For anyone who hasn’t seen this before, mechanistic interpretability solves a very common problem with LLMs: when you ask a model to explain itself, you’re playing a game of rhetoric where the model tries to “convince” you of a reason for what it did by generating a plausible-sounding answer based on patterns in its training data. But unlike most trends of benchmark numbers getting better as models improve, more powerful models often score worse on tests designed to self-detect “untruthfulness” because they have stronger rhetoric, and are therefore more compelling at justifying lies after the fact. The objective is coherence, not truth.

Rhetoric isn’t reasoning. True explainability, like what overfitted Sparse Autoencoders claim they offer, basically results in the causal sequence of “thoughts” the model went through as it produces an answer. It’s the same way you may have a bunch of ephemeral thoughts in different directions while you think about anything.

I want to point out here that people do the same: a lot of the time we don't know why we thought or did something, but we'll confabulate plausible-sounding rhetoric after the fact.

/Some/ people bullshit themselves stating the plausible; others check their hypotheses.

The difference is total in both humans and automated processes.

How are you going to check your hypotheses for why you preferred that jacket to that other jacket?

Do not lose the original point: some systems have a goal to sound plausible, while some have a goal to say the truth. Some systems, when asked "where have you been", will reply "at the baker's" because it is a nice narrative in their "novel writing, re-writing of reality", some other will check memory and say "at the butcher's", where they have actually been.

When people invent explicit reasons on why they turned left or right, those reasons remain hypotheses. The clumsy will promote those hypotheses to beliefs. The apt will keep the spontaneous ideas as hypotheses, until the ability to assess them comes.

Is that example representative for the LLM tasks for which we seek explainability ?

Are we holding LLMs to a higher standard than people?

Ideally yes, LLMs are tools that we expect to work, people are inherently fallible and (even unintentionally) deceptive. LLMs being human-like in this specific way is not desirable.

Then I think you'll be very disappointed. LLMs aren't in the same category as calculators, for example.

I have no illusions on LLMs, I have been working with them since og BERT, always with these same issues and more. I'm just stating what would be needed in the future to make them reliably useful outside of creative writing & (human-guided & checked) search.

If an LLM provides an incorrect/orthogonal rhetoric without a way to reliably fix/debug it it's just not as useful as it theoretically could be given the data contained in the parameters.

Not in math.

Yes in math. Formalisms come after casual thoughts, at every step.

It's totally different: those formalisms are in a workbench, following a set of rules that either work or not.

So, yes, that (math) is representative of the actual process: pattern recognition gives you spontaneous ideas, that you assess for truthfulness in conscious acts of verification.

What is a casual thought that you cannot explain in math?

That question makes no sense. You can explain anything in math, because math is a language and lets you define whatever terms and axioms you need at a given moment.

(Whether or not such explanation is useful for anything is another issue entirely.)

Can you explain how intuition led you to try a certain approach?

Is it enough if I hand-wave it with probability distributions, or do you want me to write out adjacency search in a high-dimensional space?

The split-brain experiment is one of my favorites! https://www.youtube.com/watch?v=wfYbgdo8e-8

A lot of the mech interp stuff has seemed to me like a different kind of voodoo: the Integer Quantum Hall Effect? Overloading the term “Superposition” in a weird analogy not governed by serious group representation theory and some clear symmetry? You guys are reaching. And I’ve read all the papers. Spot the postdoc who decided to get paid.

But there is one thing in particular that I’ll acknowledge as a great insight and the beginnings of a very plausible research agenda: bounded near orthogonal vector spaces are wildly counterintuitive in high dimensions and there are existing results around it that create scope for rigor [1].

[1] https://en.m.wikipedia.org/wiki/Johnson%E2%80%93Lindenstraus...

Superposition code is a well known concept in information theory - I think there is certainly more to the story then described in the current works, but it does feel like they are going in the right direction

Where are you seeing the integer quantum Hall effect mentioned? Or are you bringing it up rather than responding to it being brought up elsewhere? I don’t understand what the connection between IQHE and these SAE interpretability approaches is supposed to be.

Pardon me, the reference is to the fractional Hall effect.

"But our results may also be of broader interest. We find preliminary evidence that superposition may be linked to adversarial examples and grokking, and might also suggest a theory for the performance of mixture of experts models. More broadly, the toy model we investigate has unexpectedly rich structure, exhibiting phase changes, a geometric structure based on uniform polytopes, "energy level"-like jumps during training, and a phenomenon which is qualitatively similar to the fractional quantum Hall effect in physics, among other striking phenomena. We originally investigated the subject to gain understanding of cleanly-interpretable neurons in larger models, but we've found these toy models to be surprisingly interesting in their own right."

https://transformer-circuits.pub/2022/toy_model/index.html

A{rt,I} imitating life

I believe that's why humans reason too. We make snap judgements and then use reason to try to convince others of our beliefs. Can't recall the reference right now but they argued that it's really a tool for social influence. That also explains why people who are good at it find it hard to admit when they are wrong - they're not used to having to do it because they can usually out argue others. Prominent examples are easy to find - X marks de spot.

I wonder if the reference you are reaching for, if it's not the Jonathan Haidt book suggested by a sibling comment, is The Enigma of Reason by the cognitive psychologists Hugo Mercier and Dan Sperber (2017).

In that book (quoting here from the abstract), Mercier and Sperber argue that reason 'is not geared to solitary use, to arriving at better beliefs and decisions on our own', but rather to 'help us justify our beliefs and actions to others, convince them through argumentation, and evaluate the justifications and arguments that others address to us'. Reason, they suggest, 'helps humans better exploit their uniquely rich social environment'.

They resist the idea (popularized by Daniel Kahneman) that there is 'a contrast between intuition and reasoning as if these were two quite different forms of inference', proposing instead that 'reasoning is itself a kind of intuitive inference'. For them, reason as a cognitive mechanism is 'much more opportunistic and eclectic' than is implied by the common association with formal systems like logic. 'The main role of logic in reasoning, we suggest, may well be a rhetorical one: logic helps simplify and schematize intuitive arguments, highlighting and often exaggerating their force.'

Their 'interactionist' perspective helps explain how illogical rhetoric can be so socially powerful; it is reason, 'a cognitive mechanism aimed at justifying oneself and convincing others', fulfilling its evolutionary social function.

Highly recommended, if you're not already familiar.

Already before Galileo we had experiments to determine whether ideas represented reality or not. And in crucial cases, long before that, it meant life or death. This will be clear to engineers.

«Reason» is part of that mechanism of vetting ideas. You experience massive failures without it.

So, no, trained judgement is a real thing, and the presence of innumerable incompetent do not prove an alleged absence of the competent.

I think Robert Sapolsky's lectures on yt cover this to some degree around 115.

https://youtu.be/wLE71i4JJiM?feature=shared

Sometimes our cortex is in charge, sometimes other parts of our brain are, and we can't tell the difference. Regardless, if we try to justify it later, that justification isn't always coherent because we're not always using the part of our brain we consider to be rational.

People who are good at reasoning find it hard to admit that they were wrong?

That’s not my experience. People with reason are.. reasonable.

You mention X and that’s not where the reasoners are. That’s where the (wanna be) politicians are. Rhetoric is not all of reasoning.

I can agree that rationalizing snap judgements is one of our capabilities but I am totally unconvinced that it is the totality of our reasoning capabilities. Perhaps I misunderstood.

This is not totally my experience, I've debated a successful engineer who by all accounts has good reasoning skills, but he will absolutely double down on unreasonable ideas he's made on the fly he if can find what he considers a coherent argument behind them. Sometimes if I absolutely can prove him wrong he'll change his mind.

But I think this is ego getting in the way, and our reluctance to change our minds.

We like to point to artificial intelligence and explain how it works differently and then say therefore it's not "true reasoning". I'm not sure that's a good conclusion. We should look at the output and decide. As flawed as it is, I think it's rather impressive

> ego getting in the way

That thing which was in fact identified thousands of years ago as the evil to ditch.

> reluctance to change our minds

That is clumsiness in a general drive that makes sense and is recognized part of the Belief Change Theory: epistemic change is conservative. I.e., when you revise a body of knowledge you do not want to lose valid notions. But conversely, you do not want to be unable to see change or errors, so there is a balance.

> it's not "true reasoning"

If it shows not to explicitly check its "spontaneous" ideas, then it is a correct formula to say 'it's not "true reasoning"'.

Jonathan Haidt's The Righteous Mind describes this ín details.

How does the causality part work? Can it spit out a graphical model?

Interpretability research is basically a projection of the original function implemented by the neural network onto a sub-space of "explanatory" functions that people consider to be more understandable. You're right that the words they use to sell the research is completely nonsensical because the abstract process has nothing to do with anything causal.

All code is causal.

Which makes it entirely irrelevant as a descriptive term.

"Servers shall be strict in formulation and flexible in interpretation."

Incredible, well-documented work -- this is an amazing effort!

Two things that caught my eye were (i) your loss curves and (ii) the assessment of dead latents. Our team also studied SAEs -- trained to reconstruct dense embeddings of paper abstracts rather than individual tokens [1]. We observed a power-law scaling of the lower bound of loss curves, even when we varied the sparsity level and the dimensionality of the SAE latent space. We also were able to totally mitigate dead latents with an auxiliary loss, and we saw smooth sinusoidal patterns throughout training iterations. Not sure if these were due to the specific application we performed (over paper abstracts embeddings) or if they represent more general phenomena.

[1] https://arxiv.org/abs/2408.00657

I'm very happy you appreciate it - particularly the documentation. Writing the documentation was much harder for me than writing the code so I'm happy it is appreciated. I furthermore downloaded your paper and will read through it tomorrow morning - thank you for sharing it!

This seems like decent alignment-positive work on a glance, though I haven't checked full details yet. I probably can't make it happen, but how much would someone need to pay you to make up your time, expense, and risk?

Hey - Thanks for sharing!

Will take a closer look later but if you are hanging around now, it might be worth asking this now. I read this blog post recently:

https://adamkarvonen.github.io/machine_learning/2024/06/11/s...

And the author talks about challenges with evaluating SAEs. I wonder how you tackled that and where to look inside your repo for understanding the your approach around that if possible.

Thanks again!

So evaluating SAEs - determining which SAE is better at creating the most unique features while being as sparse as possible at the same time - is a very complex topic that is very much at the heart of the current research into LLM interpretability through SAEs.

Assuming you already solved the problem of finding multiple perfect SAE architectures and you trained them to perfection (very much an interesting ML engineering problem that this SAE project attempts to solve) then deciding on which SAE is better comes down to which SAE performs better on the metrics of your automated interpretability methodology. Particularly OpenAI's methodology emphasizes this automated interpretability at scale utilizing a lot of technical metrics upon which the SAEs can be scored _and thereby evaluated_.

Since determining the best metrics and methodology is such an open research question that I could've experimented on for a few additional months, have I instead opted for a simple approach in this first release. I am talking about my and OpenAI's methodology and the differences between both in chapter 4. Interpretability Analysis [1] in my Implementation Details & Results section. I can also recommend reading the OpenAI paper directly or visiting Anthropics transformer-circuits.pub website that often publishes smaller blog posts on exactly this topic.

[1] https://github.com/PaulPauls/llama3_interpretability_sae#4-i... [2] https://transformer-circuits.pub/

Thanks!

Find a latent for the Golden Gate bridge and put a Golden Gate Llama 3.2 on HuggingFace. This will get even more attention and love, more so if you include link to a space to chat with it!

Also, you didn't ask for suggestions but putting some interesting results / visualizations at the top of the README is a very good idea.

I’m really excited to see some more open SAE work! The engineering effort is non trivial and I’m going to check out your dataloading code tomorrow. You might be interested in an currently in-progress project of mine to train SAEs on vision models: https://github.com/samuelstevens/saev

Coming from a layman's perspective, a genuine question regarding: "Implements SAE training with auxiliary loss to prevent and revive dead latents, and gradient projection to stabilize training dynamics".

I struggle to understand this phrase "to prevent and revive ", perhaps this is simple speak to those that understand the subject of SAEs, but it feels a bit self contradictory to me, could anyone elaborate?

A latent that is never active and hence doesn't (seem to) represent anything. A loss term to reduce the occurrence of that, and if it does happen, push it back to being active sometimes.

So basically preventing dead latents from occurring and whenever they do occur to possibly reviving them through the use of auxiliary loss term in the loss function? Thanks btw

This is great work! Mechanistic interpretability has tons of use cases, it's great to see open research in that field.

You mentioned you spent your own time and money on it, would you be willing to share how much you spent? It would help others who might be considering independent research.

Thank you for posting this PaulPauls,

can I please ask a wacky question that I have about mech.interp. ?

we know that when we use a metric to measure humans, we game the metric, I wonder about future ai, gaming mech.interp.

for simplicity let me propose a 2d matrix to encode tokens, and hope my understanding of neural networks is good enough for this to be a valid question I understand that higher dimensions have a lot more 'space', so accidental collisions are easily avoided, but 2d makes my example easier to explain.

if we had the following tokens mapped in 2d space

  Apple 1a
  Pear  1b
  Donkey 2a
  Horse 2b

it becomes impossible to understand if the neurons 1,2,a,b, all excited represents apple and horse or donkey and pear?

I wonder if a future, vastly more competent AI overseeing its own training might use this form of ambiguity as means to create a model in which the weights are selected so the potential for encoding collisions exist, so that it is possible to deceive an mech.int. observer as to the true thoughts of the model, essentially enabling the ai to think in euphemisms?

This is awesome! I really appreciate the time you took to document everything!

Thank you for saying that! I have a much, much harder time documenting everything and writing out each decision in continuous text than actually writing the code. So it took a look time for me to write all of this down - so I'm happy you appreciate it! =)

Very cool work! Any plans to integrate it with SAELens?

Not sure yet to be honest. I'll definitely consider it but I'll reorient myself and what I plan to do next in the coming week. I also planned on maybe starting a simpler project and maybe showing people how to create the full model of a current Llama 3.2 implementation from scratch in pure PyTorch. I love building things from teh ground up and when I looked for documentation for the Llama 3.2 background section of this SAE project then the existing documentation I found was either too superficial or outdated and intended for Llama 1 or 2 - Documentation in ML gets outdated so quickly nowadays...

This is amazing, the documentation is very well organized

Fantastic work! I absolutely love all the documentation.

This is incredible!

PaulPauls, how would you like us to cite your work?

Amazing research!!