One Cadence brain (cadence-net 0.91.1, 3.6M parameters) trained once on a small self-written chat grammar: 16 kinds of line, 188 words. It is not a general chatbot and knows nothing outside that grammar.
On held-out prompts it gives a valid reply 83% of the time. A plain MLP on the same input gets 95%.
The brain runs in Python, so it can't run on a phone. Every reply here was produced by the real brain on Sean's PC and saved: all prompts the grammar can make. Each prompt was answered after the same two warm-up exchanges, because with an empty memory window it mostly babbles. Nothing here is generated in your browser: if your line isn't one it knows, you see the closest prompt it does know, and that reply.
Code: cadence-lab/chat (replies made by export_web.py).
Independent experiment by sean-mcclane832. Not affiliated with Pragma Research. Cadence: muellerberndt/cadence. More experiments.