Statistics student at ETH. Previously, thinking about evals as a research fellow at Pivotal.
Lennart Finke
Karma: 126
Thanks! From the feedback it seems like I’ll have to make a v2 and add this. Edit: And good idea, I’ve pushed Gemini’s classification as well.
How Much of AI Labs’ Research Is Safety?
> Do you have any intuition on what “ocode” means?
Oddly enough, the “ocode” token also has a child in the BPE merging rules, namely token #107224, “ocoder”. I really wonder where the tokenizer got that one from. This is probably a red hering for the embedding matrix norm though, it’s very likely from ” pseudocode”, which gets tokenized as ” pseud”-”ocode”.
> Furthermore, it is unclear to me from which GPT OSS model you take those English L2 norm embeddings from.
I believe it was 120b.
> And lastly, can you please elaborate why having the tokenizer means we can use the GPT OSS embeddings to study the token list without having to look at each token’s text content.
Because the embeddings contain all the semantic information needed to speak human language as well as gpt-oss does, which is really well! A nearest-neighbor search in embedding space gives semantically similar tokens, etc.
> The above text is quite compelling and I am currently doing ablations on reasoning and in particular I want to prevent the model from using these reasoning words and see how the reasoning degrades
Looking forward to the results!
The sum of its parts: composing AI control protocols
It is no secret that labs indiscriminately scrape from all over the internet, but usually a filter is applied to remove unwanted content. Because I assume the pretraining team would consider these strings as unwanted content, we can infer there is room to improve the pretraining filtering. I think that better pretraining filtering is useful for mitigating emergent misalignment.
What GPT-oss Leaks About OpenAI’s Training Data
SimpleStories: A Better Synthetic Dataset and Tiny Models for Interpretability
The component of ignoring two intervening characters is less mysterious to me. For example, a numbered list like “1. first_token 2. second_token …” would need this pattern. I am wondering mostly why the specific map from b’xa1′-b’xba’ to a-z is learned.
Why does Claude Speak Byzantine Music Notation?
A much appreciated update, thank you!
Good point, and I was conflicted whether to put my thoughts about this at the end of the post. My best theory is that increased persuasion abilities looks something like “totalitarian government agents doing solid scaffolding on open-source models to DM people on Facebook”. We will see that persuasive agents get better, but not know why and how. As stated in the introduction, persuasion detection is dangerous, but one of the few capabilities that could also be used defensively (i.e. detecting persuasion in an incoming email → displaying warning in UI and offer to rephrase).
In conclusion, definitely agree that we should consider closed-sourcing any improvements upon the above baseline and only show them to safety orgs instead. Some people at AISI I have talked to while working on persuasion are probably interested in this.
Thanks, fixed!
Agreed, although that it turn makes me wonder why it does perform a bit better than random. Maybe there is some nondeclarative knowledge about the image, or some blurred position information? I might test next how much vision is bottlenecking here by providing a text representation of the grid, as in Ryan Greenblatt’s work on ARC-AGI.
By “data sources you’re using”, do you mean I have to be more careful in what kinds of public data to use, and to filter by topic to remove non-LLM research, or do you mean the only way to meaningfully answer this would need access to non-public data? I’ve not worked at a lab, but I would have thought we’d at least see the successful results, with a 6 month delay or so.