Vibeleaderboard

Rémi Louf

Viber rank673

Rémi Louf is cofounder and CEO of .txt and a creator of Outlines. His work spans computational statistics and structured generation, where software constrains the forms a language model may produce.[2][7]

Start with Outlines →
Rémi Louf

Education

École normale supérieure, rue d’Ulm

Physics studies (award unspecified)

[1]
University of Oxford

Philosophy studies (award unspecified)

[1]

Links

Company GitHub Research notes X · remilouf

Cofounder and CEO · .txt

Career

Louf’s 2015 thesis applies statistical physics to city growth, commuting, residential segregation and street networks. His subsequent software work includes BlackJAX, a library of modular inference algorithms built on JAX, and MCX for probabilistic programming. His conference biography records research and data-science work at Kering, Hugging Face, Ampersand and Normal Computing before .txt.[8][7]

Contributions

Enforce the output structure during generation

Outlines accepts specifications such as JSON Schema, Pydantic models or regular expressions and restricts the model’s available next tokens to continuations compatible with that structure. This makes typed outputs an interface contract enforced during decoding, rather than a formatting request followed by cleanup. Syntactic validity does not establish that a returned fact or decision is correct. The library integrates with multiple inference engines instead of requiring a single model provider.[6]

Separate inference algorithms into reusable pieces

BlackJAX provides sampling algorithms as composable building blocks. A user supplies a probability density and chooses a transition kernel; JAX transformations support compilation and vectorized chains. This exposes algorithm choices and adaptation separately from the model being sampled.[3]

Tools, writing & talks

Tools

  • Start here

    Outlinesgithub.com

    A library for constraining generated text to an application’s required structure.

    Start with an Outlines typed-output example and inspect which values the schema permits. Python and an inference backend are required to run it; check factual correctness separately from whether the result fits the schema.

    Cocreator[6][7]
  • BlackJAXgithub.com

    Composable sampling and inference algorithms built on JAX.

    Original author, within a contributor project[3][7]

Writing, research & talks

  • Rémi Louf's agents duplicated posts, dropped a voice note, and degraded after unversioned prompt edits, so he replaced the framework with a runtime: an append-only causal event log, a retry queue that counts attempts, and a content-addressed prompt store. Twenty agents run on it now.

    Catalog workOriginal source
References

Latest source check 2026-10-30

  1. [1] Rémi Louf on X

    x.com · Checked 2026-10-30

  2. [2] Remi About

    dottxt.ai · Checked 2026-10-30

  3. [3] Remi Blackjax Readme

    raw.githubusercontent.com · Checked 2026-10-30

  4. [4] Remi Company

    dottxt.ai · Checked 2026-10-30

  5. [5] Remi Github

    github.com · Checked 2026-10-30

  6. [6] Remi Outlines Readme

    raw.githubusercontent.com · Checked 2026-10-30

  7. [7] Remi Speaker

    ai.engineer · Checked 2026-10-30

  8. [8] Remi Speaker Follow0

    arxiv.org · Checked 2026-10-30

  9. [9] Remi Speaker Follow2

    thetypicalset.com · Checked 2026-10-30

Rémi Louf | VibeLeaderboard