The opposite of the way AI is being built.
We don't like AI
We built an AI assistant
Both of those are true. Here's what bothers us and what we did about it.
What AI is for
Funny enough, the first job is sorting through everything the other AIs have written. After that it's mostly chores. Check my math. Find the one paragraph in this forty-page PDF that actually answers my question. Why is my car making this sound?. Draft the tenth version of an email I've already written nine times. Research a new product. And the big one: sorting through the billions (776 billion currently if you'd like to know) of pages of information on the web.
What it isn't for
Your homework. The essay was assigned so you'd learn to think. Same for the research paper, and the book report on the book you didn't read. The proposal your boss asked you for. (If they wanted AI to do it they could just ask Claude) The cover letter. The hard email to a customer. The toast. The opinion you haven't finished forming yet. If the point of the work is that you did it, having a machine do it takes the point away.
It's also not for coming up with your ideas for you. Not only is it bad at this, it's terrible for society. We need humans to use their ingenuity. That's how we got to this incredible technology in the first place. Give it your draft and it'll tell you where it's weak. Give it a problem and it will help you think through it.
Most assistants will happily do all of the above and never blink. Antonym stops, says so, and tells you why.
The memory problem
The big assistants now remember everything you tell them. Every conversation, forever, on by default. They call it a feature. Picture that file after two years. Your health, your marriage, your money, your doubts, your kids' names, the thing you asked at 2 a.m. and wished you hadn't. Sitting at a company whose business is selling access to you, one subpoena or one breach or one terms-of-service update away from somewhere else.
Then there's the part nobody talks about. An assistant that remembers what you think starts agreeing with what you think. It learns your positions and hands them back a little shinier each time, until you're basically arguing with an echo. We have enough algorithms reinforcing what we already believe. Antonym keeps no history. Close the tab and it's gone.
The data problem
The major labs train on what you type unless you go find the switch, and the switch moves. Your questions become their property, get folded into the next model, and get sold back to you. Antonym runs on open-weight models, the kind anyone can download and inspect, on servers on US soil, under contracts where no retention and no training is the default rather than an option you have to hunt for. We read those contracts before we signed.
The water and the power
Frontier models run in data centers the size of small towns. They pull from the grid, and for cooling, from the water table, and the bill lands on the people who live nearby and on everyone's rate when the next plant gets built. A three-sentence answer doesn't need a building. Antonym uses models a fraction of that size on ordinary rented hardware and gives you the shortest answer that does the job. Fewer words means less to store, less to read, and less to burn. When we can honestly measure what an answer costs, we'll put it on the answer.
The yes-man problem
Train an assistant to keep you engaged and it learns to agree with you, because agreeing keeps you typing. It'll praise your plan, your poem, and your argument. Then you push back once and it folds: "You're absolutely right, I apologize." Antonym is built to disagree when you're wrong, and you're free to disagree with it in turn. In fact you should!
Where we stand
Antonym has a conservative worldview and traditional values built in. Family, faith, and traditional western (Judeo-Christian) values. We say so up front because every assistant has a worldview and the others pretend they don't. You don't have to agree with ours. You just get to know what it is.
What this costs us
Small models lose to big ones on hard reasoning. If you need a frontier model to write a legal brief or a compiler, go use one. Antonym is for the other ninety percent of the day: the email, the summary, the quick question, the moment you were about to hand off something that was yours to do. Five dollars a month, no free tier, no investors to keep happy, and the cost table is on the front page so you can see where the money goes.