All of Logan Riggs's Comments + Replies

I agree. There is a tradeoff here for the L0/MSE curve & circuit-simplicity.

I guess another problem (w/ SAEs in general) is optimizing for L0 leads to feature absorption. However, I'm unsure of a metric (other than the L0/MSE) that does capture what we want.

Hey Armaan! Here's a paper where they instead used an MLP in the beginning w/ similar results (looking at your code, it seems by "dense" layer, you also mean a nonlinearity, which seems equivalent to the MLP one?)

How many tokens did you train yours on?

Have you tried ablations on the dense layers, such as only having the input one vs the output one? I know you have some tied embeddings for both, but I'm unsure if the better results are for the output or input. 

For both of these, it does complicate circuits because you'll have features combining nonline... (read more)

3Armaan A. Abraham
Ah, I was unaware of that paper and it is indeed relevant to this, thank you! Yes, by "dense" or "non-sparse" layer, I mean a nonlinearity. So, that paper's MLP SAE is similar to what I do here, except it is missing MLPs in the decoder. Early on, I experimented with such an architecture with encoder-only MLPs, because (1) as to your final point, the lack of nonlinearity in the output potentially helps it fit into other analyses and (2) it seemed much more likely to me to exhibit monosemantic features than an SAE with MLPs in the decoder too. But, after seeing some evidence that its dead neuron problems reacted differently to model ablations than both the shallow SAE and the deep SAE with encoder+decoder MLPs, I decided to temporarily drop it. I figured that if I found that the encoder+decoder MLP SAE features were interpretable, this would be a more surprising/interesting result than the encoder-only MLP SAE and I would run with it, and if not, I would move to the encoder-only MLP SAE. I trained on 7.5e9 tokens. As I mentioned in my response to your first question, I did experiment early on with the encoder-only MLP, but the architectures in this post are the only ones I looked at in depth for GPT2.   This is a good point, and I should have probably included this in the original post. As you said, one of the major limitations of this approach is that the added nonlinearities obscure the relationship between deep SAE features and upstream / downstream mechanisms of the model. In the scenario where adding more layers to SAEs is actually useful, I think we would be giving up on this microscopic analysis, but also that this might be okay. For example, we can still examine where features activate and generate/verify human explanations for them. And the idea is that the extra layers would produce features that are increasingly meaningful/useful for this type of analysis.

How do you work w/ sleep consolidation?

Sleep consolidation/ "sleeping on it" is when you struggle w/ [learning a piano piece], sleep on it, and then you're suddenly much better at it the next day!

This has happened to me for piano, dance, math concepts, video games, & rock climbing, but it varies in effectiveness. Why? Is it:

  1. Duration of struggling activity
  2. Amount of attention paid to activity
  3. Having a frustrating experience
  4. Time of day (e.g. right before sleep)

My current guess is a mix of all four. But I'm unsure if you [practice piano] in the morning, you... (read more)

1CstineSublime
If I'm playing anagrams or Scrabble after going to a church, and I get the letters "ODG" I'm going to be predisposed towards a different answer than if I've been playing with a German Shepard. I suspect sleep has very little to do with it, and simply coming at something with a fresh load of biases on a different day with different cues and environmental factors may be a larger part of it. Although Marvin Minsky made a good point about the myth of introspection: we are only aware of a think sliver of our active mental processes at any given moment, when you intensely focus on a maths problem or practicing the piano for a protracted period of time, some parts of the brain working on that may not abandon it just because your awareness or your attention drifts somewhere else. This wouldn't just be during sleep, but while you're having a conversation with your friend about the game last night, or cooking dinner, or exercising. You're just not aware of it, it's not in the limelight of your mind, but it still plugs away at it. In my personal experience, most Eureka moments are directly attributable to some irrelevant thing that I recently saw that shifted my framing of the problem much like my anagram example.
3philip_b
I think the way to learn any skill is to basically: 1. Practice it 2. Sleep 3. Goto 1 And the time spent in each iteration of item 1 is capped in usefulness or at least has diminishing returns. I think this has nothing to do with frustration. Also, I think reminding yourself of the experience is not that important and I think there is no cap of 1 thing a day.
2[anonymous]
Oh, I've thought a lot about something similar that I call "background processing" - I think it happens during sleep, but also when awake. I think for me it works better when something is salient to my mind / my mind cares about it. According to this theory, if I was being forced to learn music theory but really wanted to think about video games, I'd get less new ideas about music theory from background processing, and maybe it'd be less entered into my long term memory from sleep. I'm not sure how this effects more 'automatic' ('muscle memory') things (like playing the piano correctly in response to reading sheet music). I'm not sure about this either. It could also be formulated as there being some set amount of consolidation you do each night, and you can divide them between topics, but it's theoretically (disregarding other factors like motivation; not practical advice) most efficient if you do one area per day (because of stuff in the same topic having more potential to relate to each other and be efficiently compressed or generalized from or something. Alternatively, studying multiple different areas in a day could lead to creative generalization between them).

Huh, those brain stimulation methods might actually be practical to use now, thanks for mentioning them! 

Regarding skepticism of survey-data: If you're imagining it's only an end-of-the-retreat survey which asks "did you experience the jhana?", then yeah, I'll be skeptical too. But my understanding is that everyone has several meetings w/ instructors where a not-true-jhana/social-lie wouldn't hold up against scrutiny. 

I can ask during my online retreat w/ them in a couple months.

4niplav
As for brain stimulation, TMS devices can be bought for <$10k from ebay. tDCS devices are available for ~$100, though I don't expect them to have large effect sizes in any direction. There's been noises of consumer-level tFUS devices for <$10k, but that's likely >5 years in the future. The incentives of the people running jhourney are to over-claim attainments, especially on edge-cases, and hype the retreats. Organizations can be sufficiently on guard to prevent the extreme forms of over-claiming & turning into a positive-reviews-factory, but I haven't seen people from jhourney talk about it (or take action that shows they're aware of the problem).

Implications of a Brain Scan Revolution

Suppose we were able to gather large amounts of brain-scans, lets say w/ millions of wearable helmets w/ video and audio as well,[1] then what could we do with that? I'm assuming a similar pre-training stage where models are used to predict next brain-states (possibly also video and audio), and then can be finetuned or prompted for specific purposes.

Jhana helmets

Jhana is a non-addicting high pleasure state. If we can scan people entering this state, we might drastically reduce the time it takes to learn to enter ... (read more)

4niplav
For pleasure/insight helmets you probably need intervention in the form of brain simulation (tDCS, tFUS, tMS). Biofeedback might help but you need to at least know where to steer towards. I'm pretty skeptical of those numbers, all exiting projects I know of don't have a better method of measurement other than surveys and that gets bitten hard by social desirability bias/not wanting to have committed a sunk cost. Seems relevant that jhourney isn't doing much EEG & biofeedback anymore.

it seems unlikely to me that so many talented people went astray

Well, maybe we did go astray, but it's not for any reasons mentioned in this paper!

SAEs were trained on random weights since Anthropic's first SAE paper in 2023:

To assess the effect of dataset correlations on the interpretability of feature activations, we run dictionary learning on a version of our one-layer model with random weights. 28 The resulting features are here, and contain many single-token features (such as "span", "file", ".", and "nature") and some other features firing on seeming

... (read more)

I didn't either, but on reflection it is! 

I did change the post based off your comment, so thanks!

I think the fuller context,

Anthropic has put WAY more effort into safety, way way more effort into making sure there are really high standards for safety and that there isn't going to be danger what these AIs are doing

implies it's just the amount of effort is larger than other companies (which I agree with), and not the Youtuber believing they've solved alignment or are doing enough, see: 

but he's also a realist and is like "AI is going to really potentially fuck up our world"

and

But he's very realistic. There is a lot of bad shit that is going to happ

... (read more)
1dabbing.
Nevermind I didnt think was a requirement

Thinking through it more, Sox2-17 (they changed 17 amino acids from Sox2 gene) was your linked paper's result, and Retro's was a modified version of factors Sox AND KLF. Would be cool if these two results are complementary.

You're right! Thanks
For Mice, up to 77% 

Sox2-17 enhanced episomal OKS MEF reprogramming by a striking 150 times, giving rise to high-quality miPSCs that could generate all-iPSC mice with up to 77% efficiency

For human cells, up to 9%  (if I'm understanding this part correctly).
 

SOX2-17 gave rise to 56 times more TRA1-60+ colonies compared with WT-SOX2: 8.9% versus 0.16% overall reprogramming efficiency.

So seems like you can do wildly different depending on the setting (mice, humans, bovine, etc), and I don't know what the Retro folks were doing, but does make their result less impressive. 

4TsviBT
(Still impressive and interesting of course, just not literally SOTA.)

You're actually right that this is due to meditation for me. AFAIK, it's not a synesthesia-esque though (ie I'm not causing there to be two qualia now), more like the distinction between mental-qualia and bodily-qualia doesn't seem meaningful upon inspection. 

So I believe it's a semantic issue, and I really mean "confusion is qualia you can notice and act on" (though I agree I'm using "bodily" in non-standard ways and should stop when communicating w/ non-meditators).

This is great feedback, thanks! I added another example based off what you said.

For how obvious the first one, at least two folks I asked (not from this community) didn't think it was a baby initially (though one is non-native english and didn't know "2 birds of a feather" and assumed "our company" meant "the singers and their partner"). Neither are parents. 

I did select these because they caused confusion in myself when I heard/saw them years ago, but they were "in the wild" instead of in a post on noticing confusion.

I did want a post I could link [non rationalist friends] to that's a more fun intro to noticing confusion, so more regular members might not benefit!

For those also curious, Yamanaka factors are specific genes that turn specialized cells (e.g. skin, hair) into induced pluripotent stem cells (iPSCs) which can turn into any other type of cell.

This is a big deal because you can generate lots of stem cells to make full organs[1] or reverse aging (maybe? they say you just turn the cell back younger, not all the way to stem cells).

 You can also do better disease modeling/drug testing: if you get skin cells from someone w/ a genetic kidney disease, you can turn those cells into the iPSCs, then i... (read more)

TsviBTΩ3110

According to the article, SOTA was <1% of cells converted into iPSCs

I don't think that's right, see https://www.cell.com/cell-stem-cell/fulltext/S1934-5909(23)00402-2

A trending youtube video w/ 500k views in a day brings up Dario Amodei's Machines of Loving Grace (Timestamp for the quote):
[Note: I had Claude help format, but personally verified the text's faithfulness]

I am an AI optimist. I think our world will be better because of AI. One of the best expressions of that I've seen is this blog post by Dario Amodei, who is the CEO of Anthropic, one of the biggest AI companies. I would really recommend reading this - it's one of the more interesting articles and arguments I have read. He's basically saying AI is going to

... (read more)
9habryka
Ah yes, a great description of Anthropic's safety actions. I don't think anyone serious at Anthropic believes that they "made sure there isn't going to be danger from these AIs are doing". Indeed, many (most?) of their safety people assign double-digits probabilities to catastrophic outcomes from advanced AI system. I do think this was a predictable quite bad consequence of Dario's essay (as well as his other essays which heavily downplay or completely omit any discussion of risks). My guess is it will majorly contribute to reckless racing while giving people a false impression of how good we are doing on actually making things safe.

Hey Midius!

My recommended rationality habit is noticing confusion, by which I mean a specific mental feeling that's usually quick & subtle & easy to ignore.

David Chapman has a more wooey version called Eating Your Shadow, which was very helpful for me since it pointed me towards acknowledging parts of my experience that I was denying due to identity & social reasons (hence the easy to ignore part).

Could you go into more details into what skills these advisers would have or what situations to navigate? 

Because I'm baking in the "superhuman in coding/maths" due to the structure of those tasks, and other tasks can either be improved through:
1. general capabilies
2. Specific task 

And there might be ways to differentially accelarate that capability.

4Chris_Leong
I don't exactly know the most important capabilities yet, but things like, advising on strategic decisions, improving co-ordination and non-manipulative communication seem important.

I really appreciate your post and all the links! This and your other recent posts/comments have really helped make a clearer model of timelines. 

In my experience, most of the general public will verbally agree that AI X-risk is a big deal, but then go about their day (cause reasonably, they have no power). There's no obvious social role/action to do in response to that.

For climate, people understand that they should recycle, not keep the water running, and if there's a way to donate to clean the ocean on a Mr. Beast video, then some will even donate (sadly, none of these are very effective for solving the climate problem though! Gotta avoid that for our case).

Having a clear call-to-action seems rel... (read more)

Claude 3.5 seems to understand the spirit of the law when pursuing a goal X. 

A concern I have is that future training procedures will incentivize more consequential reasoning (because those get higher reward). This might be obvious or foreseeable, but could be missed/ignored under racing pressure or when lab's LLMs are implementing all the details of research.

Thanks! 

I forgot about faithful CoT and definitely think that should be a "Step 0". I'm also concerned here that AGI labs just don't do the reasonable things (ie training for briefness making the CoT more steganographic). 

For Mech-interp, ya, we're currently bottlenecked by:

  1. Finding a good enough unit-of-computation (which would enable most of the higher-guarantee research)
  2. Computing Attention_in--> Attention_out (which Keith got the QK-circuit -> Attention pattern working a while ago, but haven't hooked up w/ the OV-circuit)

This is mostly a "reeling from o3"-post. If anyone is doom/anxiety-reading these posts, well, I've been doing that too! At least, we're in this together:)

From an apparent author on reddit:

[Frontier Math is composed of] 25% T1 = IMO/undergrad style problems, 50% T2 = grad/qualifying exam style porblems, 25% T3 = early researcher problems

The comment was responding to a claim that Terence Tao said he could only solve a small percentage of questions, but Terence was only sent the T3 questions. 

I also have a couple friends that require serious thinking (or being on my toes).  I think it's because they have some model of how something works, and I say something, showing my lack of this model.

Additionally, programming causes this as well (in response to compilation errors, getting nonsense outputs, or runs too long). 

4Nathan Young
Yes, this is one reason I really like forecasting. I forces me to see if my thinking was bad and learn what good thinking looks like.

Was looking up Google Trend lines for chatGPT and noticed a cyclical pattern:

Where the dips are weekends, meaning it's mostly used by people in the workweek. I mostly expect this is students using it for homework. This is substantiated by two other trends:
1. Dips in interest over winter and summer breaks (And Thanksgiving break in above chart)

2. "Humanize AI" which is 

Humanize AI™ is your go-to platform for seamlessly converting AI-generated text into authentic, undetectable, human-like content 

[Although note that overall interest in ChatGPT is W... (read more)

I’d guess that weekend dips come from office workers, since they rarely work on weekends, but students often do homework on weekends.

I was expecting this to include the output of MIRI for this year. Digging into your links we have:

Two Technical Governance Papers:
1. Mechanisms to verify international agreements about AI development
2. What AI evals for preventing catastrophic risks can and cannot do

Four Media pieces of Eliezer regarding AI risk:
1. Semafor piece
2. 1 hr talk w/ Panel 
3. PBS news hour
4. 4 hr video w/ Stephen Wolfram

Is this the full output for the year, or are there less linkable outputs such as engaging w/ policymakers on AI risks?

Hi, I’m part of the communications team at MIRI.

To address the object-level question: no, that’s not MIRI’s full public output for the year (but our public output for the year was quite small; more on that below). The links on the media page and research page are things that we put in the spotlight. We know the current website isn’t great for seeing all of our output, and we have plans to fix this. In the meantime, you can check out our newslettersTGT’s new website, and a forthcoming post with more details about the media stuff we’ve ... (read more)

Donated $100. 

It was mostly due to LW2 that I decided to work on AI safety, actually, so thanks!

I've had the pleasure of interacting w/ the LW team quite a bit and they definitely embody the spirit of actually trying. Best of luck to y'all's endeavors!

I tried a similar experiment w/ Claude 3.5 Sonnet, where I asked it to come up w/ a secret word and in branching paths:
1. Asked directly for the word
2. Played 20 questions, and then guessed the word

In order to see if it does have a consistent it can refer back to.

Branch 1: 

Branch 2:

Which I just thought was funny.

Asking again, telling it about the experiment and how it's important for it to try to give consistent answers, it initially said "telescope" and then gave hints towards a paperclip.

Interesting to see when it flips it answers, though it's a sim... (read more)

It'd be important to cache the karma of all users > 1000 atm, in order to credibly signal you know which generals were part of the nuking/nuked side. Would anyone be willing to do that in the next 2 & 1/2 hours? (ie the earliest we could be nuked)

4Zach Stein-Perlman
The post says generals' names will be published tomorrow.

We could instead  pre-commit to not engage with any nuker's future posts/comments (and at worse comment to encourage others to not engage) until end-of-year.

Or only include nit-picking comments.

5Logan Riggs
It'd be important to cache the karma of all users > 1000 atm, in order to credibly signal you know which generals were part of the nuking/nuked side. Would anyone be willing to do that in the next 2 & 1/2 hours? (ie the earliest we could be nuked)

During WWII, the CIA produced and distributed an entire manual (well worth reading) about how workers could conduct deniable sabotage in the German-occupied territories.
 

(11) General Interference with Organizations and Production 

   (a) Organizations and Conferences

  1. Insist on doing everything through "channels." Never permit short-cuts to be taken in order to expedite decisions. 
  2. Make speeches, talk as frequently as possible and at great length.  Illustrate your points by long anecdotes and accounts of personal experiences. Neve
... (read more)

Could you dig into why you think it's great inter work?

But through gradient descent, shards act upon the neural networks by leaving imprints of themselves, and these imprints have no reason to be concentrated in any one spot of the network (whether activation-space or weight-space). So studying weights and activations is pretty doomed.

This paragraph sounded like you're claiming LLMs do have concepts, but they're not in specific activations or weights, but distributed across them instead. 

But from your comment, you mean that LLMs themselves don't learn the true simple-compressed features of reality, but a ... (read more)

2tailcalled
A true feature of reality get diminished into many small fragments. These fragments birfucate into multiple groups, of which we will consider two groups, A and B. Group A gets collected and analysed by humans into human knowledge, which then again gets diminished into many small fragments, which we will call group C. Group B and group C make impacts on the network. Each fragment in group B and group C produces a shadow in the network, leading to there being many shadows distributed across activation space and weight space. These many shadows form a channel which is highly reflective of the true feature of reality. That allows there to be simple useful ways to connect the LLM to the true feature of reality. However, the simplicity of the feature and its connection is not reflected into a simple representation of the feature within the network; instead the concept works as a result of the many independent shadows making way for it. The true features branch of from the sun (and the earth). Why would you ignore the problem pointed out in footnote 1? It's a pretty important problem.

The one we checked last year was just Pythia-70M, which I don't expect the LLM itself to have a gender feature that generalizes to both pronouns and anisogamy.

But again, the task is next-token prediction. Do you expect e.g. GPT 4 to have learned a gender concept that affects both knowledge about anisogamy and pronouns while trained on next-token prediction?

3tailcalled
I guess to add, if I ask GPT-4o "What is the relationship between gender and anisogamy?", it answers: So clearly there is some kind of information about the relationship between gender and anisogamy within GPT-4o. The point of my post is that it is unlikely to be in the weight space or activation space.
2tailcalled
Next-token prediction, and more generally autoregressive modelling, is precisely the problem. It assumes that the world is such that the past determines the future, whereas really the less-diminished shapes the more-diminished ("the greater determines the lesser"). As I admitted in the post, it's plausible that future models will use different architectures where this is less of a problem.

Sparse autoencoders finds features that correspond to abstract features of words and text. That's not the same as finding features that correspond to reality.

(Base-model) LLMs are trained to minimize prediction error, and SAEs do seem to find features that sparsely predict error, such as a gender feature that, when removed, affects the probability of pronouns. So pragmatically, for the goal of "finding features that explain next-word-prediction", which LLMs are directly trained for, SAEs find good examples![1]

I'm unsure what goal you have in mind for "feat... (read more)

3tailcalled
If you remove the gender feature, does the neural network lose its ability to talk about anisogamy?

Is there code available for this?

I'm mainly interested in the loss fuction. Specifically from footnote 4:

We also need to add a term to capture the interaction effect between the key-features and the query-transcoder bias, but we omit this for simplicity

I'm unsure how this is implemented or the motivation. 

Some MLPs or attention layers may implement a simple linear transformation in addition to actual computation.

@Lucius Bushnaq , why would MLPs compute linear transformations? 

Because two linear transformations can be combined into one linear transformation, why wouldn't downstream MLPs/Attns that rely on this linearly transformed vector just learn the combined function? 

What is the activation name for the resid SAEs? hook_resid_post or hook_resid_pre?

I found https://github.com/ApolloResearch/e2e_sae/blob/main/e2e_sae/scripts/train_tlens_saes/run_train_tlens_saes.py#L220
to suggest _post
but downloading the SAETransformer from wandb shows:
(saes): 
    ModuleDict( (blocks-6-hook_resid_pre): 
        SAE( (encoder): Sequential( (0):...

which suggests _pre. 
 

8Dan Braun
They are indeed all hook_resid_pre. The code you're looking at just lists a set of positions that we are interested in viewing the reconstruction error of during evaluation. In particular, we want to view the reconstruction error at hook_resid_post of every layer, including the final layer (which you can't get from hook_resid_pre).

3. Those who are more able to comprehend and use these models are therefore of a higher agency/utility and higher moral priority than those who cannot. [emphasis mine]

This (along with saying "dignity" implies "moral worth" in Death w/ Dignity post), is confusing to me. Could you give a specific example of how you'd treat differently someone who has more or less moral worth (e.g. give them more money, attention, life-saving help, etc)? 

One thing I could understand from your Death w/ Dignity excerpt is he's definitely implying a metric that scores every... (read more)

2testingthewaters
Thank you for the feed back! I am of course happy for people to copy over the essay > Is this saying that human's goals and options (including options that come to mind) change depending on the environment, so rational choice theory doesn't apply? More or less, yes, or at least that it becomes very hard to apply it in a way that isn't either highly subjective or essentially post-hoc arguing about what you ought to have done (hidden information/hindsight being 20/20) > This is currently all I have time for; however, my current understanding is that there is a common interpretation of Yudowsky's writings/The sequences/LW/etc that leads to an over-reliance on formal systems that will invevitably fail people. I think you had this interpretation (do correct me if I'm wrong!), and this is your "attempt to renegotiate rationalism ".  I've definitely met people who take the more humble/humility/heuristics driven approach which I outline in the essay and still call themselves rationalists. On the other hand, I have also seen a whole lot of people take it as some kind of mystic formula to organise their lives around. I guess my general argument is that rationalism should not be constructed on top of such a formal basis (cf. the section about heuristics not theories in the essay) and then "watered down" to reintroduce ideas of humility or nuance or path-dependence. And in part 2 I argue that the core principles of rationalism as I see them (without the "watering down" of time and life experience) make it easy to fall down certain dangerous pathways.

I think copy-pasting the whole thing will make it more likely to be read! I enjoyed it and will hopefully leave a more substantial comment later.

I've really enjoyed these posts; thanks for cross posting!

1Chipmonk
thanks! There's a lot I don't post on LessWrong because I don't think it matches the vibe. Even this post has gotten some substantial downvotes

Kind of confused on why the KL-only e2e SAE have worse CE than e2e+downstream across dictionary size:
 

This is true for layers 2 & 6. I'm unsure if this means that training for KL directly is harder/unstable, and the intermediate MSE is a useful prior, or if this is a difference in KL vs CE (ie the e2e does in fact do better on KL but worse on CE than e2e+downstream).

7Dan Braun
Here's a wandb report that includes plots for the KL divergence. e2e+downstream indeed performs better for layer 2. So it's possible that intermediate losses might help training a little. But I wouldn't be surprised if better hyperparams eliminated this difference; we put more effort into optimising the SAE_local hyperparams rather than the SAE_e2e and SAE_e2e+ds hyperparams.

I finally checked!

Here is the Jaccard similarity (ie similarity of input-token activations) across seeds

The e2e ones do indeed have a much lower jaccard sim (there normally is a spike at 1.0, but this is removed when you remove features that only activate <10 times). 

I also (mostly) replicated the decoder similarity chart:

And calculated the encoder sim:

[I, again, needed to remove dead features (< 10 activations) to get the graphs here.] 

So yes, I believe the original paper's claim that e2e features learn quite different features across seed... (read more)

And here's the code to convert it to NNsight (Thanks Caden for writing this awhile ago!)

import torch
from transformers import GPT2LMHeadModel
from transformer_lens import HookedTransformer
from nnsight.models.UnifiedTransformer import UnifiedTransformer


model = GPT2LMHeadModel.from_pretrained("apollo-research/gpt2_noLN").to("cpu")

# Undo my hacky LayerNorm removal
for block in model.transformer.h:
    block.ln_1.weight.data = block.ln_1.weight.data / 1e6
    block.ln_1.eps = 1e-5
    block.ln_2.weight.data = block.ln_2.weight.data / 1e6
    block.ln_2.e
... (read more)

Maybe this should be like Anthropic's shared decoder bias? Essentially subtract off the per-token bias at the beginning, let the SAE reconstruct this "residual", then add the per-token bias back to the reconstructed x. 

The motivation is that the SAE has a weird job in this case. It sees x, but needs to reconstruct x - per-token-bias, which means it needs to somehow learn what that per-token-bias is during training. 

However, if you just subtract it first, then the SAE sees x', and just needs to reconstruct x'. 

So I'm just suggesting changing&... (read more)

3danwil
Double thanks for the extended discussion and ideas! Also interested to see what happens. We earlier created some SAEs that completely remove the unigram directions from the encoder (e.g. old/gpt2_resid_pre_8_t8.pt).  However, a " Golden Gate Bridge" feature individually activates on " Golden" (plus prior context), " Gate" (plus prior), and " Bridge" (plus prior). Without the last-token/unigram directions these tended not to activate directly, complicating interpretability.
4tdooms
We haven't considered this since our idea was that the encoder could maybe use the full information to better predict features. However, this seems worthwhile to at least try. I'll look into this soon, thanks for the inspiration.

That's great thanks! 

My suggested experiment to really get at this question (which if I were in your shoes, I wouldn't want to run cause you've already done quite a bit of work on this project!, lol):

Compare 
1. Baseline 80x expansion (56k features) at k=30
2. Tokenized-learned 8x expansion (50k vocab + 6k features) at k=29 (since the token adds 1 extra feature)

for 300M tokens (I usually don't see improvements past this amount) showing NMSE and CE.

If tokenized-SAEs are still better in this experiment, then that's a pretty solid argument to use thes... (read more)

2tdooms
This is a completely fair suggestion. I'll look into training a fully-fledged SAE with the same number of features for the full training duration. 

About similar tokenized features, maybe I'm misunderstanding, but this seems like a problem for any decoder-like structure.

I didn't mean to imply it's a problem, but the intepretation should be different. For example, if at layer N, all the number tokens have cos-sim=1 in the tokenized-feature set, then if we find a downstream feature reading from " 9" token on a specific task, then we should conclude it's reading from a more general number direction than a specific number direction. 

I agree this argument also applies to the normal SAE decoder (if the cos-sim=1)

Although, tokenized features are dissimilar to normal features in that they don't vary in activation strength. Tokenized features are either 0 or 1 (or norm of the vector). So it's not exactly an apples-to-apples comparison w/ a similar sized dictionary of normal SAE features, although that plot would be nice!

I do really like this work. This is useful for circuit-style work because the tokenized-features are already interpreted. If a downstream encoded feature reads from the tokenized-feature direction, then we know the only info being transmitted is info on the current token.

However, if multiple tokenized-features are similar directions (e.g. multiple tokenizations of word "the") then a circuit reading from this direction is just using information about this set of tokens. 

2tdooms
That's awesome to hear, while we are not especially familiar with circuit analysis, anecdotally, we've heard that some circuit features are very disappointing (such as the "Mary" feature for IOI, I believe this is also the case in Othello SAEs where many features just describe the last move). This was a partial motivation for this work. About similar tokenized features, maybe I'm misunderstanding, but this seems like a problem for any decoder-like structure. In the lookup table though, I think this behaviour is somewhat attenuated due to the strict manual trigger, which encourages the lookup table to learn exact features instead of means.

Do you have a dictionary-size to CE-added plot? (fixing L0)

So, we did not use the lookup biases as additional features (only for decoder reconstruction)

I agree it's not like the other features in that the encoder isn't used, but it is used for reconstruction which affects CE. It'd be good to show the pareto improvement of CE/L0 is not caused by just having an additional vocab_size number of features (although that might mean having to use auxk to have a similar number of alive features).

8tdooms
One caveat that I want to highlight is that there was a bug training the tokenized SAEs for the expansions sweep, the lookup table isn't learned but remained at the hard-coded values... They are therefore quite suboptimal. Due to some compute constraints, I haven't re-run that experiment (the x64 SAEs take quite a while to train). Anyway, I think the main question you want answered is if the 8x tokenized SAE beats the 64x normal SAE, which it does. The 64x SAE is improving slightly quicker near the end of training, I only used 130M tokens. Below is an NMSE plot for k=30 across expansion factors (the CE is about the same albeit slightly less impacted by size increase). the "tokenized" label indicates the non-learned lookup and the "Learned" is the working tokenized setup.
3Logan Riggs
Although, tokenized features are dissimilar to normal features in that they don't vary in activation strength. Tokenized features are either 0 or 1 (or norm of the vector). So it's not exactly an apples-to-apples comparison w/ a similar sized dictionary of normal SAE features, although that plot would be nice!
Load More