Connor Leahy

EleutherAI

Leahy co-founded EleutherAI with Sid Black in 2020, according to Conjecture's company history, which describes EleutherAI as an open-source language-model community that developed GPT-J and GPT-NeoX. Leahy is one of the authors of The Pile, a paper posted on 2020-12-31 that describes an English text corpus constructed from 22 subsets for training language models.¹²

Leahy is listed with Sid Black, Leo Gao, Phil Wang and Stella Biderman in the citation entry for GPT-Neo, a March 2021 software release for training GPT-style models on TPUs with mesh-tensorflow; the repository has since been archived in favour of GPT-NeoX. Leahy is also a co-author of the GPT-NeoX-20B paper (April 2022), whose individual-contributions appendix lists Leahy among the contributors to its "Alignment Implications" section.⁵³⁴

Conjecture

In March 2022 Leahy, Sid Black and Gabriel Alfour founded Conjecture, which the company says grew out of EleutherAI and was set up to work on the control problem of advanced AI. The company's announcement post of 2022-04-08 introduced it as an alignment research startup founded by the three.¹⁶

Conjecture describes its current work as building language models and an approach it calls Cognitive Emulation, aimed at AI systems that are scalable, auditable and controllable. On 2024-12-02 Leahy and Gabe Alfour published "Conjecture: A Roadmap for Cognitive Software and A Humanist Future of AI", setting out the company's plans for 2025 and beyond.¹⁷

Work and contributions

• The Pile Published research

  Listed as an author of The Pile paper (2020), which introduced an 825 GiB English text corpus built from 22 subsets for training large language models.²

• GPT-NeoX and GPT-NeoX-20B Open-source code

  Co-author of the GPT-NeoX-20B paper, whose individual-contributions appendix credits Leahy with work on the alignment-implications section, and listed among the authors in the GPT-NeoX repository's citation entry for the Apache-2.0 training library.³⁴

• GPT-Neo Open-source code

  Listed among the five authors in the software citation for GPT-Neo (March 2021), EleutherAI's implementation of model-parallel GPT-2 and GPT-3-style models using the mesh-tensorflow library.⁵

Timeline

1. 2020Co-founded EleutherAI with Sid Black.¹
2. Dec 31, 2020Co-authored "The Pile", EleutherAI's paper on a text corpus for training language models.²
3. Mar 2021Listed as a co-author of the GPT-Neo software release for training GPT-style models with mesh-tensorflow.⁵
4. Mar 2022Co-founded Conjecture with Sid Black and Gabriel Alfour.¹
5. Apr 8, 2022Conjecture published its launch post introducing the company as an alignment research startup.⁶
6. Apr 14, 2022Co-authored "GPT-NeoX-20B", contributing to the paper's section on alignment implications.³
7. Dec 2, 2024Co-authored Conjecture's roadmap for cognitive software with Gabe Alfour.⁷

Documented roles

The sources do not confirm a current role; these are the roles they document.

• Co-founder (2020), EleutherAI¹
• Co-founder (founded March 2022), Conjecture¹

Companies in their history

• EleutherAI 5 people in the wiki

Related people

• Leo GaoConnected to EleutherAI
• Sid BlackConnected to EleutherAI
• Stella BidermanConnected to EleutherAI
• Katherine CrowsonConnected to EleutherAI
• Brandon DuderstadtLanguage models · Datasets and data
• Armand JoulinLanguage models · Datasets and data
• Ronen EldanLanguage models · Datasets and data
• Kyle LoDatasets and data · Language models
• Vaishaal ShankarDatasets and data · Language models
• Ce ZhangDatasets and data · Language models