ThoughtVectors — dialogue in thought-vector space
A 48M-parameter model that converses without token-level language modeling in its reasoning loop: your message becomes thought vectors, a thinker predicts response thoughts, a frozen decoder renders the reply. Trained from scratch in ~25 GPU-hours on one consumer GPU. The whole model runs in your browser — nothing you type leaves this page. Paper + code
loading…
Small talk is its domain. Decoding is greedy, so the same conversation
always gets the same replies. Known failure: it may respond cheerfully
to bad news — that's the paper's case study, go ahead and probe it.
The picker below swaps in the paper's matched token-LM baseline (same
data, tokenizer, and compute) for comparison — it has a higher
per-turn ceiling but fails differently: apology as its default
register and repetition loops it can't leave.
thinker: ~135 MB, downloads once and is cached. token-LM baseline: a further ~100 MB, downloaded the first time you pick it. Both are byte-exact with their released checkpoints (8/8 greedy-reply parity vs the PyTorch reference). Switching models mid-conversation compares them on the same transcript, same as the paper's live-chat comparison.