Hacker News Reader

Comment by jablongo | Hacker News Reader

This isnt exactly building an understanding of LLMs from first principles... IMO we should broadly be following the (imperfect) example set forth by neuroscientists attempting to explain fMRI scans and assigning functionality to various subregions in the brain. It is circular and "unsafe" from an alignment perspective to use a complex model to understand the internals of a simpler model; in order to understand GPT4 then we need GPT5? These approaches are interesting, but we should primarily focus on building our understanding of these models from building blocks that we already understand.

View on HN

I've been working in systems neuroscience for a few years (something of a combination lab tech/student, so full disclosure, not an actual expert).

Based on my experience with model organisms (flies & rats, primarily), it is actually pretty amazing how analogous the techniques and goals used in this sort of research are to those we use in systems neuroscience. At a very basic level, the primary task of correlating neuron activation to a given behavior is exactly the same. However, ML researchers benefit from data being trivial to generate and entire brains being analyzable in one shot as a result, whereas in animal research elucidating the role of neurons in a single circuit costs millions of dollars and many researcher-years.

The similarities between the two are so clear that I noticed that in its Microscope tool [1], OpenAI even refers to the models they are studying as "model organisms", an anthropomorphization which I find very apt. Another article I saw a while back on HN which I thought was very cool was [2], which describes the task of identifying the role of a neuron responsible for a particular token of output. This one is especially analogous because it operates on such a small scale, much closer to what systems neuroscientists studying model organisms do.

[1] https://openai.com/research/microscope [2] https://clementneo.com/posts/2023/02/11/we-found-an-neuron

Yup, this article on predictive coding for example is particularly interesting.

Lots of parallels to how our brains are thought to work.

https://en.m.wikipedia.org/wiki/Predictive_coding

I don't follow. Neuroscience imaging tools like fMRI are only used because it is impossible to measure the activations of each neuron in a brain in real time (unlike an artificial neural network). This research paper's attempt to understand the role of individual neurons or neuron clusters within a complete network gets much closer to "first principles" than fMRI.

Right so it should be much easier w/ access to every neuron and activation. But the general approach is an experimental one where you try to use your existing knowledge about physics and biology to discern what is activating different structures (and neurons) in the brain. I agree w/ the approach of trying to assign some functionality to individual 'neurons', but I don't think that using GPT4 to do so is the most appealing way to go about that, considering GPT4 is the structure we are interested in decoding in the first place.

All of this seems to lead to something like this paper https://journals.plos.org/ploscompbiol/article?id=10.1371/jo...

On the other hand, I find it plausible that it's fundamentally impossible to assign some functionality to individual 'neurons' due to the following argument:

1. Let's assume that for a system calculating a specific function, there is a NN configuration (weights) so that at some fully connected NN layer there is a well-defined functionality for specific individual neurons - #1 represents A, #2 represents B, #3 represents C etc.

2. The exact same system outcome can be represented with infinitely many other weight combinations which effectively result in a linear transformation (i.e. every possible linear transformation) of the data vector at this layer, e.g. where #1 represents 0.1A + 0.3B + 0.6C, #2 represents 0.5B+0.5C, and #3 represents 0.4B+0.6C - in which case the functionality A (or B, or C) is not represented by any individual neurons;

3. When the system is trained, it's simply not likely that we just happen to get the best-case configuration where the theoretically separable functionality is actually separated among individual 'neurons'.

Biological minds do get this separation because each connection has a metabolic cost; but the way we train our models (both older perceptron-like layers, and modern transfomer/attention ones) do allow linking everything to everything, so the natural outcome is that functionality simply does not get cleanly split out in individual 'neurons' and each 'neuron' tends to represent some mix of multiple functionalities.

Your last idea, that these models’ neurons are all connected in some way, makes me somewhat sceptical of this research by OpenAI. And that their technique of analysis may need to be more fractal or expansive to include groups of neurons, moving all the way up to the entire model.

> in order to understand GPT4 then we need GPT5?

I also found this amusing. But you are loosely correct, AFAIK. GPT-4 cannot reliably explain itself in any context: say the total number of possible distinct states of GPT-4 is N; then the total number of possible distinct states of GPT-4 PLUS any context in which GPT-4 is active must be at least N + 1. So there are at least two distinct states in this scenario that GPT-4 can encounter that will necessarily appear indistinguishable to GPT-4. It doesn't matter how big the network is; it'll still encounter this limit.

And it's actually much worse than that limit because a network that's actually useful for anything has to be trained on things besides predicting itself. Notably, this is GPT-4 trying to predict GPT-2 and struggling:

> We found over 1,000 neurons with explanations that scored at least 0.8, meaning that according to GPT-4 they account for most of the neuron’s top-activating behavior. Most of these well-explained neurons are not very interesting. However, we also found many interesting neurons that GPT-4 didn't understand. We hope as explanations improve we may be able to rapidly uncover interesting qualitative understanding of model computations.

1,000 neurons out of 307,200--and even for the highest-scoring neurons, these are still partial explanations.

There's little reason to think that predicting GPT-4 would be more difficult, only that it would be far more computationally expensive (given the higher number of neurons and much higher computational cost of every test).

Reply on news.ycombinator.com