All projects
Oct 13, 20262 min read

Universal sorting

A system one model that organises the world's knowledge.

Parameters
2B
Examples
10K
Hardware
L4
Training time
2 hrs

System one models provide ultrafast structured replies to complex questions. Decision models like Jev will become the default way software communicates and extracts knowledge from AI models.

We introduce a new type of system one model, lex. Lex is a universal sorting model. It can arrange pieces of information given an arbitrary criterion. Give lex a list of job titles and tell it to order by seniority, and the model returns the list sorted with scores:

1. CEO                                 score=+7.378
2. Director                            score=+3.241
3. Manager                             score=+0.901
4. Senior developer                    score=+0.160
5. Junior developer                    score=-3.228
6. Intern                              score=-5.440

Examples

Try it below. Press shuffle to scramble the input order, then sort to watch lex reorder each list. Switch criteria to see the same mechanism applied to size, urgency, and spiciness.

Criteria · Job titles

Unsorted

Rank these job titles from most to least senior

  1. CEO
  2. Director
  3. Manager
  4. Senior developer
  5. Junior developer
  6. Intern
1 / 4

How is lex different from a reranker?

A reranker answers one question: how relevant is this document to this query? It is built for retrieval, where the criterion is always relevance and the scores are used to cut a list of candidates down to the best few. Lex takes the criterion as an input. Seniority, urgency, size, spiciness — anything you can describe in a few words becomes a sort order, and the items do not have to be documents or have anything to do with a query.

Applications

Lex can be used in data-heavy websites to allow sorting with user-defined fields.

A few other places it fits:

  • Sorting your mail by lead potential.
  • Ordering a support inbox by urgency or by tone.
  • Reordering a feed or a set of search results by whatever the user types in.

Get started

Request early access to the API.

Next: AI models learning AI