We’re at the Oculus building in NYC helping the public understand how LLMs work using workbench.ndif.us. Come see us!
NDIF @ PIT Pop Up. Join me! luma.com/rwbs06zm
Adam Belfki
@adambelfki
Joined July 2018
- i know this guy what you should know is that @ndif_team is THE place for how to do interp in the vision spaceWe have our first keynote speaker @jadenfk23 from @ndif_team giving a hands-on tutorial showing how nnsight and ndif make interpretability research easy to use and less scary!
- The team has worked very hard to build a white-box hackathon infrastructure so you can investigate and control model internals, compute-free! Join the competition and help develop principled methods for understanding latent model beliefs.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.
- 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:Replying to @sheridan_feuchtTurns out, this is because MLP 18 does base-10 addition. Llama first computes a sum (e.g., "four months after October" → 4+10=14), and only in later layers applies modulo to map back to a month (14→Feb). We find that Llama re-uses this addition mechanism across several tasks.
- Fellow mechinterp researchers: is @ndif_team NNsight slow, or are you just using it wrong? 🙃 A new tutorial by @adambelfki shows how to speed up large-scale experiments with session, batching, caching and skipping for 130x speedup! 🔥 Check it out ⬇️ nnsight.net/blog/2026/04/3…
