I wrote up the project but it is Matt's

I've been looking for someone to do this for a while, and I found this person from the Sentient Futures fellowship. They were recommended to me, and I worked with them on the shape of this project.

It's very important to know how the animal agriculture industry is adopting technology so we can determine if any of them are very bad and we can build solutions that are better.

I also paid back a $4000 loan so the total size of the grant is $64000

Hello, am curious about this project. I sent you an e-mail under nicholas@voltarecovery.org.

I wrote this proposal with an AI assistant, and the flag is correct about the prose. English is not my first language. I arrived in the US in 2020 with none of it, and I still think faster than I write. The project, the 1,175 saved answers, the 30 businesses I measured and every decision here are mine. I gave the material and the corrections; the assistant put it into English. Flagging the prose is fair. It is not a statement about the work, and I would rather say this myself than have you find it.

I'm now steering introductory SLT material development in partnership with Iliad.

@ajskateboarder (the DCT post notes that the output directions are likely not features on their own, however i believe this won't be necessary in order to factor DCTs)

What progress have you made since your last update?

One way I tried going about the problem was investigating the structure of deep causal transcoding, which shows strong generalization abilities for language models (i.e. finding many interesting generalizations given less sampling). The post gives partial evidence for the 'persistent shallow circuits' hypothesis, which if true could considerably alter the difficulty of calculating expectations. It could make the problem simpler since shallow circuits are comparatively simpler structures to locate, and to improve on Pareto frontier for estimating generalizations, all that roughly needs to be done is to enumerate the circuits on the DCT. It could also be true that LLMs and other nets do mechanistically implement generalizations through distributed shallow circuits, and we can't rely on the structure of deep sequential circuits to make the problem easier (and the problem would be harder if this was false). Overall, being able to factor the structure of the DCT seems important for making better estimations.

It seems hard to extract shallow causal structure from DCTs where many features interact to activate one feature, and even harder vice-versa; most of these circuits are representative of the DCT, but not the original LLM/transformer/net. However, it does seem easier to extract circuits that map from one input feature to an output feature in a highly data-efficient manner.

https://www.lesswrong.com/posts/8iG7orf6q8hNHT4Xh/ajskateboarder-s-shortform?commentId=XNcbxWHhqYHpMzXXr

What are your next steps?

Continue looking into how to best factor DCTs into ensembles of shallow circuits, and how 'surprising' the generalizations these circuits encode are in context of other interp methods (hopefully by relying on some scaling laws/theory/well-reasoned assumptions)

Is there anything others could help you with?

I think the DCT method has broadly not been used very much, and it would be good if people looked more into how they work/generalize (the original authors seem to have made a DCT follow-up in weight-space https://arxiv.org/pdf/2606.29604)

Hi everyone, I am Wanfu Zhuang, creator of AxiomForge. I want to proactively address the Pangram automated AI flag on this proposal.

As an independent researcher whose primary language is not English, I used an LLM to help translate and refine my technical draft into clear English prose.

However, 100% of the mathematical conception, algorithmic architecture, and code execution are my own original work:

1. Fully Verified Codebase: All source code is public under the MIT License at [github.com/Zwf5458-Py/AxiomForge](https://github.com/Zwf5458-Py/AxiomForge).

2. Automated CI Suite: 16 unit tests passing across Grassmannian slice projections, affine space validations, and AST syntax sanitizers.

3. Reproducibility: Features a working 3D interactive slice visualizer and reproducible benchmarks exploring finite-field Cap Sets via LLM program evolution.

I encourage reviewers and regrantors to inspect the repository directly and run the benchmarks locally. The $4,000 grant will be used strictly for compute and API credits to scale search into 6D–8D combinatorial spaces.

I am eager to answer any technical questions or feedback here in the discussion!

@AdnanSlef Much appreciated 🙌

@AdnanSlef Support Doom Debates. It hosts the most clearheaded discussions of AI, with a laser focus on the actual problem: we don't know how to align superintelligence, so we must not try to build it.

On AI assistance: Yes, I used AI tools to develop and edit this proposal. That is also part of the research practice being studied here. I use models to help structure language, test distinctions, and challenge drafts; I make the research decisions, verify the claims, revise the text, and take responsibility for the final proposal. I’d rather state that plainly than let an “AI-generated” flag imply a simpler authorship story than the one that actually occurred.

For your reference, we have uploaded AGEP v3.14—which is scheduled for peer-review submission to JMLR next month—to Zenodo.

https://doi.org/10.5281/zenodo.22716189

@caroline hi, Austin here -- I'm making a public note that the above is a fake/joke account, not tied to Caroline Ellison. we'll likely rename your account and username if you don't do so yourself.

I believe it's very important to have pathways to whistleblowing, I deeply support the cause.

I like Rubi's work, I think he made non-zero progress on corrigibility. It's plausible he and his team will have more good ideas. I ultimately don't think the approach of directly applying it to RL experiments makes much sense, but maybe it'll spark more ideas.

NOT AI generated.

Approving and making a small grant in support! Generally happy to fund EA events around the world.

(at Fumi's request, we lowered the min funding bar down from its initial $1.4k; Fumi asked, "we have more than enough since we downscaled with less people, but we still need money to hit our initial minimum budget to get our funds.")

I support your work - keep it up!

@raelifin https://www.lesswrong.com/posts/CKArJ4GAQGFjnhjx6/corrigibility-research-fund-grantees-round-1

I am excited about Rubi's work, and Principled Agents more broadly. They're doing real alignment work, which we desperately need more of. As manager of the Corrigibility Research Fund I awarded Rubi's team a $20k grant as a vote-of-confidence, partly to give them a little more runway, but mostly to nudge other, bigger funders to send more substantial support to their newborn research org.

Thanks David!

I wrote up the project but it is Matt's

I've been looking for someone to do this for a while, and I found this person from the Sentient Futures fellowship. They were recommended to me, and I worked with them on the shape of this project.

It's very important to know how the animal agriculture industry is adopting technology so we can determine if any of them are very bad and we can build solutions that are better.

I also paid back a $4000 loan so the total size of the grant is $64000

Hello, am curious about this project. I sent you an e-mail under nicholas@voltarecovery.org.

I wrote this proposal with an AI assistant, and the flag is correct about the prose. English is not my first language. I arrived in the US in 2020 with none of it, and I still think faster than I write. The project, the 1,175 saved answers, the 30 businesses I measured and every decision here are mine. I gave the material and the corrections; the assistant put it into English. Flagging the prose is fair. It is not a statement about the work, and I would rather say this myself than have you find it.

I'm now steering introductory SLT material development in partnership with Iliad.

@ajskateboarder (the DCT post notes that the output directions are likely not features on their own, however i believe this won't be necessary in order to factor DCTs)

What progress have you made since your last update?

One way I tried going about the problem was investigating the structure of deep causal transcoding, which shows strong generalization abilities for language models (i.e. finding many interesting generalizations given less sampling). The post gives partial evidence for the 'persistent shallow circuits' hypothesis, which if true could considerably alter the difficulty of calculating expectations. It could make the problem simpler since shallow circuits are comparatively simpler structures to locate, and to improve on Pareto frontier for estimating generalizations, all that roughly needs to be done is to enumerate the circuits on the DCT. It could also be true that LLMs and other nets do mechanistically implement generalizations through distributed shallow circuits, and we can't rely on the structure of deep sequential circuits to make the problem easier (and the problem would be harder if this was false). Overall, being able to factor the structure of the DCT seems important for making better estimations.

It seems hard to extract shallow causal structure from DCTs where many features interact to activate one feature, and even harder vice-versa; most of these circuits are representative of the DCT, but not the original LLM/transformer/net. However, it does seem easier to extract circuits that map from one input feature to an output feature in a highly data-efficient manner.

https://www.lesswrong.com/posts/8iG7orf6q8hNHT4Xh/ajskateboarder-s-shortform?commentId=XNcbxWHhqYHpMzXXr

What are your next steps?

Continue looking into how to best factor DCTs into ensembles of shallow circuits, and how 'surprising' the generalizations these circuits encode are in context of other interp methods (hopefully by relying on some scaling laws/theory/well-reasoned assumptions)

Is there anything others could help you with?

I think the DCT method has broadly not been used very much, and it would be good if people looked more into how they work/generalize (the original authors seem to have made a DCT follow-up in weight-space https://arxiv.org/pdf/2606.29604)

Hi everyone, I am Wanfu Zhuang, creator of AxiomForge. I want to proactively address the Pangram automated AI flag on this proposal.

As an independent researcher whose primary language is not English, I used an LLM to help translate and refine my technical draft into clear English prose.

However, 100% of the mathematical conception, algorithmic architecture, and code execution are my own original work:

1. Fully Verified Codebase: All source code is public under the MIT License at [github.com/Zwf5458-Py/AxiomForge](https://github.com/Zwf5458-Py/AxiomForge).

2. Automated CI Suite: 16 unit tests passing across Grassmannian slice projections, affine space validations, and AST syntax sanitizers.

3. Reproducibility: Features a working 3D interactive slice visualizer and reproducible benchmarks exploring finite-field Cap Sets via LLM program evolution.

I encourage reviewers and regrantors to inspect the repository directly and run the benchmarks locally. The $4,000 grant will be used strictly for compute and API credits to scale search into 6D–8D combinatorial spaces.

I am eager to answer any technical questions or feedback here in the discussion!

@AdnanSlef Much appreciated 🙌

@AdnanSlef Support Doom Debates. It hosts the most clearheaded discussions of AI, with a laser focus on the actual problem: we don't know how to align superintelligence, so we must not try to build it.

On AI assistance: Yes, I used AI tools to develop and edit this proposal. That is also part of the research practice being studied here. I use models to help structure language, test distinctions, and challenge drafts; I make the research decisions, verify the claims, revise the text, and take responsibility for the final proposal. I’d rather state that plainly than let an “AI-generated” flag imply a simpler authorship story than the one that actually occurred.

For your reference, we have uploaded AGEP v3.14—which is scheduled for peer-review submission to JMLR next month—to Zenodo.

https://doi.org/10.5281/zenodo.22716189

@caroline hi, Austin here -- I'm making a public note that the above is a fake/joke account, not tied to Caroline Ellison. we'll likely rename your account and username if you don't do so yourself.

I believe it's very important to have pathways to whistleblowing, I deeply support the cause.

I like Rubi's work, I think he made non-zero progress on corrigibility. It's plausible he and his team will have more good ideas. I ultimately don't think the approach of directly applying it to RL experiments makes much sense, but maybe it'll spark more ideas.

NOT AI generated.

Approving and making a small grant in support! Generally happy to fund EA events around the world.

(at Fumi's request, we lowered the min funding bar down from its initial $1.4k; Fumi asked, "we have more than enough since we downscaled with less people, but we still need money to hit our initial minimum budget to get our funds.")

I support your work - keep it up!

@raelifin https://www.lesswrong.com/posts/CKArJ4GAQGFjnhjx6/corrigibility-research-fund-grantees-round-1

I am excited about Rubi's work, and Principled Agents more broadly. They're doing real alignment work, which we desperately need more of. As manager of the Corrigibility Research Fund I awarded Rubi's team a $20k grant as a vote-of-confidence, partly to give them a little more runway, but mostly to nudge other, bigger funders to send more substantial support to their newborn research org.

Thanks David!