Asked a bunch of mechanical turkers what one question they would ask to determine if they were talking to a human or AI.
fave reply: When is the last time your teeth felt like they had little sweaters on them?
New project is live: gen.studio
a collaboration with @metmuseum & @microsoft
For each object at the Met, we can
> find a corresponding object in GAN latent space
> traverse latent space between objects to discover artworks that could have been made, but never were
MAIA (A Multimodal Automated Interpretability Agent) is here! 🧵
📝New paper: arxiv.org/abs/2404.14394
🌐Website: …imodal-interpretability.csail.mit.edu/maia/
Agents like MAIA advance automated interpretation of AI systems from one-shot feature description into an interactive regime where hypotheses
So, update, I'm founding @TransluceAI with @JacobSteinhardt. In thinking about where I’d go next with the automated interpretability work I’ve been doing, a few things were important:
(1) to scale our ability to understand AI systems, both their internals and their behaviors
(2)
Announcing Transluce, a nonprofit research lab building open source, scalable technology for understanding AI systems and steering them in the public interest.
Read a letter from the co-founders Jacob Steinhardt and Sarah Schwettmann:
transluce.org/introducing-tr…
Saved this from Passover but you know what? Most of these holidays are about something like this, and if your practice has grown too stiff, maybe it’s the perfect moment for a little re-examination
(originally from @alananewhouse)
🚨 NEW PAPER 🚨
Understanding increasingly large and complex neural networks will almost certainly require other AI models (themselves uninterpretable!)
How should we evaluate automated interpretability methods?
Introducing our new benchmark, FIND: huggingface.co/papers/2309.03…
I have a new paper out!
To reason physically about the world, the brain encodes abstract representations of key variables (like mass) that generalize across scene differences.
This invariance is what we'd expect to support a physics engine in the brain:
A Multimodal Automated Interpretability Agent
This paper describes MAIA, a Multimodal Automated Interpretability Agent. MAIA is a system that uses neural models to automate neural model understanding tasks like feature interpretation and failure mode discovery. It equips a
