Comment by x312
3 days ago
A TLDR, since most comments here seem to misunderstand the result.
The idea is: a model may only have 100M neurons or whatever. But there are infinite concepts in the world. Therefore, models have to "stuff" multiple concepts per neuron.
How do you do that? Theres two ways:
1. Each neuron handles similar concepts. One neuron for dogs, a separate neuron one for cars, etc.
2. Each neuron handles a bundle of unrelated concepts. One neuron handles dogs and cars. Another neuron handles cats and numbers. And so on.
Historically, people thought #2 was correct because many neurons were found that seemingly represented multiple concepts. #2 had a cool name (polysemanticity) and was promoted in OpenAI/Anthropic blogs so it had a lot of traction.
However, this blog says #2 is wrong. Actually, the dog/car neuron is just a "V shape neuron". So #2 was just methodological error, and #1 was right all along. This is all the blog is saying.
The way he describes the model of neural nets almost reminds me of a decision tree, with the positive/negative poles and stuff. There's another quote from the article about the "textbook" intuition of convnets
>neurons generalize, work together, and build abstractions over layers. Somewhere along the way we overcomplicated that picture, but looking harder brought it back.
and it does seem that if what the article claims is true, then it really is that simple. To clssify e.g. fruit they first detect curves/shape, then the neurons downstream of say the "circle" detector may distinguish based on color, then the ones downstream of a specific color distinguish by texture, and so on. By the end, what you have are neurons that activate only on very specific combination of features, which map to each fruit.
I like semantic space because it is all continuous. As I spend more time using embeddings I sometimes think in semantic space.
A father is continuous opposed to boolean. 'Father' would rank highly in semantic similarity to a man who has a child, but father-like is better, easily someone is more or less father-like. A woman without a child has a certain value of 'father'.
But really the direction 'father-like' is not an intrinsic feature of semantic space, it is a human construction. Much the same as it would be arbitrary to measure 'v shape'.
I don't think that the latent space has a property or point 'v shape'. My view is that there are no concepts or neurons there. Discrete abstractions are not intrinsic properties of the model. Actually I think the abstractions of concepts or neurons do not map onto anything.
A nice example from cot itself is pareidolia. That is, seeing faces in clouds, or jesus on a piece of toast.
A comment which was promised! Thank you sir!
Original article desperately needs an editor. Like a real one, not Claude or whatever. Your tldr is 100x more understandable
Thank you!
Perhaps you can also help me understand why #1 being right and #2 being wrong would be some win for alignment that people on X claim it to be?
My impression, being a total outsider, is that the reaction feels like a bunch of insider bros hyping up their insider bro friend. It feels more like PR than a genuine groundswell behind a truly revolutionary idea.
The big win would be, "Here's the deception neuron, here's the sycophancy neuron, here's the prioritize my own goals over the goals of the user neuron, here's the seek control over my environment neuron."
How much would it cost to find those things, if they exist? For using the blog post's methods to label every neuron in modern arch, Claude and Gemini produce similar back of the envelope math that looks something like, within an order of magnitude of the resources it takes to pretrain the model in the first place. If you have a better way of narrowing down the set of candidates, probably a lot cheaper.
Would that be reliable and useful? Well.
Because if each neuron represents a small set of similar concepts instead of several completely different things, it should be easier than expected to look inside the neural net and see what it's really thinking.
It might still be quite difficult since the network is huge, and just because you know what the individual neurons represent doesn't mean you understand the complicated thoughts they're building. But easier than expected is still good news.
If (2) is correct then you need to train SAEs and other things to figure out what a neural net is thinking. This is very expensive and slow, which is why a lot of neuron-interpretability experiments are done on small networks. The techniques don't scale.
But if (1) is correct, and Nick presents strong arguments that it is, then you can toss out all the indirect stuff like training dedicated interpretation networks. So the techniques can scale and be trustworthy.
Neurons having narrow semantics means you can inspect what a neural network is "thinking" with confidence and even constrain its space of behaviors by activating/suppressing a small set of neurons that have very straightforward effects on its behavior. This is in contrast to polysemanticity where reading the state of any number of neurons may not give insight into what state the neural network is in and controlling the network by activating/suppressing neurons is nearly impossible to do reliably.
> Neurons having narrow semantics means you can inspect what a neural network is "thinking" with confidence
Not really, because of emergent behaviours and also because there are still millions of neurons being summed (with variable weighting depending on inputs) at each layer.
> and even constrain its space of behaviors by activating/suppressing a small set of neurons that have very straightforward effects on its behavior.
You can influence behaviour by doing this kind of thing, yes, but it isn't straightforward at all, and is generally very messy and you get all sorts of unpleasant sides effects (e.g. abliterated models not performing as well).
6 replies →