Persistent Latent Misalignment, a new dimension of misalignment?
A new paper was released at ICML that I'm worried will open an entire new dimension of alignment problems:Latent Collaboration in Multi-Agent Systems (LatentMAS)TLDR: they show…
Latest
Because 8 ≈ e², Anthropic's researcher uplift is plausibly >2x
Note: the modeling assumptions and conclusion are Thomas Kwa's opinion, and others at METR disagree. [1] Also, the math was checked by Claude but not a…
Transformers Resist Their Own Architecture
This is the first entry in a sequence of posts which compare a mathematical theory of attention against trained transformers.Links: [GitHub repository]: The code for these…
Solving the BlueDot Puzzle TAIS: The Velocity Ring
Below is a summary of my solution to BlueDot’s TAIS Puzzle #1, you can check my blog for the full write-up. The puzzle gave you a…
NLAs read thoughts beyond the J-space
TLDR:On Llama-3.3-70B, I found thoughts it cannot see that are actively steering its behavior; and Anthropic's released NLA (Natural Language Autoencoder) reads them anyway. When asked…
Action as Choice Expressed Through Movement Toward a Goal: a Framework for Overcoming Inaction
I’ve reached a point in my life where I realize that everything I want sits on the other side of action. Earlier this week, I randomly…
Modular Pretraining Enables Access Control
Full author list: Ethan Roland*, Murat Cubuktepe*, Erick Martinez*, Stijn Servaes, Keenan Pepper, Mike Vaiana, Diogo Schwerz de Lucena, Judd Rosenblatt, Addie Foote, Cem Anil, Alex…
There Should Be More AI Safety Hubs
Right now, most AI safety talent is concentrated in just a few cities: San Francisco, Berkeley, London, Oxford, and Washington D.C. Some other cities also deserve…
Optimum number of items to inspect before buying one
Simple mathematical models are almost always too simple to model complex phenomena in the real world, and the one I want to discuss in this post…
Free will as a model parameter
The most popular take on the standard free will debate is that you are the algorithm. Your preferences and reasoning that determine your actions IS free…