Reading The Butcher's Masquerade

Last Read Gate of the Feral Gods

More Experiments with Claude

I continue to experiment with Claude, and like many others have said, the difference between the 4.6 models (particularly Opus) and the previous models doesn’t feel massive, but it just seems so much more competent.

I’ve mostly been building personal helper apps and automations:

• a personal bookmarking/read later solution that works on iOS/iPadOS/macOS
• a scraper that goes to the websites of all the local music venues and pulls down the upcoming concerts so I can quickly look at what is coming up and decide if I’m interested in checking anything out
• some games to entertain my kids that I know aren’t going to lead them into a morass of in-app purchases or down the algorithm rabbit hole

As I’ve built these over the last few months, there was a lot more hand-holding for the prompting. Claude would constantly overestimate its abilities, and break Xcode projects, or create features and tests that it insisted it were working, but neither the feature nor test actually worked. After a while, it’d just tell me I didn’t need that feature, and to move on.

With the 4.6 models (particularly Opus 4.6), it is worlds different. Far more frequently I can say what I want, and it’ll have a 90% working solution by the end of my session. A couple of iterations, and it’s basically done (like a marble race JS/HTML game I made for my kids was nearly perfect from the start).

With Opus 4.6, I pointed it at the work Claude had previously done, and asked for improvements, and it made what I would call legitimate improvements to simplicity, performance, and good idiomatic usage. Good enough that at the end of the process, I even had Claude create MCP servers for the bookmarking and concerts tools I’ve built, so that I can ask desktop Claude about shows at a venue or any articles I’ve bookmarked related to certain topics, and it can bring them back in context (and tie them into the context of other things I might be doing, like adding buying a concert ticket to my todo list, or incorporating articles into some broader research I’m working on).

Life changing? No. Worth $20/month? For me, probably. For everyone? I don’t think so yet. It is still so much more heavily weighted towards having some inkling on things you’d want to build and automate, which is not every computer user.

These are small, toy apps. Useful to me, but nothing I’ve built would I feel comfortable sharing to anyone but friends and family. Incorporating this into existing, commercial apps is going to be harder than I think a lot of folks believe, because as good as Claude (and Codex) have become, they still frequently make meaningfully bad decisions around things like security. For businesses that are in an industry with high compliance needs, you will need good processes to ensure you’re not leaking information that could cause you legal or reputational issues. For folks launching apps on Lovable (or the like), I think the models are so good, that they give too much confidence that they’ve buttoned everything up, but they’ll make mistakes (just like humans do), but without some developer to take a look and try to catch it before it burns you.

Trying Out Some More Claude Use Cases

I’ve been using Claude (generally via Claude Code, but also via the Desktop and iOS apps) pretty regularly, trying out different use cases. These are all on the Pro Plan because I don’t wanna spend $100/mo. (Or, really, I don’t think it would be worth it, in my current life.)

Coding, in my experience, feels like a pretty solved use case. Nearly anything I’ve tried I’ve been able to get Claude to succeed at. In some cases, with little input from me to get 80% of the way there. In other cases, there’s a lot more trial and error to get what I want, but it’s still pretty darn good. I could likely get that success rate up using some more tools/MCPs to help it see what it’s building (usually it’s UX stuff that’s the gotcha) more easily. Still, if this is your common use case, it’s going to go pretty well.

Bespoke data analysis has been good, but maybe not quite as good. A while back I took my Netflix ratings data (which I’d saved like 2000 ratings) and uploaded it to Letterboxd. I then exported that data plus more recent ratings from Letterboxd, and uploaded it as a CSV to Claude. I told it to analyze my movie tastes, and then I’ve asked it to make recommendations in certain areas on movies I haven’t seen yet. That’s been surprisingly successful. It’s recommended a few movies I’ve really enjoyed (that I’d not heard of, or had forgotten about), and the recommendations have been pretty reliable. This is probably limited only by Claude’s data set recency and lack of which streaming service a movie might be on (but it tries). I did a similar exercise by exporting my Apple Music/iTunes Library into Claude. This one was ok. I think the amount of data was probably a bit too overwhelming, took a lot more finagling to get it to fit the variety of limits, and then it started making some pretty mundane recommendations (“You like Jenny Lewis? Have you heard of Rilo Kiley?”) and I think it was just too much context to maintain. This is another place that having an MCP or the like directly into Apple Music probably would have worked more effectively to keep all that data from polluting the context window on every request. More on that in a bit.

In this same bucket, I have these lists of top songs of 2025 from some music sites I follow (my friend who posts at Midnight Snark, local music blogger If It’s Too Loud, and internet legend Said the Gramophone). Usually, I’ll find a time I want to check out new music, pull up a page I’ve saved, listen to some. Go back later, forget what I listened to, rinse repeat. Eventually, I give up and the posts stay in my saved archive for a few years.

What a great job for AI, right? Go to those sites, grab the songs, normalize and rank them, and create an Apple Music playlist for me. It has not gone quite as well.

First, Claude tried to access those sites. Of the 3, could only access 1 over the web to pull the data. It tried a bunch of random things (spawning a browser via a tool/MCP was one of them), but was being stymied and blocked. All the while, just burning through tokens and continually needing to compact the convo (losing the context of what it had already done). Then I said “cool, let me just throw these into PDFs for you, then you can read those”. I do that, and now I’m over the context window for the chat I’m in.

I switch to Claude Cowork, point it at the directory the PDFs are in, and it says “thanks, let’s do it”, and then fails ostensibly because it couldn’t install the tool to read PDFs. Except, that’s not the actual reason. When you dig in to that particular set of logs, it’s actually because I hit my limit 😬

We’ll see how this does when I pick it back up again later tonight, but this is where the UX for this process is still too complicated for most people. Managing context windows, managing tool failures, trying 3 or 4 different ways to massage things into the box to get the right outcome. It’s early, and many of these rough edges are going to get smoothed away, but I still think the more complex topics are trickier than something your average user would do.

I also wonder if people sometimes aren’t even realizing that what they asked the computer to do didn’t actually get done. Instead, they set it off and it said “cool” and people have been trained over decades to expect the computer will do the thing they said, and aren’t trained to assume the computer may just completely botch the job.

Still, the pace of advancement is pretty remarkable, and the barrier to solve personal data/technical challenges is basically non-existent now, if you’re persistent and can afford $20/mo.

This iPod Classic: Nearly 24 Years Old and Still Kicking

We’re prepping to have our basement finished (no small task in a small footprint home in an urban area), so I’ve been steadily chipping away at finding places to put things. Sometimes that means revisiting other things that are stored elsewhere in the house to see if it’s time to part ways to make room for whatever object has been squirreled away in the basement.

Opening up one of my electronics treasure bins, there, sitting on top, was this beauty:

That’s my iPod Classic 2nd gen, which I’m almost certain I bought back in 2002. Plug in a set of headphones, and I can listen to whatever music I was listening to back in 2002 (and at least up through 2006, as I found a podcast from May of 2006).

It still works; maybe the audio jack is a bit more finicky, but that’s about it. Scroll wheel scrolls, display works fine.¹

Amazingly, it weighs less than my current iPhone 17 Pro (though, sure, it doesn’t do 90% of what the phone does). I could drop this in my pocket and listen to some music today, same as I could nearly 24 years ago. There’s something about finding a device like this, when it’s still highly functioning, that is inspiring. It’s like when you find an old radio (or TV with rabbit ears) in someone’s garage or basement, you turn it on and the same music you can hear today streaming over the internet or coming digitally through your car stereo comes blasting out. There’s nothing amazing about that inherently, but our (the big our) ability to have built devices that withstand the decades of time, building them on standards and formats that also last through the decades is pretty awe inspiring.²

1. I do think I replaced the battery in it a while back, which is probably why it can hold a charge. ↩

2. Of course, I say this as the latest version of macOS no longer supports FireWire, which is what this iPod uses to sync music over. I’m probably going to spend a stupid amount of time trying to figure out a way to fix that (even if the simplest solution is to dual boot into an older version). This is where that Claude Code subscription comes in handy … ↩

All the Information You Need is on the Task

I’ve been watching a lot of (some might say too much) Taskmaster lately, and it has seeped into my thinking. It lead to a momentarily epiphany groggily waking up to make my kids breakfast.

Working with the Apple and Google App Stores is like competing on Taskmaster.

• You’re given a task (in our metaphor, successfully submitting an app to the App Store) that is seemingly straightforward, but, actually, fraught with complexity
• The complexity is buried in a task that is written in such a way as to ensure you’re not quite sure what the rules actually are, and enough latitude that the rules could be altered at any point in time
• When you have a question or clarification and reach out to the App Store overlords, just like the Taskmaster’s assistant Little Alex Horne, the response is “All the information is on the task”, ensuring you get no more clarity and are left to just flounder around hoping to figure it out
• With each subsequent attempt, you either get more furious at your lack of success, or slightly more unscrewed and unhinged as you try changing random bits hoping something gets you through the process
• Ultimately, you’re arbitrarily judged by the Taskmaster, and maybe whatever you tried this time was good enough to get you points and published to the store

I don’t think it’s a good thing that the App Store process is reasonably paralleled by a comedy game show whose objective is to drive the contestants crazy and make them look like fools, but it at least the idea that Greg Davies and Alex Horne are overseeing it makes the absurdity marginally more tolerable.

Apple Goes Third Party to Jumpstart Siri →

In December, Apple moved to consolidate its AI leadership under Federighi, completing a transition that had begun earlier in the year when responsibility for ‌Siri‌ was removed from the AI group and brought under Federighi’s software division. In January, Apple announced plans to use Google’s Gemini AI models to power future AI upgrades, including an improved version of ‌Siri‌. In Federighi’s view, integrating a third-party model would allow Apple to finally ship a revamped ‌Siri‌ later this year after controversially postponing the update in 2025.

However, the report also outlines internal concerns about the implications of placing AI under Federighi’s control. People who have worked closely with him described him as highly cost-conscious and skeptical of investments with uncertain returns. This approach stands in notable contrast to rivals such as OpenAI, Meta Platforms, and Google, who invest tens of billions of dollars in data centers, chips, and AI researchers.

Given how AI and LLMs are developing, I don’t think this is the wrong choice in the short-to-mid term, and maybe even in the long run. There’s not a world where we end up with four-or-five-basically-identical-for-general-use LLMs. Every few months, there’s a new leader for general use, and then everybody catches up. It’s a market rapidly heading to commoditization (though on a time scale likely impacted by the insane CPU and power costs). Spending billions of dollars to be at par (or worse) isn’t terribly prudent and also likely wouldn’t be underwritten by investors (meaning they’d be doing it out of cash).

Instead, Apple gets to use a good (or great) LLM without investing billions to build one. Given how easy it is to switch LLMs (even setting aside some of the privacy nits, which any provider would agree to for Apple to pay them a large sum of money each year), it can swap providers very easily. Apple can be aggressive on the M&A front when the market inevitably shakes out and some (many?) of these companies who are valued at 10 to 100 times what they are actually worth are forced to reckon with an unclear path forward when the money tap dries up.

The thing Apple needs to be focused on is how to get ahead on what comes next to consumers, which is probably how to run bespoke, targeted models (rather than the generalized LLMs) and how to get them running on devices with less power and memory than an army of servers. That’s a game Google is certain to be attacking, and one that the larger players in the market (like OpenAI and Anthropic) are not as likely to. Ending up being boxed out by your #1 frenemy when models go small is a much bigger risk.

New Year, New Look

It’s been a while (i.e. almost 8 years) since I updated the look and feel of this here site. I decided this would be a fun job to see what Claude Code could come up with. I’ve done a good bit of work with Claude Code on the command line, but this time I thought I’d try it via the desktop app.

Aside from a few hiccups around using a git worktree and how that made running a local preview a little tricky, it ended up being a pretty easy exercise. I’m a horrible designer, so while I’m sure this looks better, I’m also sure there are plenty of ways that it could look better.

Still, it’s nice to have a fresh coat of paint. It also forced me to update my setup so that I could actually preview my site, as I hadn’t locally installed the gems needed to run this since … (checks watch) … before I got an Apple Silicon mac. 😬

It's been a minute

So, turns out, I haven’t posted in like 5 years. I’ve had some interesting things going on recently, so I thought I’d dump them here for posterity. Some will be useful for me to find later, when inevitably I forget what I fixed to make Time Machine backups work reliably again. Some Most will be useful for no one.

• About 6 months ago, I left my job. I’d been at it for nearly 8 years (after doing my previous gig for nearly 12), and I was just out of gas. My nervous system was basically in perma-“fight or flight” mode, and it was impacting me in ways I noticed (i.e. I was sleeping even less than I normally do), and I’m sure impacting me in ways I wasn’t noticing. It was a pretty big life decision, which my family was incredibly supportive of, as it’s given us lots more time together.
• It’s also given me time to build a shed. I mean, who doesn’t want to build a shed?
• I’ve also used some of the time I’ve had to dig into some technologies a bit deeper than I could day-to-day at the job. I’ve been building toy apps in Claude Code, which have been great for me to learn from (particularly iOS development, which has not been something I’ve spent a lot of my own time on), but also to build out loads of automations for my computer to make my life easier. I think that’s been a pretty big unlock for me—taking some process that used to take lots of steps and loading it up as a Shortcut, or AppleScript, or some automation that now takes less steps. Claude (and, presumably, the other coding tools) are very very good at that.
• One of the things I was playing with was trying to see how far I could get replacing Docker with Apple’s native container. Some stuff worked out of the box, but some other stuff I used (like nearly anything related to Docker Compose) did not. But, it’s been fun to toy around with it to see how far these native containers have come.
• Interestingly, from that experimentation, I learned a hard lesson. Shortly after I started playing with Container, I started having massive problems with Time Machine finishing backups. This was happening both locally and to my NAS backup. It would get about midway through, and then time out and restart. After a while, I finally pumped the logs and some other ephemera into Claude, and Claude helped me pinpoint the issue: the container snapshots that get downloaded for Container to leverage are sparse volumes, and they’re basically all 512GB volumes. Time Machine sees those as 512GB and parses through them, even though they’re actually tiny (maybe just a few MB) on disk. Excluding the snapshot directory for Containers (or, really, the whole directory ~/Library/Application\ Support/com.apple.container/) fixed the backups back to their normal, working state.
• I’ve gotten super into the UK show Taskmaster(I know, late to the party) and have worked my way up to Series 11. It is really quite good.

4 Years of Solar

In February of 2017, we decided to get solar panels. Our city was running a program to give low interest rate loans to get solar. For us, we figured it would be both a good environmental choice, but also a good economic choice.

With the heat wave we’re experiencing, felt like a good time to see how the panels have paid off.

For our home, the install covered about half our roof (a tree covers the other half). The quoted price was roughly $22k for the parts and install. The installer laid out some assumptions and made the case for a 5 year payback.

Well, here we are in year 4:

• Over the first 4 years, we’ve gotten $4373.29 in SREC credits
• Rough justice, after looking at historic electric bills, I think we’ve saved $1000 per year. This is probably conservative, as it doesn’t account for reduced usage in the colder months. But, it makes the math easy. So, $4000 in savings over the first 4 years of our panels.
• We got a $2500 rebate for the purchase.
• And a $6225 Federal Tax credit¹.

That puts us at $17098.29, so roughly ~$4900 left to cover.

Over the rest of 2021 through February 2022 (which will bring us to 5 years), we’ll likely pick up:

• Another $700 in SREC credits².
• Another $1000 in electricity savings

This is all without accounting for potential increase in home value.

If my conservative electricity savings is off by, say $250/year, you more or less close the gap on a 5 year payout.

That’s promising for solar, in general. A 5-7 year payback makes an investment far more economical for most people³, and if you assume that there’s some moderate increase in home value, this makes a lot of sense for almost any home owner.

1. This is going to vary year over year, administration by administration, but also might be augmented by a state credit. ↩

2. The SREC market fluctuates, but this is likely to shrink in many geographies, as more alternative energy sources come online. ↩

3. This assumes the privilege of having the means to own a home. Obviously, not everyone has that opportunity. ↩

Interview with Simpsons' Great John Swartzwelder →

John Swartzwelder, writer of some of the greatest episodes of The Simpsons (like “Itchy & Scratchy & Marge”, “Homer at the Bat”, “Krusty Gets Kancelled”, “Itchy & Scratchy Land”, and many many many many more), which contained some of the greatest lines from the show, like:

“Y’ello? You’ll have to speak up. I’m wearing a towel.”

“We are now approaching our final destination, Itchy and Scratchy Land. The amusement park of the future where nothing can “possa-bly” go wrong. Uh, possibly go wrong. That’s the first thing that’s ever gone wrong.”

“My eyes! The goggles do nothing!”

If you like The Simpsons, those make you laugh. And you will like this interview with Swartzwelder, who responded to this interview with the same comedic wit that he wrote his episodes or his insanely funny detective novels.

John Hughes was another writer who was working in Chicago advertising at the time. He has been credited with the famous credit-card shaving test, for Edge. Did you know John?

John and I had a few mutual friends, so I knew who he was, but the only time I ever sat down with him was when he tried to hire me to work for him at Leo Burnett, one of the biggest, richest, and boringest—to me—advertising agencies in town. Charlie the Tuna, Tony the Tiger, that sort of thing. I almost took the job, because the money was good and the view was terrific, but I discovered I wouldn’t have an office of my own. I would have to work in a kind of horse-stall setup, in the middle of a huge open area full of similar horse stalls. See those tragic figures down there? One of them is going to be you.

Well, I’d always had my own office, so I said no. Later, when John was making a million dollars per second directing movies, it occurred to me that maybe I should have taken that job, after all. When he went to Hollywood, I could have hung onto his leg. Nothing wrong with horse stalls, when you think about it. Horses like them.

I agree. I’m looking through your window as I type this next question: What do you make of the compliment “Swartzweldian”?

I guess I understand what they’re driving at, and it all sounds very complimentary, and I thank everybody for that, but I can’t help thinking “Swartzweldian” is about the most awkward-sounding word in the English language. I mean, I thought “Oakleyesque” and “Vittiriffic” [after “Simpsons” writers Bill Oakley and Jon Vitti] were bad, but “Swartzweldian”!

So how would you describe your sense of humor, your comedic sensibility?

Swartzweldian.

Working Through the Backlog

Apropos of nothing, working from home has had a couple of interesting benefits. Well, that and my body’s adjustment to spring (thanks 5am wake up time!)

I find myself keeping up with podcasts more readily, and TV and movies put production (rightfully) on hold due to the pandemic. So, music exists again!

Thanks to my dogged usage of OmniFocus, it turns out that I’ve got stuff on my “to listen” list going back to 2016. It’s a little strange to think “oh, I remember hearing this song and wanting to listen more” and realizing that the hearing this song part was 3 or 4 years ago.

Turns out I missed a lot of good stuff in the past few years.

← Older Newer →