Conw AI

Conw AI is an independently served language model that learns directly from the people who use it. Instead of routing every request to an external answer API, the project runs its own model on its own machine and folds verified user teaching back into the system through what it calls a guarded learning loop. When you correct a reply, rate it, or explain a word the model got wrong, that lesson can eventually become part of the model itself, and the whole process stays visible along the way. The pitch is a participatory alternative to conventional chatbots, an AI whose knowledge you can actually shape and audit rather than a black box trained on scraped data you'll never get to inspect. It's a small operation making deliberately modest claims, and that modesty turns out to be central to what it's selling.

The current model is Conway-Retrain 12B, a retrain built on a Gemma 4 base and served through MLX inference on the project's own hardware. That hardware is a single 16GB iMac, which handles both answering live conversations and running the background learning jobs. An earlier and much smaller model, Conway-Omega 188M, has been deprecated inside the product but remains open source on Hugging Face for anyone who wants to study it. None of this is dressed up. The site is unusually blunt about the scale of the operation, and the team's stated position is that most AI marketing is fiction while theirs can't afford to be, because users would notice the lie within a day. Capability claims and environmental claims both stay grounded for the same reason.

Learning happens in three stages. Teaching comes first. You correct replies, rate them, or explain terms, and confirmed meanings enter your private memory immediately, so the model can draw on them in your future conversations without waiting for anything to retrain. Filtering comes second. Lessons that are sensitive, unsafe, badly formatted, or down-rated stay out of the shared pool. Promotion comes last. Actual weight updates run as background processing during idle server time and have to pass checks before going live, so a training job never gets in the way of a conversation. The design principle underneath all of it is that confirmed teaching works right away while weight updates never block a reply, which means your corrections pay off instantly even though the deeper model changes take longer to land.

Vocabulary gets special treatment. When users explain words, those meanings can be verified against a dictionary, and dictionary-verified terms can become shared knowledge across the whole user base instead of staying locked inside one person's sessions. That builds a communal layer of vocabulary that grows through collective contribution rather than scraping. The gates matter just as much here. Unsafe, invented, or poisoned lessons are kept out of the shared model, and anything sensitive or down-rated stays private to the chat where it happened. You also get visible proof of what was learned and what was rejected, which is the piece that separates this from systems that quietly absorb feedback with no accounting anyone can check.

Conw AI is for people who want to understand exactly what an AI has learned and to have a real hand in shaping it. It's explicitly not competing with the frontier models on raw capability, and the site says so in plain terms. A 12B model on a single desktop machine won't match the big labs on benchmarks, but it also doesn't come with their opacity, and the practical guardrails reflect the constrained setup honestly. Short-answer response budgets keep output from being wasted, and repetition guards stop runaway generation before it burns compute the machine doesn't have to spare. If you're comfortable trading some headroom for transparency, efficiency, and a system you can actually influence, that's the trade on offer, stated openly rather than buried in fine print.

Access is freemium. The free plan covers casual weekly use, while the Pro and Max plans raise the weekly allowance and help fund the operation. Every tier gets the same answers, and the site is explicit that there are no ads and no selling of user data, which keeps the incentives pointed at the people paying rather than at advertisers or data brokers. There's also an API priced at £1.50 per million tokens counting input and output together, with a £5 minimum top-up, for anyone who'd rather build against the model directly than work through the chat interface. Questions go to hi@conw.ai, an address that reaches the people actually running the machine.