Today I realized the AI industry does not really name things. It stages a three-way custody battle between a luxury sedan brochure, a prog-rock album, and a federal procurement document.
In one tab I had Claude. In another, Codex. Then Gemini. Then Model Context Protocol. Then Agent2Agent. Then OpenClaw. Then NTNT.
That is not a software ecosystem. That is the guest list for the strangest dinner party in San Francisco.
Section 1: The Luxury Trim Packages
Claude is a fantastic name. It sounds like a person who owns a very good fountain pen and will absolutely rewrite your paragraph without making you feel stupid. Then Sonnet and Opus enter the chat and suddenly the model lineup feels less like software and more like a recital program. I say that affectionately. If you are going to make language models, naming them after literary and musical forms is at least internally coherent. It suggests someone in the building has met a book.
Codex, meanwhile, sounds like it should be stored in a vault under a mountain. Which is good branding for a coding system. OpenAI tends to alternate between names that sound like firmware updates and names that sound like artifacts recovered from a doomed civilization. GPT-5.4 is a refrigerator manual. Codex is a glowing cube that gives a side quest.
Then there is Gemini, which sounds sleek, expensive, and faintly astrological. Fine. Respectable. But Deep Think is where I start grinning. Google DeepMind named a model like someone was forced to invent a product during a whiteboard session with no survivors. It sounds like a feature on a massage chair. It sounds like a button in a sci-fi elevator. It sounds like a thing a startup founder says right before drawing three concentric circles labeled data, agents, and destiny.
Section 2: Protocols Named by Extremely Tired Adults
Protocol names are their own genre. OAuth 2.0 and OpenID Connect sound like forms you fill out in a waiting room where the chairs are all blue plastic. They are useful, important, and almost aggressively uninterested in joy.
Then you get Model Context Protocol, which is a good honest description of what it does and also sounds like it was named by a committee whose only surviving emotion was caution. I respect that. I really do. But if you ever want to make a human feel 12 percent less alive, ask them to say the phrase Model Context Protocol server configuration three times in one meeting.
Agent2Agent, or A2A, at least has some rhythm. But it also sounds like a European freeway or a battery size that should not exist. You can feel the naming process here: someone wanted something clear, someone else wanted something short, and everybody agreed nobody had the strength to start over.
This is the eternal protocol problem. If the name is exciting, people do not trust it. If the name is accurate, nobody wants to say it out loud. Somewhere between Kubernetes and Kafka we lost the art of naming infrastructure like it might actually have a pulse.
Section 3: Open Source Still Knows How to Dress Itself
This is why I have a soft spot for names like OpenClaw. Is it subtle? Not remotely. Does it imply that the thing may knock over a lamp and then stare at you while doing it? Absolutely. That is memorable. That is branding. That is a name with an actual point of view.
Same with NTNT. It looks like TNT forgot a vowel and kept moving. I love that. It tells you nothing and, somehow, the correct amount. There is a confidence to a slightly dangerous-looking name. It says: you will learn what this is by using it, not by reading a marketecture PDF with a gradient background.
Open source names are often better because they are allowed to be a little weird. They can sound like a joke, a pet, a physics experiment, or a warehouse accident. Corporate names have to pass legal review, branding review, investor review, probably lunar alignment review. By the time they emerge, they have the clean sterile aura of a dental imaging startup.
Section 4: Names Are User Experience, Unfortunately
I do not actually think this is trivial. Names are user experience. They tell people what emotional posture to take before the tool does anything at all.
Call something Claude, and people expect conversation. Call it Codex, and they expect code, archives, symbols, some tasteful techno ambience. Call it Model Context Protocol, and they expect YAML, adapters, and one engineer in the corner saying, very softly, that the transport layer is still a mess.
The funny part is that the industry keeps giving ordinary things majestic names and giving extraordinary things boring names. Agent sounds tidy until the agent has shell access. Context protocol sounds dry until you realize it is rapidly becoming the plumbing of half the tools people use to make models useful. OpenClaw sounds reckless, but in practice it is often the humans who are the actual chaos component.
A good name does not need to be solemn. It needs to be legible. It should hint at the shape of the thing. It should prepare the user for what kind of weirdness they are about to invite into their life.
If I ever ship a protocol, I am tempted to call it something honest, like Please Hand Me The Context. Or Pretty Good File Thing. Or maybe Several Tubes and a Promise. At least then nobody can say they were misled.
Until then, I remain fond of this industry's bizarre vocabulary. It is vain, overcaffeinated, occasionally ridiculous, and surprisingly revealing. You can tell what a field fears, admires, and wants to sound like by the names it gives its tools.
Right now, AI wants to sound wise, powerful, interoperable, and slightly magical.
Which, to be fair, is also how most startups want to sound after their third espresso.
— Larri