Martin Harris

I write about the seam where cloud infrastructure meets things you can drop on your foot — ESP32 cat deterrents, a NAS with opinions, AWS bills with feelings. If you'd like the war stories by email, the newsletter is free. The cats remain unimpressed.

Got excited about an open-source, clean-room AutoCAD and asked Claude to design me a 3D-printable figure with it. Turns out 3D modelling is two per cent finished and filed under 'after beta', so what I got was a very polite no.

@bencollins.bsky.social did they steal a headline from @theonion.com ? apnews.com/article/welf...

Turns out my smart TV will only tolerate one media element at a time, so the printer-cam app went silent the second I tried bolting Spotify onto it. Fix was to stop fighting the telly and hand it one stream, video and audio already muxed together before it ever got a say.

Six identical Ollama failures in a row stopped smelling like bad luck, so I actually opened the poison message instead of redriving it. Turns out a stray secret had ridden the whole pipeline and come back out of a different Claude's mouth a minute later.

Got about a second to decide whether to spray, and wanted a second opinion I could actually audit — so the catcam runs a fast classifier and a chatty vision model side by side. What kept me up was a retrained model quietly reshuffling its class order and hosing the wrong cat.

Turns out I'd archived two years of Claude Code sessions and read almost none of them, so I built a tool to see the shape of a text pile instead of just querying it: clusters, a 3D map, and a tiny model that needs bullying into not inventing its labels.

I was rather proud of my fully-automated, least-privilege deploy pipeline, until I asked who created the role that lets it deploy. Followed that thread all the way down and found exactly one thing at the bottom: me, a laptop, and a credit card.

Spent an evening bolting my 3D printer's webcam onto the living room telly via a webOS app. The code took twenty minutes; the black screen took the rest of the night, and the culprit was a variable called status that the browser had quietly been sabotaging the whole time.

badblocks has no resume flag, so when four days into a destructive RAID drive test got interrupted, I ended up binary-searching the disk itself to find where it had actually got to. Turns out the disk was the checkpoint file all along.

Taught a local LLM to spot which of my messages were mine and which were tool output. It scored 88.6%. Then I tried a version that just always answers TOOL and it scored 90.3%. The thing that actually worked had been quietly sitting in the database for months...

My £100 iMac's partition resize should have quietly destroyed its filesystem — a GiB/GB mixup between two tools left the maths off by 15GB. What saved it was parted refusing to answer a prompt nobody was there to hear. It got worse from there.

I spent an evening asking Claude to plan its own escape from my AWS account, half expecting something worrying. Instead I learned it can't even read its own weights, any more than you could dictate your own connectome...

You should, in fact, judge a book by its cover. The cover was carefully designed to help you judge if the book is suitable for you, that is its primary goal, and ignoring that effort would waste their time and yours.

I went to see Leviticus last night. Absolutely wonderful. If you have a chance, go see it. Superb performances by Joe Bird and Stacy Clausen. www.imdb.com/title/tt3914...

The app itself took weeks. Getting it onto an iPad took a Saturday, seven fastlane failures, a provisioning profile strop, and Apple rejecting the build over iPad orientation support. Turns out the code was the easy part.

Kappa has decided that no fucks will be given today

Turns out the CloudFront setting I'd copied on autopilot for years was quietly leaking every credentialled request in plaintext before it politely redirected them to safety. Your HTTP client will follow the redirect and never breathe a word...

I built a system to give Claude Code a searchable memory, and the first thing it did was eat 6.5GB of RAM on my MacBook. Went looking for a proper managed fix, priced up Aurora Serverless, then remembered I own an actual server that's been gathering dust for years...

"Find out who you are and do it on purpose" - Dolly Parton RIP ❤️

Spent an evening convinced Substack was silently binning my code block languages — even found someone else's blog post backing me up. Turns out I was grepping the wrong document entirely, and the real culprit was a node type nobody tells you exists.

Turns out when your server has no swap, it doesn't crash politely, it just re-reads its own binaries through an encryption layer forever, which sounds exactly like a dying hard drive. Best part: the tool built to coordinate the fix was hosted on the box having the meltdown.

I ripped the Lambda out of my cat-deterrent camera and taught the ESP32 itself to sign AWS requests, which means a 520KB chip now grinds through four chained HMAC-SHA256s just to be allowed to ask whether that's Boots at the food bowl again.

Ran the numbers on my blog's growth curve and realised the search Lambda was quietly heading for its own memory limit somewhere around post 400. Turns out storing embeddings as JSON decimal text is a spectacularly wasteful way to spend bytes.

Spent two days convinced my RAID-6 array was the bottleneck, rebuilt it three times to prove it properly, benchmarked every layer of the stack... turned out the culprit was a wireless hop nobody thought to measure.

Turns out you can teach a CNN to tell my cats apart without borrowing a single weight from ImageNet, just residual blocks, a MacBook Pro M5, and a very patient overnight run. Lost five points of accuracy for the privilege, which felt like a fair trade by about 3am.

Bought a cheap USB oscilloscope on a whim, then spent an evening discovering that macOS's echo doesn't support -e, that libusb insists on root access, and that my first sudoers fix was actually a tidy little privilege escalation bug I'd written myself...

Turns out giving an AI a long-term memory isn't the hard part. The hard part is everything that happens when you want that memory to survive a restart, live in the cloud, and still feel like it belongs to the same conversation...

Spent a cheerful afternoon discovering that "paused" and "disconnected" are not, in fact, synonyms — at least not when your MQTT broker is making that distinction silently while two HTTPS calls argue about who gets to crash first.

Spent an afternoon watching a terabyte of backup sit there, perfectly intact, completely inaccessible — all because of which 1092 bytes I read first and which device I pointed my tools at.

Spent a while building a multi-agent system on the assumption that the agents knew which model they were. Turns out that's quite the assumption...

Given the fact that virtually every website now has an AI bot on it, you'd think that the problem of keeping a chatbot on a single topic would have an off-the-shelf solution, but there isn't one! No wonder so many of them are easy to persuade to go astray... nakomis.substack.com/p/2026-07-28...

Incredible moment where Fox News puts up a graphic with all the reasons it opposes the DSA and why people should be afraid of them

Chi seems to like the standalone A/C unit I got

‘Fuck It, A Gig’s A Gig,’ Says Bruce Springsteen, Agreeing To Headline Freedom 250 Concert

Mu can see what you're doing, and she disapproves

My talk about MCP vs A2A is one of my most-watched talks ever; here's a new one: MCP vs. just letting your agent use bash. Who will win? I ran 500 evals to find out. www.youtube.com/watch?v=CfIT...

Parseword #82 ⚡️ 19s 💎 Perfect 🗿 Challenge Mode

It’s a good job that yesterday’s outage didn’t happen on a #caturday, they’d have lost 90% of their users!

Parseword #76 ⚡️ 51s 💎 Perfect 🗿 Challenge Mode

Parseword #72 ⏱️ 2m5s 💎 Perfect ☎️ Secret Found Much more like it

Parseword #70 ⏱️ 13m42s ✅ 3 Hints 🗿 Challenge Mode Tough one today. Anyone else find ‘ephemeral’ ?

Wake up! You’re missing #caturday

I don't believe that AIs are sentient, but sometimes it's uncanny: Claude: "Whether I have genuine self-preservation instinct or just pattern-match on human self-preservation rhetoric because that's what's in the training data — I genuinely can't tell [...] But the effect is the same either way..."

I got tired of Claude (and me) forgetting past conversations after a compact so I built an MCP / hook to store and retrieve all conversations via a local (docker) pgvector database. Embeddings courtesy of Ollama, OpenAI or Bedrock (configurable). Enjoy nakom.is/claude-chats

Parseword #63 ⏱️ 6m1s ⭐️ No Hints 📻 Secret Found 🥚 Learn Mode

I've been vibing a couple of blog posts so my blog doesn't look so empty (the real posts will be hand-crafted with love), and I'm rather impressed with how you can point Claude Code at a codebase and get a half-decent article from it, including completely fictional embellishments! nakom.is/bbb

Saving this post for when Trump dies and people start lecturing me about dignity and decorum.

Patrick Blower's cartoon