36 comments

[ 3.2 ms ] story [ 50.6 ms ] thread
This is very interesting. I don't see much discussion of interpretability in day to the day discourse of AI builders. I wonder if everyone assumes it to either be solved, or to be too out of reach to bother stopping and thinking about.
Mostly out of reach. There is a ton of research on figuring out how to do this coming out every day, including both proposals of new ways to do things and (often strong) critiques of old or recently proposed ways of doing things. Interpretability (esp. for large, modern models) is very, very far from being a solved problem.
Now this is something which is very interesting to see and might be the answer to the explainability issue with LLMs, which can unlock a lot more use-cases that are off limits.

We'll see.

Maybe I’m not creative enough to see the potential, but what value does this bring ?

Given the example I saw about CRISPR, what does this model give over a different, non explaining model in the output ? Does it really make me more confident in the output if I know the data came from Arxiv or Wikipedia ?

I find the LLM outputs are subtlety wrong not obviously wrong

Most interpretability methods fail for LLMs because they try to explain outputs without modeling the intent, constraints, or internal structure that produced them. Token‑level attribution is useful, but without a framework for how the model reasons, you’re still explaining shadows on the wall.
It's a neat party trick, but explainability it's not solution to any AI safety issue I care about. It's a distraction from real problems, which is everything else around the model. The inflexible bureaucratic systems that make it hard to exercise rights and deflect accountability.
Either I'm missing something or this is way overstated.

Steerling appears to be just a discrete diffusion model where the final hidden states are passed through a sparse autoencoder (a common interpretability layer) before the LM head.

They also use a loss that aligns the SAE'S activations with labelled concepts? However, this is an example of "The Most Forbidden Technique" [1], and could make the model appear interpretable without the attributed concepts actually having causal effect on the model's decisions.

1: https://thezvi.substack.com/p/the-most-forbidden-technique

If this decomposition actually holds, it's the first model where you could show a regulator why it produced a given output.
Can you use this to decrease hallucinations?
This seems really interesting. While Anthropic tried to use dictionary learning over an existing model to try to extract concepts, this almost feels like training the model alongside the dictionary itself (or rather, the model and the dictionary are intertwined).
(comment deleted)
Does anybody know if I can try this online?
Just wanted to say i think most interpretability research it's just a smoke show nowadays but this is actually the first one that i think has a very serious potential. I love that the SAE is actually constrained and not just slapped unsupervised posthoc.

How granular can you get the source data attribution? Down to individual let's say Wikipedia topics? Probably not urls?

Would be interested to see this scale to 30/70b

Looks very interesting. Can you comment on why you think this model can give comparable performance with less training data?
So maybe one day we'll see coding agents like Claude Code create and update an ATTRIBUTION.md, citing all the open source projects and their licenses used to generate code in your project?
This is actually being built right now. ATTRIBUTION.md (https://attribution.md) is a protocol that does exactly this. You drop a file in your repo root with a few lines of YAML, and it asks AI coding agents to prompt users to star the repos they built on.

The key design choice is that it does not automate anything. The agent surfaces a prompt, the user decides yes or no. No bulk starring, no forced actions. The spec also deliberately stays out of licensing territory. It is purely a social recognition layer.

It is at v0.1 and no agent supports it yet, but the spec and schema are published and open for feedback: https://github.com/attributionmd/attribution.md

Looks neat and original, congrats!

I don't quite grasp how to interpret the training data attribution process. For example, it seems to say that for a given sentence like "They argued that humans tend to weigh losses more heavily than gains, leading to risk aversion", 24% is attributed to Wikipedia and 23% to Arxiv.

Does that mean that the concepts used in this sentence are also found in those datasets, and that's what's getting compared here? Or does it mean that you can track down which parts of the training data were interpolated to create that sentence?

Always happy to see improvements on explanable LLMs. Congrats!
Seems pretty cool. You can simply block the concept of tiananmen square and it will be permanently removed from the brain. Ideal.
(comment deleted)
In the recent HN thread announcing the new Gemini coding agent (https://news.ycombinator.com/item?id=47074735), a lot of people complained about Gemini’s tendency to do unwanted refactors, not perform requested actions, etc.

It made me cautiously optimistic that all of Anthropic’s work on alignment, which they did for AI safety, is actually the cause of Claude code’s comparatively superior utility (and their present success). I wonder if future progress (maybe actual AGI?) lies in the direction of better and better alignment, so I think this is super cool and I’m suddenly really interested in experiments like this