Deng Nan

[Deng Nan @deng_nan]

@jakieuh8i @GithubProjects MatrixOne stores data as immutable objects on object storage, and branching is copy-on-write. Creating a branch only writes metadata, so it's near-instant and near-zero cost regardless of table size. After that, you pay only for data the branch actually changes.

[Deng Nan @deng_nan]

Tested TypeSafe's Jev on real resume screening. Same CV scored 5x: score drift ≤2 pts vs up to 31 for our LLM pipeline. ~0.6s vs ~23s per CV. Accuracy on par. For text judgment it sits between LLMs (flexible but slow, inconsistent) and NLP models (consistent but narrow).

[Deng Nan @deng_nan]

@tamarajtran @typesafeai The scoring is the easy half. The part that surprised us: dropping items mid-context breaks the KV cache prefix, so the tokens you save on the request can get spent back on a cache miss. Does Jev score into a stable prefix, or does it recompute the window each turn?

[Deng Nan @deng_nan]

The question was 17 tokens. The model received 41,820, and 11,269 of them were tool definitions.
Astra is an open-source agent runtime that prints an EXPLAIN ANALYZE plan for every model call: what went in, what it cost, what got dropped.
https://t.co/lzvriMezER