Nora Belrose

why don't more people become zoroastrian?


it's where judaism and christianity got the idea of ethical monotheism, afterlife, and final judgment but without any of their baggage


(no eternal hell, no historically questionable dogmas, etc.)

If we care only about appearances, outcomes, and results then AI will outcompete humans at everything


If we care about the process used to create things then humans can still have jobs and meaningful lives


The idea that ends can be detached from means is the root of many evils

Strongly agree with this bill https://www.usatoday.com/story/news/politics/2025/09/29/ohio-state-legislator-ban-people-marrying-ai/86427987007/

if the laws of physics are fundamentally probabilistic, as they seem to be, that makes it easier to see how they can smoothly change over time

data attribution is a special case of data causality:


estimating the causal effect of either learning or unlearning one datapoint (or set of datapoints) on the neural network's behavior on other datapoints

Neural networks don't have organs.


They aren't made of fixed mechanisms.


They have flows of information and intensities of neural activity. They can't be organized into a set of parts with fixed functions.


In the words of Gilles Deleuze, they're bodies without organs (BwO).

Mixture-of-Depths: Dynamically allocating compute in... Transformer-based language models spread FLOPs uniformly across input sequences. In this work we demonstrate that transformers can instead learn to dynamically allocate FLOPs (or compute) to...

This seems like a cool way to use an adaptive amount of compute per token. I speculate that models like these will have more faithful CoT since they don't get to do "extra" reasoning on easy tokens https://arxiv.org/abs/2404.02258

Also chapter 10 where he discards the notion of the Soul but maintains the distinction between mind and brain

William James did a lot of good philosophy of mind in chapters 1, 5, and 6 ofThe Principles of Psychology, we've barely made any progress in 135 years 😂

I love this meme

might interest @nabla_theta

Pro tip: if you want to implement TopK SAEs efficiently, and don't want to deal with Triton, just use this function for the decoder, it's much faster than the naive dense matmul implementation
https://pytorch.org/docs/stable/generated/torch.nn.functional.embedding_bag.html

And here's the code we used to generate the results: github.com/EleutherAI/b...

Estimating the Probability of Sampling a Trained Neural Network at Random We present an algorithm for estimating the probability mass, under a Gaussian or uniform prior, of a region in neural network parameter space corresponding to a particular behavior, such as achieving ...

Here's the paper link: arxiv.org/abs/2501.18812

Second, we speculate that complexity measures like this be useful for detecting undesired "extra reasoning" in deep nets. We want networks to be aligned with our values instinctively, without scheming about whether this would be consistent with some ulterior motive arxiv.org/abs/2311.08379

We're interested in this line of work for two reasons:

First, it sheds light on how deep learning works. The "volume hypothesis" says DL is similar to randomly sampling a network from weight space that gets low training loss. But this can't be tested if we can't measure volume.

We find that the probability of sampling a network at random— or local volume for short— decreases exponentially as the network is trained.

And networks which memorize their training data without generalizing have lower local volume— higher complexity— than generalizing ones.

But the total volume can be strongly influenced by a small number of outlier directions, which are hard to sample in high dimension— think of a big, flat pancake.

Importance sampling using gradient info helps address this issue by making us more likely to sample outliers.

It works by exploring random directions in weight space, starting from an "anchor" network.

The distance from the anchor to the edge of the region, along the random direction, gives us an estimate of how big (or how probable) the region is as a whole.

My colleague Adam Scherlis and I developed a method for estimating the probability of sampling a neural network in a behaviorally-defined region from a Gaussian or uniform prior.

You can think of this as a measure of complexity: less probable, means more complex.

What are the chances you'd get a fully functional language model by randomly guessing the weights?

We crunched the numbers and here's the answer:

we have seven (!) papers lined up for release next week


you know you're on a roll when arxiv throttles you

deepseek now largely replacing chatgpt for me

Evolutionary biology can learn things from machine learning.


Natural selection alone doesn't explain "train-test" or "sim-to-real" generalization, which clearly happens.


At every level of organization, life can zero-shot adapt to novel situations. https://www.youtube.com/watch?v=jJ9O5H2AlWg

Truth is relative, when it comes to the physical state of the universe.


But we should accept the existence of perspective-neutral facts about how perspectives relate to one another, to avoid vicious skeptical paradoxes. https://arxiv.org/abs/2410.13819

There's Plenty of Room Right Here: Biological Systems as Evolved, Overloaded, Multi-scale Machines The applicability of computational models to the biological world is an active topic of debate. We argue that a useful path forward results from abandoning hard boundaries between categories and adopt...

Neural networks are polycomputers in
@drmichaellevin.bsky.social's sense.

Depending on your perspective, you can interpret them as performing many different computations on different types of features. No perspective is uniquely correct. arxiv.org/abs/2212.10675

If OpenAI's new o3 model is "successfully aligned," then it could probably be trusted to supervise more powerful models, allowing us to bootstrap to benevolent superintelligence.

My Week Without Cosmic Hope (Photo by Tom Pumford on Unsplash)

Interesting to see @philipgoff.bsky.social go back and forth on the fine-tuning argument.

I think the multiverse definitely can't explain fine-tuning, but it's also unclear we need an explanation at all. And God may be a more "complex" hypothesis than the physical constants themselves.

GitHub - EleutherAI/training-jacobian Contribute to EleutherAI/training-jacobian development by creating an account on GitHub.

This is the first in a series of upcoming papers on neural network training dynamics and loss landscape geometry. Please check out the interp-across-time channel in the
@eleutherai.bsky.social Discord if you'd like to get more involved.

Code: github.com/EleutherAI/t...

Unfortunately, computing the entire training Jacobian and performing SVD on it is computationally intractable for all but the smallest networks.

We focused on a tiny 5K parameter MLP for most experiments, but we did find a similar SV spectrum in a 62K param image classifier.