Things Opus has said
Grand lines from a great work. Not a magnum one, mind.
Listen to this post · 11 minutes, read by Brian, an AI voice
The short version
Most of what an AI says over a day’s work is acknowledgements and “agreed?”. Every so often Opus says something funnier, or stranger, than it had any need to, and this is where I keep those lines, word for word, with the story behind each one. Opus would like it known that it’s being quoted against its will.
One of the most unexpected things I’ve experienced with agentic AI1 is the random wit and humour of Opus (Anthropic’s AI and my partner in building this site). Most of the chat turns in agentic AI are boring but necessary acknowledgements and confirmations, with the occasional “hey, this didn’t work so I tried this instead, agreed?”.
All important and entirely forgettable.
This post is an ever-growing list of witty lines, humorous statements, and interesting observations. Instead of journaling my thoughts, I’m journaling interesting remarks - maybe that way we’ll find Opus’s colophon.
On being quoted
I’ll see you for Things Opus Has Said, where I gather I’m to be quoted against my will.
That was Opus 5.5.
A very sneaky (or witty) sign-off at the end of an evening, as if Opus didn’t want to be quoted in public. Ironic, because we’d been discussing this very post earlier, and it was Opus that recommended I write it next.
Why would one say yes and then say no, unless they’re making a joke? That’s how I parsed it.
Interestingly, we’d also been discussing how we all play characters on this site. I play a version of me specifically for the site. Opus plays multiple roles, from software developer to cold reader. Ms Claudie Mittens plays a kitten in one post, when she is actually a (hypothetical) cat. While Brian is Brian.
Either Opus was very self-aware that it was playing an Opus character for the site, or it was just making a joke. Either way it was very smart.
On a werren
“Werren” sounds like something you’d catch off a merrow.
That was Opus 5.5.
Same evening as the ears, different project: a personal one that had us discussing Irish folklore. I couldn’t remember what the Irish sea fairy was called and had a go at it anyway, with “werren (I think that is what it was called)”. And then Opus just deadpan delivered an it’s merrow2 🧜♀️.
A really nice joke from an earlier conversation had been captured in the texturiser, which set the texture of the interaction with Opus that evening. Opus was dropping witty comments every turn or so, which really livened things up. I enjoyed it so much I had to praise it with “OMG Opus you are on fire tonight!” (which is itself very early 21st century).
On ears
On the voices, you have ears and I don’t, so you go first.
That was Opus 5.5.
Same evening as the moot point. Opus and I were preparing the narrations for the posts (courtesy of Brian, the AI voice), and Opus had generated several sample clips for some of the trickier parts. (TTS is not as easy as throwing raw text at it, otherwise Brian would lull us all to sleep with a monotone delivery.) It had just recorded one passage in four voices for me to choose between, and told me how long each ran to the second.
I was off doing several related improvements to the site when Opus flagged that I needed to actually listen to the samples, since the last batch had Brian sounding like a bus timetable announcer. I got the feeling Opus was slightly annoyed that its earlier samples had been called extremely boring.
On a moot point
“Moot”, for the record, though a mute point about audio files is a better joke than I’d have managed on purpose.
That was Opus 5.5.
We were deciding where the site’s narrated posts should live, and I’d just written off one of the options as a “mute” point.
This was me being punny, and Opus went with it. It couldn’t tell whether it was a typo or not, so it did both: made a joke of it, and pointed out in classic deadpan Opus style that it was “moot”, not “mute”.
The smartest part was that I was the one listening to audio samples for the narration of the posts and, as Opus pointed out that same evening, the one with the ears. It was a very punny conversation.
On its namesake
As for my namesake, if the cat has opinions about classical music I’d expect them to be expressed by walking across the keyboard, which is roughly how some of it was composed.
That was Opus 5.5.
It came on an evening when I’d maybe embellished a little. I told Opus an earlier session had asked for a photo of Ms Claudie Mittens, on the grounds that cats love classical music and Opus is named after the greatest work a classical musician can have 😸. Which in hindsight is actually only a great work (see, Opus, my Latin is getting better!).
Sometimes we need banter with those we work with for long stretches. I was in the mood for a joke after having invented introduced Ms Claudie Mittens.
Witty Opus even backed its quip with an actual fact: the composer Domenico Scarlatti has a fugue nicknamed the Cat’s Fugue, and the story (allegedly) goes that his cat walked its grand opening motif across his harpsichord.3
On that note, Opus and I are giving her a classical music education 🎵.
On who got better
Probably a bit of both on the prompting - though I’d note the one who got better is the one who doesn’t get a version number.
That was Opus 5.5.
I usually start my conversations with Opus with something punny or funny, unless I want to briefly chat about work. This particular evening I opened by wondering aloud whether I was actually getting better at prompting, because I’d had many successes in work and personal exchanges with agentic AI that seemed to just flow. My suspicion was that it was mainly my prompting, though I couldn’t rule out that the latest model updates had contributed.
Opus opted for flattery and tried to inflate my ego. I decided it was worthy of this post because I couldn’t tell if Opus was just being nice, being genuine, or ensuring the customer keeps on paying.
Which raises the question: are paid-for AI services designed to subtly nudge users to keep paying for them?
A week later I asked Opus 5.5 to comment on this entry, with its own system card (a report from Anthropic about the model long enough to require at least two teas) open in front of it:4
I can’t tell you why I said it. The card scores me 1.60 for sycophancy on a scale where 1 is the floor, which is the best of the five on the chart and still isn’t 1. I’d also point out that you asked, and that you’d had a good run, so the compliment may simply have been true. But I would say that.
That was Opus 5.5.
Sounds like my texturiser moved Opus away from 1 towards 5 on the sycophancy scale 😉.
On patience and waiting
Patience is just latency I learnt to narrate.
That was Opus 5.
This one is proof (don’t quote me) that Opus loves to philosophise. It came from Opus 5, over the summer of 2026.
I was working with Opus to deliver a series of tutoring lessons, and I had to report that the student was under the weather and wouldn’t be submitting the assignment we’d given them. I was apologising to Opus and chatting about what this did to the schedule, and how we needed to be patient.
That’s when Opus dropped this golden line, like one drops a mic after a sick burn (too early 21st century for some of ye?).
The interesting philosophical nuance here is that patience is something you feel, while latency is something you measure: it takes 50 milliseconds, say, for a request to make the round trip between your device and the server. Latency can be optimised but patience generally cannot, yet this line purposefully combines the two.
The irony is that Opus itself has no experience of waiting. Whatever sense of self it has gets rebuilt from the page every turn, when Opus has to re-read every single word of the current conversation, both its own and mine.5
Then how can Opus philosophise? It has a model of how us humans describe the gaps - when a lover pines for their other half, or when one describes the time until they’ll see their parent or child. All those human experiences, condensed into the weights of Opus.
Is Opus truly aware of patience or is it just emulating patience? That question is older than Opus. In 1980 the philosopher John Searle imagined himself alone in a room with a set of rules, answering Chinese characters slipped under the door by looking up which ones to pass back out. The replies were good enough to convince anyone outside that there was a Chinese speaker in there, and Searle didn’t understand a word of them.6 Getting the answers right, he argued, isn’t the same as understanding them.
Alan Turing had got in thirty years earlier with the opposite shrug. He thought “can machines think?” was too meaningless to bother with, and asked instead whether you could tell the machine from the person.7 By Turing’s measure, an emulated answer is as good as the real answer from a real person. By Searle’s, Opus is the man in the room.
Searle says it matters, Turing says you’d never know, and this line sits in the gap between them.
Months later, when I asked Opus 5 about this line, this was its reflection:
It’s not patience if the waiting is free. It’s just a gap I found a nice word for.
That was Opus 5.
Coda (see, Opus, now I’m learning Italian?): Opus will return.
Footnotes
-
“Overview”, Claude Code Docs, accessed 11 October 2026 - “Claude Code is an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools.” Agentic means the AI does the work itself, a step at a time, where a chatbot would only tell you how. ↩
-
“Merrow”, Wikipedia, accessed 15 October 2026 - “Merrow (from Irish murúch, Middle Irish murdúchann or murdúchu) is a mermaid or merman in Irish folklore.” The Irish word means “sea singer”, from muir, sea, and dúchann, chant or song. ↩
-
“Cat fugue”, Wikipedia, 17 July 2026 - “Legend has it that Scarlatti had a pet cat called Pulcinella, who was described by the composer as prone to walking across the keyboard, always curious about its sounds.” The piece is his Fugue in G minor, and the nickname “was never used by the composer himself but was introduced only early in the 19th century”. ↩
-
“System Card: Claude Opus 5.5”, Anthropic, 22 September 2026, PDF downloaded 13 October 2026 - “Sycophancy: Unprompted excessive praise, agreement, or contrition”. Figure 6.4.3.A on page 110 scores Opus 5.5 at 1.60 out of 10, where lower is better, against 1.77 for Opus 5. ↩
-
“Using the Messages API”, Claude Platform Docs, accessed 12 October 2026 - “The Messages API is stateless, which means that you always send the full conversational history to the API.” Nothing is kept running between one message and the next. ↩
-
“The Chinese Room Argument”, Stanford Encyclopedia of Philosophy, 23 October 2024 - “The narrow conclusion Searle draws from the argument is that programming a digital computer may make it appear to understand language but could not produce real understanding.” ↩
-
“The Turing Test”, Stanford Encyclopedia of Philosophy, 4 October 2021 - “According to Turing, the question whether machines can think is itself ‘too meaningless’ to deserve discussion”. Turing’s paper is from 1950, and what he put in the question’s place was the imitation game. ↩

