conw.ai — The AI that learns with its users.
Usefulchatsteach it.Seewhatstuck.<br>Conw is independently served. Confirmed teaching works immediately, while safe and useful chats can become reviewed learning candidates — you can see what was saved and tell us which replies are worth keeping.
No card required 30-second setup Free to use
Verified learningIndependently servedLocal MLX inferenceLearns from useful chatsLiving-room iMacFree to useVerified learningIndependently servedLocal MLX inferenceLearns from useful chatsLiving-room iMacFree to use
How it learns<br>You teach. It filters. It proves what stuck.
A guarded learning loop that improves from useful signals without letting one bad chat rewrite the system.
01<br>You teach<br>Correct a reply, rate it, or explain a new word. Confirmed meanings enter your private memory immediately so Conw can use them in the same conversation.
02<br>It filters<br>Sensitive, unsafe, badly formatted, and down-rated replies stay out. Only short, useful examples can enter the reviewed learning queue.
03<br>It proves it<br>Background updates wait for enough clean examples and an idle server. A candidate must pass checks before it can replace anything live.
Confirmed teaching works immediately. Weight updates never block your reply.
The night shift
When the iMac is quiet, it reviews.
Eligible examples wait in a guarded queue. Training starts only after there is enough clean material and serving is idle. The result is a candidate, not an automatic live update.
SignalOnly safe, useful, short replies survive. Down-rated and contaminated replies stay out.
ScheduleBackground work waits for the iMac to be idle so live chat keeps priority.
PromotionA candidate must load and pass its checks before anything live can change.
guarded queue
live chat first · learning second
Principles
Three things we can say with a straight face.
01.<br>Most AI products hide the learning loop. Conw shows whether a lesson passed or failed, and keeps the useful memory even when the weight update is rejected.
Self-taught<br>It learns from the people who use it.<br>Confirmed teaching is written to memory immediately. Dictionary-checked meanings can become shared vocabulary; unverified or unsafe lessons stay inside the original chat. Any future weight change still has to pass safety and quiz checks first.<br>Confirmed meanings work immediately from memory<br>Dictionary-verified vocabulary can help every user<br>Unsafe, invented, or poisoned lessons stay out<br>0gates before a learned weight is promoted
02.<br>Conw is a workshop, not a monument: the serving model can change while memory, safety, verification, and user ownership stay intact.
Independent by design<br>Our serving stack. Our hardware. A replaceable checkpoint.<br>Conway-Retrain 12B is now live, retrained on top of a Gemma 4 base and served through MLX on our own machine, not an external answer API. The earlier 188M Conway-Omega checkpoint is deprecated in-product, but stays published, open-source, on Hugging Face. The checkpoint stays behind an interface, so memory and safety rules survive future model changes.<br>Inference is served by Conw, not forwarded to another model API<br>Checkpoint-specific thresholds live in calibration, not product code<br>The learning loop survives future checkpoint replacements<br>0local serving host for the current product
03.<br>The greenest token is the one we never need to generate. Efficiency starts with useful answers, bounded output, and training that never starves the person waiting.
Light by design<br>Efficiency we can measure, not invent.<br>Conway-Retrain 12B currently serves from one 16GB iMac. Short-answer token ceilings avoid waste, repetition guards stop runaway output, and background learning yields whenever chat traffic needs the machine. We will not publish a carbon number until we can meter it properly.<br>One 16GB serving machine instead of a GPU cluster<br>Response budgets stop needless token generation<br>Background learning defers when inference is busy<br>0GBmemory budget for serving and learning
Platform API<br>Build on Conway-Retrain from your own code.<br>Pay-as-you-go access to Conway-Retrain 12B and Conway-Omega 188M, live at api.conw.ai — the same model behind the chat product, reachable from any OpenAI SDK.<br>View the API<br>Drop-in OpenAI-compatible endpoints — swap base_url and go.<br>£1.50 per 1,000,000 tokens, input and output, either model.<br>Named keys, usage graphs, and a £5.00 minimum top-up.
Why we tell the truth
“Most AI marketing is fiction. Ours can't afford to be — you'd notice tomorrow if we lied today.”
A model that improves in public has nowhere to hide.
FAQ
Straight answers.
Anything we missed? [email protected] — a human writes back.
Does Conw really learn from my chats?Yes, but not by blindly training on everything. Confirmed teaching enters memory immediately. Dictionary-verified meanings can become shared vocabulary, while sensitive, unsafe, invented, badly formatted, and down-rated lessons stay out. Any future weight update still needs...