The National Deep Inference Fabric, an NSF-funded computational infrastructure to enable research on large-scale Artificial Intelligence. https://nitter.cf/t.co/STsQ707an3
Boston, MA
Joined May 2024
- Tweets135
- Following0
- Followers584
- Likes0
We convened in Boston last month to announce the winners, compare approaches, and debate what the field should try next. Stay tuned for a comprehensive report, and check out our blog post detailing the competition: aletheias-quest.github.io/wi…
Congratulations to the winners, including 1st place teams GMU-Veritaserum (@Skywalkerx47) and FlaiChi (Serhii Honcharenko). Thanks to organizing team (@walterlaurito, Kieron Kretschmar, @emmabortz, @jadenfk23, @gsarti_, @peterbhase, @davidbau, Olivia Floody, and Yik Siu Chan)!
Announcing our bi-weekly reading group focused on meta-learning and cognition, led by @ericwtodd and Arjun Khurana.
We read the 1986 backpropagation paper together last week! Check it out:
youtu.be/vgbQh5lx9yU?si=PG7m…
Join us next Wednesday, 9/23 at 3PM ET for our second installment: luma.com/ji8qc5n3
We'll be reading the J-lens paper (transformer-circuits.pub/202…)
Check out IBM's new Steerability toolkit! Would you be interested in using this with NDIF's backend? 👀
We're excited to announce the Steerability toolkit!
There are a ton of methods on steering out there but not a lot of cohesion in how they are built. One of our main goals with Steerability is to provide some structure to this space. We've spent a good amount of time designing abstractions that make sense. We also allow for steering methods to be composed into chained model operations (which we call steering pipelines).
The toolkit is compatible with PEFT/TRL (for training) and Inspect (for evals). We also provide support for steering via vLLM through vLLM-Hook (github.com/IBM/vLLM-Hook).
There are many more features that we're actively working on so keep an eye out.
Let us know what you think!
Repo: github.com/generative-comput…
Docs: generative-computing.github.…
New NNsight pre-release out now! vLLM optimizations make high-throughput, scaled experiments easier than ever before.
Check out Jaden's thread/blog post for more info on what's inside. 👀
Blog post: nnsight.net/blog/2026/09/09/…
The application deadline for our AI lie detection competition with @cadenza_labs is approaching on June 21st! Multiple suspect model organisms to interrogate and probe, $50K prize pool, and no local GPU required (it all runs on NDIF).
Application: forms.office.com/r/BXVwPEweJ…
Website: aletheias-quest.github.io/
Can you tell when an AI model is lying?
Announcing Aletheia's Quest, an AI lie detection challenge running this summer, organized by @cadenza_labs and @ndif_team.
Multiple model organisms to interrogate and probe, $50K prize pool, no local GPU required.
We'll have a $50,000 total prize pool, split across the highest-ranking methods and handpicked judge awards.
Best teams will also get recognition as co-authors in the competition report and an invite to our final in-person convening in Boston.
Organized by @cadenza_labs and @ndif_team, with support from @schmidtsciences and AWS.
Join the Discord 👉 discord.gg/4NX3sUC4ge
Apply by June 21 👉 aletheias-quest.github.io/
Bring your best lie detector!
Check out @adambelfki's exploration of @sheridan_feucht and @tal_haklay's addition mechanism findings at @GoodfireAI in Workbench, no code required.
Try it yourself using workbench.ndif.us! Post your Workbench experiments and tag us; we'll send you some swag 😎
When @sheridan_feucht first told me about these results I was kind of skeptical, until they mentioned I can trace the modulo base-10 addition in Llama 8B just using Logit Lens. 🔍
So I opened workbench.ndif.us to check it out myself, and this is what I saw:
Interested in optimizing your experiments on NDIF? Check out our new blog post by @adambelfki detailing how he helped one of our users speed up their experiment by 130x!
nnsight.net/blog/2026/04/30/…
