We recently had a new addition to the family. His name is Cabbage.
Named by Charlize, my youngest, who in a moment of inspiration, also put forward two other names for contention: Constipation and Chlamydia. I like Constipation but dreaded the other because I will not appreciate responding to, “the doctor will see the owner with Chlamydia now.” during my future sojourns to the vet.
Cabbage is a cotton furball of a dog, impossibly small and fluffy. He entered this household that already contains Cora, a domestic shorthaired and short-tempered orange tabby. Some say she has the resting face of a pure evil Mona Lisa.
We have been doing the introduction cautiously. Both animals have been permitted to observe each other from behind the safety of their respective cages. Cora hisses. Seven-week old Cabbage stares back with the blissful vacancy of someone who doesn’t yet understand consequences.
And while the hissing has been getting less frequent on the second day, which I choose to read as progress, I do secretly fear that Cora is merely conserving energy for the main event when they do meet. I have seen what Cora is capable of when she meets another neighbourhood cat or a trip to the vet. I do have my reasons (and scars) to be cautious.
All of this has gotten me thinking about the AI tools that I use daily.
My daily driver is Claude. Somewhere along the way I’ve realised that Claude is, unmistakably, a dog. Specifically, an overly enthusiastic Golden Retriever. Eager to please, boundlessly energetic, genuinely multi-talented: it can code, it can think, it can manage files, draft documents, debug itself mid-task and then circle back to fix what it broke while fixing the previous thing. It wants to help. It really wants to help.
But if you have ever watched an excited Golden Retriever move through a small, densely packed room, you’ll understand the risks. I’ve watched it try something, then try something else, then revisit the first thing, then confidently announce it is done before discovering a new edge case and setting off again. You have to hold the leash tight constantly.
Meanwhile, OpenAI is like my cat Cora: cool, composed and operating entirely on its own schedule. It is undeniably impressive when it decides to perform and it will sit there in the background with a kind of elegant detachment, occasionally glancing at you with judgmental eyes.
You don’t prompt OpenAI so much as make a request and wait to see if it deigns to engage. Most of the time, it does something generically AI. You leave it alone until the next major update (when it shows up) and does a spectacular new party trick and everyone says, “well, that’s cute”.
The truth about AI tools is that their output is probabilistic. A small tuning decision by some engineers in Anthropic or OpenAI can cause significant behaviourial shifts. This has become the unfortunate reality when working with frontier models. Some models are good for certain things. And when we can get them to work together, harmony ensues.
Which is also a reasonable description of integrating a new puppy into a household with an established cat.
The best-case scenario is that Cabbage and Cora find their respective lanes, reach some kind of détente, and maybe, eventually, become something like the best of friends. The worst case is that they amplify each other’s worst qualities and the house descends into a state of continuous, low-grade chaos. Same odds, really, as any two AI tools you’re trying to get to work in concert.
As for Charlize’s readiness to take care of a dog, only time will tell.
But she did settle on Cabbage instead of Chlamydia, so I remain cautiously hopeful.