We develop our own AI models, computing tools, and research prototypes. Explore what each system does, its current stage, and the results recorded so far.
Word representations, short-text generation, and sentence-pair classification. Each model has a specific input, output, and evaluation.
01 / WORD REPRESENTATIONS
Noun Collapse
Turns words into numeric representations that can be compared by meaning. It learns the relationships between nouns from Wikipedia text using one geometric representation per word.
INPUT
A word or a sequence of words.
OUTPUT
Word vectors and similarity scores for comparing words.
68.92%Reported accuracy on the SNLI sentence-pair test
Parameters
~12.9M
Output classes
3 relationships
Above hypothesis-only
+7.42 points
CPU encoder latency
0.33 ms
Trained research model
Technical evaluation & benchmark comparisons
Noun Collapse
The 0.3616 noun-similarity result covers 662 of 666 noun pairs in
SimLex-999. The model trained on roughly 7.5% of Wikipedia; a
random same-shape control scores about 0.022. This is a
lexical-similarity evaluation. Equal-corpus, multi-seed training
comparisons remain future work.
Premise Generator
The recorded token accuracy is 52.71%. The roughly 6 ms CPU
measurement applies to short replies in the recorded setup. The
model generates contextual text; improving consistency with the
requested relationship label remains an active task. The example
above is a recorded model output, not a live response.
Supervised Collapse NLI
Reported SNLI results: GloVe average 60.70%, hypothesis-only
61.50%, collapse 68.92%, and a stronger frozen-embedding MLP probe
70.13%. Matched end-to-end comparisons across multiple seeds are
needed to isolate the contribution of collapse dynamics. The 0.33
ms figure measures already-tokenized encoder latency.
Memory systems for storing and retrieving information, recording changes, checking retrieval quality, and testing reversible read and write operations.
Geometry-controlled circuit simulation, state-vector and tensor-network experiments. These are software simulation projects with separate implementations and tests.