Walking the dog with Claude

Written with LLM assistance.
The interview format is genuine; the prose is lightly tidied from voice notes.

I had lunch with a pal yesterday, and we got onto the subject of why so much technical material is either accurate-but-impenetrable or polished-but-slightly-wrong. It’s a gap I think about a lot, partly because I make videos that try to land in the middle of it and don’t always succeed. The conversation stuck with me on the drive home.

This morning, whilst walking the dog, I tried something a bit silly: I asked Claude (dictating to the mobile app¹) to interview me about it. Voice-to-text on one end, an LLM playing journalist on the other, and a black lab setting the pace. What follows is that conversation, lightly tidied – paraphrased in places, verbatim in others. The LLM is the interviewer; I’m the subject.

The words are mine, the questions aren’t.

Claude: Let’s start wide. Your friend’s comment was about the gap between people who know things and people who can explain them. Where do you see that gap most clearly?

Matt: Almost everywhere, honestly. The people with the information may not be great presenters, and the great presenters don’t necessarily have the information. You end up with two failure modes: dry, accurate material that nobody wants to watch, or polished material that gets the details slightly wrong in ways that make experts twitch. Both are common; the overlap is rarer than I’d like.

Claude: Who do you think bridges it well?

Matt: Grady Hillhouse on Practical Engineering – he actually is a civil engineer, and it shows. Devin Stone, LegalEagle, is a working lawyer. And the Brady Haran shows – Numberphile, Computerphile – solve it from the other end: Brady and Sean Riley are well-prepared “straight man” interviewers, and the experts they bring in are carefully chosen. The format itself is the trick. I’m biased on Computerphile, mind, since I’ve done a few.

Claude: Let’s talk craft. You’ve mentioned the Usborne robots before. What are they, and why do they still matter to you?

Matt: There were these late-70s / early-80s Usborne books on how computers worked, and they used little cartoon robots to show information moving around inside the machine. As a kid that just clicked for me. The thing I find quietly remarkable is that I still use it. When I’m reasoning about out-of-order execution, register renaming, modern microarchitecture – somewhere underneath, there are still little robots shuffling things between pigeon-holes. I’m not consciously picturing them anymore, but the shape of the model is the same.

Claude: That sounds like a beginner’s model that you’ve never quite outgrown.

Matt: Maybe you’re right? The thing I keep coming back to is that what’s easy for me to reason about as an expert² is often exactly what’s easy for a beginner to grasp. The mental model and the explanation aren’t two things – they’re the same object. I think people assume good explanations are simplifications of the expert’s real model. Sometimes they’re just the expert’s real model.

Claude: You’ve talked about a “tell my mum” test. Where does that come from?

Matt: When I was a teenager with a BBC Micro, if I figured something out I’d run downstairs and tell my mum. She’d listen and either follow or politely pretend to. Somewhere along the way that became my actual yardstick: could I explain this to her? Simple enough to follow, but not dumbed down. True. And underwritten by the fact that I actually understand the thing properly, even if I don’t expose all of it. If I can’t clear that bar, I’m not ready.

Claude: And the prep that goes with it?

Matt: More than people probably realise. For the early Computerphiles I’d practise on paper, out loud, two or three times, even though the finished video looks off-the-cuff. The off-the-cuffness is the prep working.

Claude: You skipped a Computerphile recording last week. What happened?

Matt: Yeah, that’s been nagging at me. The topic was how computers tell the time – it’s fallout from my C++Now keynote on benchmarking⁵. I wanted to talk about rdtsc and rdtscp on x86, how raw counter ticks get turned into wall-clock time, and the out-of-order execution wrinkles that show up when you read the clock. Good material. I just couldn’t find a visual hook. Computerphile’s whole stick is hand-drawn, handwritten stuff on lined paper, and I hadn’t worked out how to draw this one. I’m now mulling writing a tiny x86 assembly snippet long-hand and using that to show out-of-order behaviour. But the honest reason I bailed is that I hadn’t cleared the mum test yet. I hadn’t reduced it cleanly enough.

Claude: Why do you do these at all? Be honest.

Matt: Two layers, I think. The first is that I’m a show-off. My mum’s side of the family are performers – singers and dancers, mostly – and I did school musicals, played instruments, all of that. It faded in my twenties and I’ve sort of rediscovered it through conference talks and the videos. I like being on stage. I’m not going to pretend otherwise.

Claude: And the second layer?

Matt: I was giving a guest lecture recently and made the showing-off joke to a friend afterwards. He said, “I don’t think that’s all of it.” He’s being kind, but he’s also right. The bigger thing is the moment someone gets it. “I know this stuff, and now so do you – isn’t that cool?” That penny-dropping moment is the actual fuel. My territory is microarchitecture, compilers, assembly – how computers really work, underneath. Advent of Compiler Optimisations 2025 was probably the most concentrated dose of all of this in one go.

Claude: You’ve been quieter on the video front lately. Why?

Matt: Because they cost. And I want to be careful, because the costs aren’t the same shape. Something like Advent of Compiler Optimisations is genuinely expensive in hours – between the blog posts, filming and editing, I reckon about ten hours per 5–10-minute video, and there were twenty-five of them. Computerphile is much lighter on time – an hour or two of prep once the idea’s ready. The harder cost there isn’t hours, it’s the conceptual distillation. I can’t take it to Sean until it clears the mum bar. Last week’s no-show is the example.

Claude: Final question, and it’s a deliberate provocation. Compiler Explorer takes the most opaque part of writing C++ – what the compiler actually does – and makes it inspectable in real time. Is that the same impulse as the videos and the talks, just made into a tool?

Matt: Do you know what? I’d never actually thought of it that way. But – yeah, I suppose there is a bit of that going on. There’s the immediacy I like: it compiles as you type, rather than you having to kick off some long process. And I try³ to keep pushing it to be more accessible, more instinctive. Put like that, there’s clearly overlap. The tool, the videos and the talks have something of the same flavour I suppose.

That last question was quite a surprise; I genuinely hadn’t thought of CE that way before.

I want to do more of these – more Computerphiles, more talks, more written-up material – and I keep not doing them as often as I’d like. The honest reason is the one above: sometimes the cost is hours, sometimes it’s getting the idea clean enough to be worth saying⁴. I’m trying to give myself permission for the output to be infrequent without it being stopped. Those are different things, and I keep conflating them.

And – is this the future? An LLM as an interviewer while you walk the dog? I’m not sure, but I will say: I came home with a thought I didn’t leave with, which is more than I can say for most dog walks. Monty remains unconvinced (though he’s now snoozing happily on the sofa).