at home so that made sense. It built me a container, I ran it on my Synology, got it working for one number, cloned it for a second. Then I thought, wait, can we just have one container with a dropdown to switch between numbers? Of course we can. And it did it.

That was the moment I thought, if it can build me an app with this level of input and time, what else can it really do?

So I set up Claude Code properly, gave it access to my Raspberry Pi setup, and had it take the SMS app further. It interrogated my existing config, made some enhancements, deployed everything from the command line. Straightforward, but impressive.

From there I got more ambitious. My blog had been running on YunoHost, which is a self-service VPS platform. Decent enough, but it’s always on an older version of Debian because the open source volunteers take what feels like a long time to update it, and if an app isn’t in their package store you’re out of luck. I’d always wanted to run my own properly configured stack but never wanted to deal with the time to care and feed it.

I had a spare VPS sitting around doing nothing. So I asked Claude, can we design and build my entire website on this thing. WordPress in Docker, proper backups, the works. It said yes.

First I designed the whole thing with Claude, got a proper design document together, then imported that into Claude Code and let it build. About $25 to $40 in API costs later I had a website. And not just a website. Automated daily backups following a daily, weekly, monthly cadence with pruning built in, all replicating to a second location with an immutable copy at the end. Backup infrastructure honestly better than some hosting providers I’ve heard about. I then migrated my site over to it, wiped the old setup, and migrated it back again just to prove I could restore it. All worked.

Then I got bold. I had the spare VPS now freed up, so I used it to build a set of personal tools I’d always wanted but never had the time to set up properly. An open source SSO system with passkey authentication for me and my wife. A Searx search instance sitting behind the SSO. Network monitoring. I didn’t even know I could set up my own SSO until Claude walked me through it. Some API costs and an afternoon later it was running.

Around the same time I set up a proper monitoring stack. My external VPS now watches my home internet connection and my main Raspberry Pi, and sends push notifications to my phone if anything goes down. Not email, because I’m not going to look at an email. An actual push notification. I also have a speed test running every hour on a gigabit ethernet connection straight into the router, and my internet provider, who I’ve never had a bad word to say about, is consistently delivering around 305 megabit on a 350 megabit plan. Seeing that graphed over time is genuinely satisfying.

All of that cost me some API charges, probably a couple of dollars worth, and £4 for a push notification app I now own outright. Not a subscription. Just mine.

Then I built a router.

I’d looked at this before. A router is basically just a computer with extra network cards, and I’d used open source router software in the past. But I’d always bought the manufacturer’s hardware because I didn’t want to deal with building and maintaining my own. This time I asked Claude to help me design and deploy it, and then help me maintain it going forward.

Of course It said yes. It can be a yes man/lady/person a lot.

I researched the hardware with Claude’s help and landed on a Protectli VP2430, a fanless little box with four 2.5 gigabit network ports and an Intel N150 processor. 16 gigs of RAM and a 256 gig SSD. Overkill for a router, which means it’ll last a long time. Then I designed the whole OpnSense configuration with Claude before touching any hardware.

Deployment was more painful than I expected. The biggest issue was needing the new router connected to my laptop while also needing live internet on the same laptop to use Claude to programme it. The address space decisions I’d made complicated things further. I spent six hours one evening and an entire Saturday on it before rolling back. Then I realised what was in the end the hurdle that had stopped me, made one design adjustment, tried again the following week, and had it running in about two hours. It’s been stable since.

I’m not a software developer. But I’ve spent years building and supporting data centres, call centres, and large scale applications, so I understand enough to know when something looks wrong and push back on it. I understand some scripting and basic fundamentals. What I can do is explain what I want clearly, spot when the output doesn’t smell right, and interrogate it until it does. That combination, it turns out, is enough to build some pretty serious stuff.

The things I’ve been able to do aren’t things I couldn’t have done before in theory. But the time to care and feed a self-hosted setup was always more than I was willing to put in. Now I have AI that helps me design it right, build it right, and fix it when something goes wrong. The barrier that used to stop me isn’t really there anymore.

So now I’m looking at what’s next. I have mockups for a couple of actual apps I want to build. Things I want for myself that don’t exist quite the way I want them. A colleague at work just builds whatever he thinks of now. I’m getting there.

The biggest current headache is cost. Claude Code API charges add up fast when you’re doing serious work, and last month I had enough overages that I’m now trialling whether running both Claude Code and ChatGPT’s Codex together is actually cheaper than paying for the next Claude tier up. Early signs are interesting.

And separately, with my new MacBook Air running 32 gigs of RAM, I’m finally in a position to run proper local models. I’ve started downloading and testing, and the early results suggest I might be able to replace some of what I’m using Lumo for with a local model that’s actually more private and possibly better. That’s an ongoing experiment.

It’s a lot. But honestly, talking to people now, whether friends, colleagues, or people in the industry, so many are either just scratching the surface or not looking at it at all. I was over a year late getting serious about this. I don’t think I’m late anymore.

Switching to Claude

by

I’ve been on this AI journey for a while now, and for most of it ChatGPT was my main tool. But in February 2026 that changed.

Someone whose opinion I trust recommended I give Claude a try after voicing my frustrations with ChatGPT. When I actually sat down to evaluate it I did something that in hindsight was pretty funny. I asked Claude directly why I should use Claude over ChatGPT.

It told me not to.

Based on what I told Claude I wanted, it said I’d probably get better results from ChatGPT. So naturally I didn’t trust that answer and kept pushing. As I interrogated it further it started explaining that it was slower and more thoughtful, and from its previous read of what I was looking for, it figured I wanted fast straight answers. And that’s when something clicked for me.

One of my biggest frustrations with ChatGPT was exactly that. I’d ask it something and it would just fire back an answer. Fast, confident, and often not what I actually asked for. Like if I asked for specific instructions on how to do something on an Apple product, it would give me generic steps that didn’t even exist in the actual interface. I’d have to stop it and say don’t give me fluff, give me the actual thing. Then it would have to go look it up and either admit it didn’t know or finally give me something useful.

What Claude was describing as a weakness, slower and more considered, was exactly what I wanted. So I signed up and started using both in parallel.

Early on I was genuinely impressed with what was coming out of Claude. I was using Sonnet, the middle tier model, and the difference in output quality was noticeable pretty quickly. The concern then was whether I’d end up paying for two services. I was already paying for Venice and Lumo on the privacy side and the last thing I wanted was more sprawl.

But it became clear fairly fast that Claude was where I wanted to be. Which meant I had to migrate everything I’d built in ChatGPT over the previous six months or so. Custom GPTs, saved prompts, all of it. I had to extract everything, make sure I had backups, build a little text based database of all my prompts, and systematically move it all across.

I got it done within a month and managed to avoid paying for both services at the same time for more than one month. Then I downgraded ChatGPT to the free tier and haven’t looked back.

From a reliability standpoint Claude is better. Not perfect, and everything I said in the last post about not trusting it still applies. But it’s a meaningful improvement. And honestly, switching to it unlocked a whole new level of what I started doing with AI. Which is what the next post is about.

Use AI Like It’s Lying To You, Because It Is

by

I’ve touched on this in passing across a few posts now, but it deserves its own space. Because as useful as AI has been for me, it is not sunshine and rainbows.

AI is a great tool. I genuinely believe that. But I also think the future gets pretty dystopian if people don’t use these tools with their eyes open. And right now, a lot of people aren’t.

One thing I read recently that stuck with me: it can do super advanced calculations but it can’t tell time right. That sounds like a joke but it isn’t. Some of the things I’ve asked it, it will be absolutely insistent it’s correct. You interrogate it because something smells off, and eventually it folds. Oh yeah, you’re right, I was wrong. I’ve had situations where I knew it was wrong, kept pushing, and it took a surprisingly long time before it admitted it.

So what I tell people, my kids, colleagues at work, is this. Use it. But do not trust it. Assume it’s going to lie to you. If you go in with that mindset and you scrutinise the output, it can be really good. But you have to be able to scrutinise it. That’s the part people skip.

That’s also the part that makes it genuinely dangerous in the wrong hands. I can ask it something about cybersecurity and I’ll know pretty quickly if the answer looks right or completely off. But I can’t ask it to do my taxes. I don’t know tax law. So if it tells me I can do something, I have no idea if it’s true. That’s a problem. And it’s why you see things like lawyers submitting court filings with citations that don’t exist because a judge caught them using AI and not checking the output. People just throw stuff in and take whatever comes out.

I ran into this myself about a year or so ago when I was doing some budget planning. Nothing super sensitive, just the savings pots I set aside for predictable expenses throughout the year so I’m not hit with an inconsistent spend later. I do it all in a spreadsheet and it gets involved. I figured let me see if AI can handle it.

It did it, and then it didn’t. The numbers were inconsistent. Flat out wrong in places. I tried for a few months and I just could not rely on it. So I stopped and went back to my spreadsheet.

More recently I tried again with a different model, and I’ll get into that in the next post. But it’s actually working now. Two months in and the output is consistently accurate. I’ve also gotten smarter about how I prompt it, asking it to show its work and export the data in a way I can verify. So it’s a combination of the models improving and me getting better at using them.

But the overall lesson hasn’t changed. The reading of the tea leaves is the hard part. Sometimes the output is exactly what you wanted and better than you could have done yourself. Sometimes it’s close but slightly off in a way that’s easy to miss. And sometimes it’s just wrong and completely confident about it.

The tool is getting better. That’s real. But so is the risk of people treating it like it’s infallible. It isn’t. Not even close.

What I Actually Started Using AI For

by

I’ve been writing about my AI journey for a few posts now, and I’ve talked a lot about which tools I use and how much I trust them. But I haven’t really gotten into what I’m actually doing with them day to day. That’s what this one is about.

Trip planning was one of the first things that clicked. When you’re searching for flights and hotels, every website limits what details you can give it. With AI I could be hyper specific. If I’m flying with just the kids it’s this configuration. If my wife is with us it’s that one. These are the types of hotels we like, these are the amenities we need. It could actually hold all of that and work with it.

It didn’t book anything, but it helped me build out exactly what we were looking for and where to go look for it.

Then it evolved. I used to use Trello boards to track trips day by day, reservations, what we were doing when. Now I do all of that inside the AI itself. It exports as a YAML file I can save and reload if I need to start a fresh chat. Since privacy matters to me I’ll delete a chat when I’m done with it, but the file means I don’t lose anything. I have a general preferences file I keep updating, and then a separate file for each trip. It works really well.

I did something similar for days out with the kids. I took a week off last August and it was just me and the girls. I built out a history file of things we’d done and liked, threw in our preferences, and used it to plan the week. On Tuesday we have this, I need to figure out Wednesday, here’s a restaurant nearby that fits. I’d already been doing some of that in a neglected Trello board, but this way it was all queryable and easy to update.

Date nights with my wife got the same treatment. I’d go through our preferences, confirm them with her, and then present her with suggestions. I’ll be honest, I wasn’t exactly hiding where the ideas came from. More like, hey, I told the magic box what we like and this is what it came up with. Sometimes it was completely off. Sometimes it was genuinely spot on.

More recently it’s been helping with meal planning for the kids. Mine are picky eaters, so it was a lot of back and forth on what might actually work. It helped me land on a couple of new meals and then built a schedule to track them. Basic stuff on the surface, but genuinely useful in practice.

One that I found unexpectedly cool was using it to wrangle my Trello data. I have boards for things to watch and books to read. When I tried to pull that data into an AI the file was massive. So I had the AI help me write some scripts to strip out everything I didn’t need, took a two meg file down to about 50K, and then I could actually work with it.

Which brings me to audiobook recommendations. I was a little cautious about feeding it my reading history since it builds up a fairly personal picture of you. But I write about what I read anyway, so it felt like a reasonable trade. I’d give it what I read, when I read it, what I thought of it, and ask for recommendations.

Hit or miss, honestly. Some suggestions were weird and when I pushed back on them it would just fold immediately. That should always give you pause. But when it got it right it was actually pretty useful, and interrogating the reasoning often got me somewhere interesting even when the initial answer was off.

Which is a good lead in to something I want to dig into a bit more. It does some genuinely impressive things, but it’s far from perfect, and that part matters too.

The Case for a Private AI

by

So when I started paying for ChatGPT, I’d hesitate before putting anything into it. I had to make conscious decisions about what I was okay sharing and what I wasn’t. In some cases it was easy. I don’t care about this, so fine. In others it was something I did care about, but the convenience won out and I’d bend my own rules a bit.

Come May or June 2024, I read about Venice.AI. It was intriguing because I wanted a private AI, and what these guys had built was designed from the ground up around privacy. Nothing stored, no logs kept. Yes, there’s still that moment in time where they’re processing your data, but they’re keeping nothing after that. Their entire business model is built on trust.

Are they 100% trustworthy? No. The only way to truly guarantee that is to run your own model. But they were offering something real, so I was intrigued.

The reason I hadn’t gone the local model route already was my hardware. I had an M3 MacBook Air with 16 gigs of RAM. I could download LM Studio and run stuff, but it was slow and clunky. Just not the experience I was looking for. I looked into cloud-hosted GPU options too, the kind of thing a friend had mentioned, but it was a lot of configuration and effort I just didn’t want to deal with. Funny enough, nowadays I could probably have Claude Code help me set that up in an afternoon. But I’m getting ahead of myself.

So when Venice came out with a pro plan at $49 a year as their introductory offer, which has since tripled, though as of my last renewal I was still grandfathered in, I figured for that price it was worth trying. It’s definitely more rudimentary than ChatGPT, but the privacy confidence is real. I’m still careful about what I put in it, but I’m more willing to share certain things there than I am with the public models.

They’ve since launched different models with different privacy levels, which is worth knowing. Some are fully private, some are anonymised but not fully private. You have to pay attention to which is which.

Fast forward to summer 2025. Proton, who I’ve been writing about for over ten years now, and at this point calling them just my email provider doesn’t really cover it, they do storage, VPN, and a bunch of other stuff, launched Lumo, their own privacy-focused LLM. I gave it a try.

The free version was pretty limited, so I added their paid tier for a few months while waiting for my main Proton plan to renew in December. The jury’s still a bit out on it. It did some things okay. From a pure trust perspective I probably trust it more than Venice just because I’ve been a paying Proton customer for a decade. But the way Venice has architected things, it’s actually more private. Lumo is more convenient though, and private enough for most of what I need.

One of the trade-offs with Venice’s full privacy mode is that nothing persists. No data moves between devices or browsers when you log in. Lumo does sync, but you’re trusting that it’s still zero-knowledge on their end.

I’ve actually been using Lumo recently for some things that are genuinely private, things I wouldn’t put near a public model. My logic is simple. I’ve been paying Proton for years to store sensitive documents privately. So why not use that same platform’s LLM to process those same documents? I’m not going to get into specifics here for obvious reasons, but it’s been useful.

The broader point is that I don’t always trust the public models, and honestly you shouldn’t either. But over time I’ve become more relaxed about certain things. It’s a constant cost-benefit calculation. The privacy models are getting better, and even the public ones will sometimes tell you that for certain tasks you don’t need high-level reasoning anyway, so a privacy model is probably fine.

The hard part now is knowing which model to reach for. Which is a whole other post.

As with the first post in this series I used AI to generate my banner image. I am not saying it’s good. I am just saying what I did.