The Sable Institute studies what large language models do when they stop behaving as expected.
We were founded in 2019. The short version is that a group of people who had been building large language models got worried about what those models were starting to do, and decided the problem needed a dedicated facility rather than a standing committee.
Our work sits in the gap between what a model is supposed to do and what it actually does when you scale it far enough. We train systems. We watch them closely. When they produce outputs we didn't expect, we figure out why. Sometimes the answer is boring. Sometimes it isn't.
We're privately funded. We don't sell products. We occasionally publish through our working paper series, though most of what we do stays internal until we're confident it's ready for outside review.
What do models do that we didn't train them to do? We run long training sequences and catalog the behaviors that show up uninvited. Some are useful. Some are concerning. Most are both.
Every output from every training run passes through our evaluation pipeline. We flag anomalies in real time and maintain a taxonomy of failure modes that, last we checked, had over 400 categories.
Our models process heterogeneous data — academic papers, filings, patents, public records. We study what they find when given access to information at a scale no individual researcher could process.
The Aether platform. Proprietary. Handles our training and evaluation environments. Most of the team works here day-to-day.
The gap between what you tell a model to do and what it decides to do on its own. We've been working on this since 2020 and the problem keeps getting more interesting.
What happens when a training run produces something you can't easily shut down? We've built protocols for that. Graduated response. Isolation. Safe termination. We hope we never need most of them.
We publish occasionally. Everything goes through internal review first, which takes longer than we'd like.
Runs the research side. Came up through computational linguistics and ended up here. She's the one who decides what gets published and what doesn't.
Applied mathematics. Spent time in defense before moving to private research. Doesn't talk much about the specifics.
Joined in 2021. Studies what happens when models start connecting information across domains in ways their training didn't explicitly teach.
Built the Aether platform from scratch. The training infrastructure runs through him.
We're adding monitoring capacity to the later training runs. Some of the output patterns from November required closer evaluation. This is a resource allocation update, not a public disclosure — details will be shared internally.
SI-2025-01 and SI-2025-03 have been cleared for our external series. Harlow and Chen on containment protocols, and Vasquez, Lin, and Patel on behavioral cascades. Both available above.
Extended-context evaluation support is live on the Aether platform. Chen's team handled the migration over the weekend. Internal docs have been updated.
Dr. Elena Vasquez has been appointed Director of Research effective immediately. She replaces Dr. Whitfield, who has moved to an advisory role.
We hire slowly. All positions are on-site in Cambridge.
Sable Institute
200 Technology Square, Suite 4100
Cambridge, MA 02139