Benjamin Laufer

@laufer

PhD student at Cornell Tech. bendlaufer.github.io

201 Siguiendo281 Seguidores79 Publicaciones

Bluesky

¿Es tu perfil?Reclámalo y laufer.dirty.social es tuyo.Reclamar perfil

Benjamin Laufer@laufer· 6h

I found that if you’re going to make the horizontal cut, make it low, roughly 20% of the way up the onion. At that height, the cut attacks the few giant pieces that vertical cuts tend to leave behind. In the simulations, it can dramatically reduce the range of piece sizes.

Intuition showing that not doing a horizontal cut leaves giant pieces of onion.

Benjamin Laufer@laufer· 6h

As a researcher, every day I think about questions I don’t have the answer to. It used to be that most of them just weren’t worth answering because they’d take too much time and resources. I think AI is changing that. Here’s one example: what is the best way to cut an onion?

Diagram showing the model of an onion as concentric layers, and showing intuition for horizontally cutting the half-onion at about 20% of its height.

Benjamin Laufer@laufer· 8d

I am delighted to share that I'll be joining the University of Washington Information School as an Assistant Professor!

Benjamin Laufer@laufer· 22d

One intuition behind many AI policy proposals is that downstream AI applications -- the companies deploying AI in healthcare, finance, education, customer service, etc. -- should bear responsibility for ensuring safety. Our paper asks: What incentives does that create for the firms building AI?

Illustrative example of our game-theoretic model. This numeric instance of the game consists of one general-purpose producer and three domain-specialists. Each player has a different utility in performance-safety space which dictates the path of development. Within this setting, the no-regulation game (Upper Left) reveals the players’ investment efforts when no floor is imposed on safety. Regulating the domain-specialist alone (Upper Right) exhibits backfiring for all three domains, meaning the regulated safety level is lower than it would be without regulation. In this particular example, the same floor is assumed for all three domain-specialists. Regulating the generalist alone (Lower Left) improves the safety level slightly across all three domains, compared to no-regulation. Finally, a regime that targets both generalists and specialists with regulation (Lower Right) is able to 1) retain the improved safety performance from regulating the generalist, 2) improve the safety level of least-safe domain-specialist, while 3) avoiding backfiring. The purpose of this figure is to visualize the model’s incentive mechanisms; none of these panels represent real regulations.

Benjamin Laufer@laufer· Aug 14

We describe these details in the paper, e.g. in the schematic below.

A schematic showing the different model relations in a given family (or connected component) in the Hugging Face graph.

Benjamin Laufer@laufer· Aug 14

Big picture: By treating ML models like organisms in an ecosystem, we can: 🌱 Understand the pressures shaping AI development 🔍 Spot patterns before they become industry norms 🛠 Inform governance & safety strategies grounded in real data

Benjamin Laufer@laufer· Aug 14

We found optimal evolutionary orderings over traits: 🔹 Feature extraction tends to be upstream from text generation. Text generation is upstream from text classification. 🔹 Certain license types precede others (e.g., llama3 → apache-2.0) Here we show the top-20 licenses transitions over fine-tunes.

A directed graph representing the typical direction of drift. For example, a directed edge pointing from gemma to mit suggests that more models mutate from gemma to mit than from mit to gemma. Color coding suggests that the general trend is commercial licenses mutate to other types of licenses. Of all creative commons licenses, the two most permissive varieties are downstream from those with more terms and restrictions.

Benjamin Laufer@laufer· Aug 14

In biology, traits get passed from parent to child — mutations are slow & often modeled as random. In AI model families, mutations are fast and directed. Two sibling models tend to resemble each other more than they resemble their shared parent.

Some of our reported measurements showing that two related models have considerable more similarity in their data than any two (not-necessarily-related) models. Similarities are measured for pairs of models belonging to different subtree topologies. Interestingly, pairs of siblings are more related than parent-child pairs, even though they're further in distance over the tree. A diagram attempts to explain this as fast, directed mutation of traits.

Benjamin Laufer@laufer· Aug 14

We measured “genetic similarity” between models from snippets of text - the metadata and model cards. Models in the same finetuning family do resemble each other… but the evolution is weird. For example, traits drift in the same directions again and again.

The "diff" in model data between two snippets of metadata, one (top) belonging to a base model and the other (bottom) belonging to one of its finetunes. Insertions, deletions and substitutions are shown in green, red and yellow respectively. The figure is supposed to evoke DNA sequencing.

Benjamin Laufer@laufer· Aug 14

In a new paper with @didaoh and Jon Kleinberg, we mapped the family trees of 1.86 million AI models on Hugging Face — the largest open-model ecosystem in the world. AI evolution looks kind of like biology, but with some strange twists. 🧬🤖