@Ericfrasersf Thanks so much Eric! I am still struglling to build the deck mainly because as you know, Japanese presentation documents have too many words while Western tend to have a few.

And Yes, ganbarimasu!

What this grant produced

This grant, in combination with my other Manifund grant, funded my transition into AI safety research and led to the formation of Lyptus Research, a not-for-profit AI safety research group based in Australia.

The published output during the grant period was our research note on offensive cyber time horizons. which extended METR's time-horizon methodology to offensive cybersecurity, grounded in a human expert study with 10 professional security practitioners.

Earlier in the grant period I also explored attack selection and monitor capability evaluation. That direction did not eventuate into a finished product.

Lyptus Research continues to work on dangerous capability evaluation using similar methods, and on technical research into alignment.

Cost breakdown

Across the two grants, approximately $70,000 USD supported the published cyber work

- Salaries: ~$50,000

- API credits: ~$10,000

- Contractors (study expert participants): ~$10,000

- Infrastructure, tools, and ops: ~$5,000

A further ~$20,000 USD supported my earlier transition into the field

- About 1.5 to 2 months of independent AI control research

- Australian domestic travel

- A couple of conferences in Australia

The project ran longer than originally scoped, and the shortfall was covered through our subsequent bridge funding grant.

What this grant produced

This grant, in combination with my other Manifund grant, funded my transition into AI safety research and led to the formation of Lyptus Research, a not-for-profit AI safety research group based in Australia.

The published output during the grant period was our research note on offensive cyber time horizons. which extended METR's time-horizon methodology to offensive cybersecurity, grounded in a human expert study with 10 professional security practitioners.

Earlier in the grant period I also explored attack selection and monitor capability evaluation. That direction did not eventuate into a finished product.

Lyptus Research continues to work on dangerous capability evaluation using similar methods, and on technical research into alignment.

Cost breakdown

Across the two grants, approximately $70,000 USD supported the published cyber work

- Salaries: ~$50,000

- API credits: ~$10,000

- Contractors (study expert participants): ~$10,000

- Infrastructure, tools, and ops: ~$5,000

A further ~$20,000 USD supported my earlier transition into the field

- About 1.5 to 2 months of independent AI control research

- Australian domestic travel

- A couple of conferences in Australia

The project ran longer than originally scoped, and the shortfall was covered through our subsequent bridge funding grant.

Erik, thanks so much for acknowledging my work! Gradually I meet people who share the similar idea/work from different cultures and languages. This reminds me of "Stand Alone Complex" from Ghost in the Shell (one of my favorite anime from Japan, and I analyse the society with the series). I believe Japan can and should offer more than what we see now with cultural frame. @eleklem

Hello everyone,

I decided to start-down Coordinal. There are a number of factors that ended up at this point, I've posted a more detailed postmortem in a few different places, including the EA Forum: https://forum.effectivealtruism.org/posts/uTffD74jJrQvHJuBY/coordinal-a-postmortem

Broadly, I made a lot of mistakes and burned out trying to do too much with too little, and was getting pulled in all sorts of directions to try to fundraise. Please feel free to reach out if you have any thoughts/questions!

Symbolical contribution because there is a lot of potential and need to innovate and just start things in this space.

@gleech Thanks Gavin. Quick clarification for the record — the directions on the site are the long-horizon Φ program, not paper commitments tied to this $15K cycle, which catalyzes three ICLR 2027 papers (Φ-Arena, mechinterp, energy-bounded). Will tighten the readme framing to make that distinction clearer. ：）

Truly supporting this novel idea

Had a brief exchange with Yuchen and like the energy (four real papers submitted simultaneously, four more promised). The readme [here](https://github.com/phi-monster/phi-monster.github.io#what-we-work-on) is a better explanation than the above. The work is cool and indeed nearly completely neglected. Luck!

Evidence note:

Selected public records for this project:

- Medium profile: https://medium.com/@office.dosanko

- GitHub: https://github.com/dosanko-tousan

- Zenodo records: https://zenodo.org/search?q=metadata.creators.person_or_org.name%3A%22Takeuchi%2C%20Akimitsu%22&l=list&p=1&s=10&sort=bestmatch

- Qiita: https://qiita.com/dosanko_tousan

- Zenn: https://zenn.dev/dosanko_tousan

Private evidence available on request:

- Cohere Labs Catalyst Grant acceptance email and API credit dashboard

- GLG independent consultant confirmation letter

- Discover Psychology / Springer Nature dashboard showing technical check passed and currently with editor

Every day I see how speciesist these AI models are. I feel it is of the utmost importance to align them not only with human interests but also with those of other animals.

Güzel bir proje olmuş dostum iyi şanslar dilerim 👍

@Lisa-Intel @Austin hmm

Animal welfare matters deeply to me, and projects at the intersection of AI and animals are underfunded.

As AI is deployed across labs and industry, animals should be represented from the start. Marcus brings the thoughtfulness and urgency this space needs, and I’m excited to support the Falcon Fund.

Tomoko-San,

I know you will impress everyone with your insights

Ganbatte Ne!

Eric

Tomoko - wishing you well in this effort. I remember speaking with you during the Apart research Technical AI Governance hackathon and have been impressed by your work since then. We need more Japanese AI safety / risk voices in the community, and your cross-cultural work is important and underserved from my perspective.

Thanks @NeelNanda!

I’m funding two of my MATS scholars for two months each for work done on a paper explaining what’s going on with subliminal learning. This was during a period when they were not funded by the main MATS programme, paid at the same rates as MATS. While there is an obvious conflict of interest, since I supervised the paper, I do not financially profit and my bar for mentoring people is noticeably higher than my bar for funding them, so I am not too worried about this. And I think this is a great paper! Subliminal learning is a mysterious phenomenon of wide interest, and I think the forthcoming paper provides a satisfying explanation. Basic science on mysterious phenomena in LLMs is valuable work in a bunch of hard to foresee ways.

Project update (May 2026):

Over the past several weeks, I continued externalizing parts of the research into public-facing reproducible artifacts and governance-oriented documentation.

New additions now include:

- multiple Zenodo DOI records covering behavioral audit, interaction ecology, and epistemic governance directions

- a public reproducibility layer (minimal repro package, schema definitions, baseline comparisons)

- governance-oriented framing for long-term human-AI interaction continuity and epistemic drift

- expanded GitHub documentation and public-safe architecture separation

The project is gradually evolving from isolated observational notes into a more structured longitudinal governance research line focused on:

- runtime interaction continuity

- behavioral drift

- cross-model ecology

- AI-mediated knowledge production

- long-horizon epistemic governance

Importantly, this work is not presented as a finalized theory or validated consensus. The current stage is best understood as an ongoing observational and protocolization effort attempting to document governance gaps emerging from sustained human-AI interaction.

Hi @Jesse-Richardson, may I ask you to take a look at my project? Grateful if you do.

Best,

Pedro

Thank you, Jesse! We could not do our work without the dedication and sacrificial giving of donors like you. Thank you for your contribution to our common mission - mitigating extreme risks to humanity from transformational AI.

@Ericfrasersf Thanks so much Eric! I am still struglling to build the deck mainly because as you know, Japanese presentation documents have too many words while Western tend to have a few.

And Yes, ganbarimasu!

What this grant produced

This grant, in combination with my other Manifund grant, funded my transition into AI safety research and led to the formation of Lyptus Research, a not-for-profit AI safety research group based in Australia.

The published output during the grant period was our research note on offensive cyber time horizons. which extended METR's time-horizon methodology to offensive cybersecurity, grounded in a human expert study with 10 professional security practitioners.

Earlier in the grant period I also explored attack selection and monitor capability evaluation. That direction did not eventuate into a finished product.

Lyptus Research continues to work on dangerous capability evaluation using similar methods, and on technical research into alignment.

Cost breakdown

Across the two grants, approximately $70,000 USD supported the published cyber work

- Salaries: ~$50,000

- API credits: ~$10,000

- Contractors (study expert participants): ~$10,000

- Infrastructure, tools, and ops: ~$5,000

A further ~$20,000 USD supported my earlier transition into the field

- About 1.5 to 2 months of independent AI control research

- Australian domestic travel

- A couple of conferences in Australia

The project ran longer than originally scoped, and the shortfall was covered through our subsequent bridge funding grant.

What this grant produced

This grant, in combination with my other Manifund grant, funded my transition into AI safety research and led to the formation of Lyptus Research, a not-for-profit AI safety research group based in Australia.

The published output during the grant period was our research note on offensive cyber time horizons. which extended METR's time-horizon methodology to offensive cybersecurity, grounded in a human expert study with 10 professional security practitioners.

Earlier in the grant period I also explored attack selection and monitor capability evaluation. That direction did not eventuate into a finished product.

Lyptus Research continues to work on dangerous capability evaluation using similar methods, and on technical research into alignment.

Cost breakdown

Across the two grants, approximately $70,000 USD supported the published cyber work

- Salaries: ~$50,000

- API credits: ~$10,000

- Contractors (study expert participants): ~$10,000

- Infrastructure, tools, and ops: ~$5,000

A further ~$20,000 USD supported my earlier transition into the field

- About 1.5 to 2 months of independent AI control research

- Australian domestic travel

- A couple of conferences in Australia

The project ran longer than originally scoped, and the shortfall was covered through our subsequent bridge funding grant.

Erik, thanks so much for acknowledging my work! Gradually I meet people who share the similar idea/work from different cultures and languages. This reminds me of "Stand Alone Complex" from Ghost in the Shell (one of my favorite anime from Japan, and I analyse the society with the series). I believe Japan can and should offer more than what we see now with cultural frame. @eleklem

Hello everyone,

I decided to start-down Coordinal. There are a number of factors that ended up at this point, I've posted a more detailed postmortem in a few different places, including the EA Forum: https://forum.effectivealtruism.org/posts/uTffD74jJrQvHJuBY/coordinal-a-postmortem

Broadly, I made a lot of mistakes and burned out trying to do too much with too little, and was getting pulled in all sorts of directions to try to fundraise. Please feel free to reach out if you have any thoughts/questions!

Symbolical contribution because there is a lot of potential and need to innovate and just start things in this space.

@gleech Thanks Gavin. Quick clarification for the record — the directions on the site are the long-horizon Φ program, not paper commitments tied to this $15K cycle, which catalyzes three ICLR 2027 papers (Φ-Arena, mechinterp, energy-bounded). Will tighten the readme framing to make that distinction clearer. ：）

Truly supporting this novel idea

Had a brief exchange with Yuchen and like the energy (four real papers submitted simultaneously, four more promised). The readme [here](https://github.com/phi-monster/phi-monster.github.io#what-we-work-on) is a better explanation than the above. The work is cool and indeed nearly completely neglected. Luck!

Evidence note:

Selected public records for this project:

- Medium profile: https://medium.com/@office.dosanko

- GitHub: https://github.com/dosanko-tousan

- Zenodo records: https://zenodo.org/search?q=metadata.creators.person_or_org.name%3A%22Takeuchi%2C%20Akimitsu%22&l=list&p=1&s=10&sort=bestmatch

- Qiita: https://qiita.com/dosanko_tousan

- Zenn: https://zenn.dev/dosanko_tousan

Private evidence available on request:

- Cohere Labs Catalyst Grant acceptance email and API credit dashboard

- GLG independent consultant confirmation letter

- Discover Psychology / Springer Nature dashboard showing technical check passed and currently with editor

Every day I see how speciesist these AI models are. I feel it is of the utmost importance to align them not only with human interests but also with those of other animals.

Güzel bir proje olmuş dostum iyi şanslar dilerim 👍

@Lisa-Intel @Austin hmm

Animal welfare matters deeply to me, and projects at the intersection of AI and animals are underfunded.

As AI is deployed across labs and industry, animals should be represented from the start. Marcus brings the thoughtfulness and urgency this space needs, and I’m excited to support the Falcon Fund.

Tomoko-San,

I know you will impress everyone with your insights

Ganbatte Ne!

Eric

Tomoko - wishing you well in this effort. I remember speaking with you during the Apart research Technical AI Governance hackathon and have been impressed by your work since then. We need more Japanese AI safety / risk voices in the community, and your cross-cultural work is important and underserved from my perspective.

Thanks @NeelNanda!

I’m funding two of my MATS scholars for two months each for work done on a paper explaining what’s going on with subliminal learning. This was during a period when they were not funded by the main MATS programme, paid at the same rates as MATS. While there is an obvious conflict of interest, since I supervised the paper, I do not financially profit and my bar for mentoring people is noticeably higher than my bar for funding them, so I am not too worried about this. And I think this is a great paper! Subliminal learning is a mysterious phenomenon of wide interest, and I think the forthcoming paper provides a satisfying explanation. Basic science on mysterious phenomena in LLMs is valuable work in a bunch of hard to foresee ways.

Project update (May 2026):

Over the past several weeks, I continued externalizing parts of the research into public-facing reproducible artifacts and governance-oriented documentation.

New additions now include:

- multiple Zenodo DOI records covering behavioral audit, interaction ecology, and epistemic governance directions

- a public reproducibility layer (minimal repro package, schema definitions, baseline comparisons)

- governance-oriented framing for long-term human-AI interaction continuity and epistemic drift

- expanded GitHub documentation and public-safe architecture separation

The project is gradually evolving from isolated observational notes into a more structured longitudinal governance research line focused on:

- runtime interaction continuity

- behavioral drift

- cross-model ecology

- AI-mediated knowledge production

- long-horizon epistemic governance

Importantly, this work is not presented as a finalized theory or validated consensus. The current stage is best understood as an ongoing observational and protocolization effort attempting to document governance gaps emerging from sustained human-AI interaction.

Hi @Jesse-Richardson, may I ask you to take a look at my project? Grateful if you do.

Best,

Pedro

Thank you, Jesse! We could not do our work without the dedication and sacrificial giving of donors like you. Thank you for your contribution to our common mission - mitigating extreme risks to humanity from transformational AI.