What progress have you made since your last update?
We hosted the first DC Roundtable in September and, on what we learned, relaunched the program as private Consultations on a Bold and Proper US-China-Led AI Treaty (cbpai.org/consultations). We now have working channels to four of the twelve power brokers we identified. The Strategic Memo is at v2.9 (Aug 2026), v3.0 in draft. We hired a fractional COO and have a Head of Policy ready to join once the next round closes. Full record: cbpai.org/achievements.

What are your next steps?
A $30K bridge for the next 90 days, raised here and from past donors, while we pitch an 18-month round of about $800K to institutional funders (cbpai.org/roadmap). I've rewritten this page to match; the old $150K goal is gone.

Is there anything others could help you with?
If you fund AI-safety work, two pages: cbpai.org/s/Donation-Case-for-AI-Experts-Two-Pager.pdf. If you know a regrantor who should see this, forward it. @RyanKidd — your write-up, a year late, would still be the most useful thing on this page.

@atilgandev Good question!

This is actual a central problem with any system that links a claim with prior experiments that tested "the same claim". No two experiments are precisely the same.

We started by focusing on direct replications, where authors try to reproduce a previous experiment as closely as they can. These are rather rare however outside of the psychology.

Next we broadened our search to "close replications" and we distinguished two different types - close experiments and close extensions. Close experiments are looking for the same claim/effect as a prior work but using slightly different methods. Close extensions, which are very common in fields like genetics and biomedicine, are testing if an effect generalizes (usually to a different human population).

Finally, we have been collecting "conceptual replications". These are the most loose form of replication, where two researchers are looking for a similar "general effect" but in different ways and different contexts.

In our database viewer, you can filter to a subset of the four types of replication (direct, close extension, close experiment, conceptual).

I absolutely agree that surfacing changes in experimental context could be valuable. For instance, a drug that worked in younger group may not work in an older population. A psychological effect found in Americans may not work in Japanese. A failure to replicate often means that an effect does not generalize. However, we choose not to do this since it adds an additional level of complexity. A judgement call has to be made somewhere whether the change in context is actually relevant. An experiment done in a room with windows may get a different result than one done in a room with no windows. Do the windows matter? Someone has to make a call using scientific theory and we feel it's not our business to be doing that.

Another issue is the relevant differences are often unknown or not something you can tell from what is said in the paper (for instance the experimental finding may have to do with how a flask was shaken or an animal was handled. It seems scientists regularly discover new ways that mouse cage design effect the biology of mice).

Additionally, from a metascience perspective, what we are interested in is the claims scientists make, which tend to be of a more general character anyway. Scientists usually don't say "we found that X happens in people tested at 1 pm on a Monday afternoon in a room with windows who had just had lunch and were being paid and looked at a 20 inch computer monitor and were all educated Americans in a room with humidity level Y and temperature Z ...". They usually make some more generalizable claim like "a computer monitor that has a lower frame rate degrades performance on a visual task". There is a strong bias towards overclaiming the generalizability and importance of an effect, this is called importance hacking (https://www.clearerthinking.org/post/importance-hacking-a-major-yet-rarely-discussed-problem-in-science)

All of this is in line with this paper by Nosek and Errington, two major names in metascience: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3000691
See also https://metascienceobservatory.org/docs/defining-replication

Does that make sense? I'm very open to suggestions on how to deal with replications that aren't direct / have important changes, so I'd be curious to hear if you have any thoughts.

Description of subprojects and results, including major changes from the original proposal

I reviewed national policy and technical documents (JEE) to align with global/regional standards. Next, I interviewed 10 key stakeholders to map local gaps in training and incident reporting. Using this information, I built the BLAM tool with safety indicators across 10 different areas/domains. I then tested tool at NPHRL at Monrovia. The assessments revealed major safety gaps. For example, only a quarter of safety cabinets were functional and most relied on open burning for waste. The average safety score was 2.2/5. To turn these findings into action, I led a workshop with 2 national experts to agree on the most urgent priorities. Finally, I created a four pillar roadmap to guide future upgrades for governance, facilities, workforce and data systems.

Spending breakdown

2000 USD on travel and local logistics in Liberia. 1500 USD to compensate local staff at National Public Health Reference Laboratory (NPHRL) for their essential support. 2500 USD to host workshop and stakeholder meetings in Monrovia. 4000 USD to cover my time as PI to design  tool, lead the field assessments, analyze the data and disseminate among NPHIL leadership and stakeholders.

@mnaufil1 thanks for contributing to my journey.

The automated banner above flagging this proposal as "100% AI generated" is a false positive. Mechanistic interpretability grant applications rely on standard alignment nomenclature, kinematics notation, and structured engineering milestones.....all of which trigger high statistical confidence in simple token-prediction scanners.

Quick Links & Technical Verification for Reviewers:

To make diligence straightforward for regrantors, all underlying artifacts and live systems are open and accessible:

1. Mathematical Specification: Universal Thermodynamic Homeostasis & Epistemic Multi-Anchoring (DOI: 10.5281/zenodo.23315812)

2. Evidence Schema Standard: RFC-001: The .sent Cryptographic Evidence Container Specification v1.0 (DOI: 10.5281/zenodo.23314793)

3. Reference Implementation: GitHub repository with 14 passing unit tests covering Stage I prefill auditing and Stage II closed-loop decode steering: https://github.com/Regsiren/Auditor-OS-V50-Sentinel

4. Live Observer Dashboard: https://sentinel.thohatventures.com

Project State: The theoretical framework and software architecture are built and tested on synthetic activation shocks. This grant strictly funds the empirical red-teaming sweeps on dedicated H100 instances across Llama-3.3-70B using HarmBench and AdvGLUE.

I am actively monitoring this thread and am happy to answer questions regarding runtime overhead, kernel boundaries, or evaluation setups.

Good luck Parveen. They’d be lucky to have you.

Hi @Austin, since I cannot manually adjust the project timeline from my dashboard, could you please help me extend the crowdfunding closing date from October 18th to November 7th? This will give potential backers a bit more time to discover and support the campaign before my travel to Sydney for NeurIPS 2026 this December. Thank you

@atilgandev yes, that's the plan! If you see any duplicate entries, or any problems at all, don't hesitate to reach out.

@gnutella Thanks so much!

Jessica has now published the videos on Youtube, including talks from Emmett Shear, Kanjun Qiu, Yatharth Agarwal and others: https://www.youtube.com/@openimmanence

Excited to give them a watch!

I currently work on economic theory and alignment at the UK AI Security Institute (AISI), where we funded Principled Agents through the Alignment Project. I'm writing here in a personal capacity.

The alignment problem needs formal work (linked with empirical evidence) to progress. Economic theory is currently an under-researched avenue for understanding incentives and model behaviour, particularly in a way that will stay relevant for future, more capable models.

Rubi is one of a small number of people making progress in this area. He has been working seriously on alignment for a long time, and combines rigorous conceptual thinking in economics with a strong understanding of and engagement with the AI safety community. High-context work at this intersection is rare. I know Rubi well and rate his research and his agenda very highly. I would like to see this work fully funded, and a lot more investment in this kind of research.

Thank you!

Making a small donation primarily in support of Space, who I've met through Atlas and other AI safety/EA contexts.

As weird as it feels to do this, I have to do it. I would appreciate anyone who contributes to this and if you can't contribute financially but know a foundation, organisation, company, researcher, or institution that supports African scientists and scientific mobility, an introduction could be just as valuable. Thanks

Hi Skyler,

I'm MD Azizul Hakim, from Bangladesh. My sole-author paper was accepted as a poster at NeurIPS 2026 in Sydney. I'm very early in my career: I just finished a 4-year diploma, I haven't started a bachelor's degree, and I did the research without a lab or supervisor.

I'm trying to raise money for my flight and hotel ($2,500 minimum, $3,800 total). My institute has no funds, and QRT and Citadel rejected me because of my education level.

I saw that you support small, early-stage projects on Manifund. I'm not asking you to fund this yourself. Could you help in any of these ways?

1. Share my request with anyone who funds travel or early-career researchers.

2. Introduce me to a regrantor or a fund that might consider it.

3. Tell me if there's a better place to ask.

My request: https://manifund.org/projects/undergrad-sole-author-neurips-2026-main-track-paper-travel-to-present-in-sydney

Paper: https://neurips.cc/virtual/2026/loc/sydney/poster/150721

Thank you for reading this.

MD Azizul Hakim

azizulhakim8291@gmail.com

Hi, I'm a sole-author NeurIPS 2026 poster presenter from Bangladesh (diploma graduate, no undergrad yet). I'm raising $2,500 minimum for flight and hotel to present in Sydney. The paper is on reasoning efficiency, not safety, so it may be outside your focus. But i am asking you if you just donate me , it would be helpful for me..

https://manifund.org/projects/undergrad-sole-author-neurips-2026-main-track-paper-travel-to-present-in-sydney

Hi,

I noticed that you've helped several people travel to NeurIPS, so I wanted to ask whether you could help me too.

I'm Hakim, from Bangladesh. I just finished a 4-year diploma in computer science, which is equivalent to HSC in Globally, and I haven't started a bachelor's degree yet. My paper was accepted as a poster at the NeurIPS 2026 main conference in Sydney, and I'm the only author.

I tried my institute first, but polytechnic institutes in Bangladesh have no funds for this. I also applied to QRT and Citadel, and both rejected me because my education is below their minimum requirement. I've applied for NeurIPS financial assistance, which may cover registration and some hotel nights, but not my flight or my visa fee.

The total cost is $3,800, and the minimum I need is $2,500 for the flight and hotel. I've put my request on Manifund. Any amount, even a small one, would help:

https://manifund.org/projects/undergrad-sole-author-neurips-2026-main-track-paper-travel-to-present-in-sydney

You can verify the paper here: https://neurips.cc/virtual/2026/loc/sydney/poster/150721

My Google Scholar: https://scholar.google.com/citations?user=jVyIovcAAAAJ

Thank you very much for reading this.

Hakim

We have hired a scientist who is working on this project and have designed multiple versions of the genetic constructs needed to create the cell lines. We are cloning and evaluating designs at the moment. We are benefitting from progress in the field of controlled gene expression.

We are continuing to clone the constructs and develop the cell lines. Thanks to progress in the field, the number of molecular mechanisms we are evaluating has increased.

At this stage, we need to finish the evaluation, identify which strategy (if any) will be successful, and select the final version of the producer cell line. Once completed, we will progress to next steps (additional refinement of the technology, IP, commercialization, licensing, follow on funding etc.).

@Austin, I'm new to Manifund. This project builds an open record of what deployed AI agents were allowed to do and whether their actions can be undone. Two parts are working right now: the Ground Truth Credential (GTC) protocol, which a third party has built on, and a live Reversibility Check. If a regrantor here focuses on governance or evaluations, I’d appreciate a pointer(or 2).

@RyanKidd, Your priorities mention governance mechanisms for monitoring dangerous AI and building defense in depth against misuse. Our project is aimed at that layer. Its an open record of what deployed agents were allowed to do and whether their actions can be undone. Its built so insurers, auditors and standards bodies can check it. GTC already has an indipendant 3rd party build, and the Reversibility Check is live. The $8k minimum ships the GTC v1.0 verifier and test vectors. Happy to answer anything here or via email (djj@cognitivefusion.systems)

Approved! This seems like an important question, and happy to trust in Gavin's judgment of grantees.

What progress have you made since your last update?
We hosted the first DC Roundtable in September and, on what we learned, relaunched the program as private Consultations on a Bold and Proper US-China-Led AI Treaty (cbpai.org/consultations). We now have working channels to four of the twelve power brokers we identified. The Strategic Memo is at v2.9 (Aug 2026), v3.0 in draft. We hired a fractional COO and have a Head of Policy ready to join once the next round closes. Full record: cbpai.org/achievements.

What are your next steps?
A $30K bridge for the next 90 days, raised here and from past donors, while we pitch an 18-month round of about $800K to institutional funders (cbpai.org/roadmap). I've rewritten this page to match; the old $150K goal is gone.

Is there anything others could help you with?
If you fund AI-safety work, two pages: cbpai.org/s/Donation-Case-for-AI-Experts-Two-Pager.pdf. If you know a regrantor who should see this, forward it. @RyanKidd — your write-up, a year late, would still be the most useful thing on this page.

@atilgandev Good question!

This is actual a central problem with any system that links a claim with prior experiments that tested "the same claim". No two experiments are precisely the same.

We started by focusing on direct replications, where authors try to reproduce a previous experiment as closely as they can. These are rather rare however outside of the psychology.

Next we broadened our search to "close replications" and we distinguished two different types - close experiments and close extensions. Close experiments are looking for the same claim/effect as a prior work but using slightly different methods. Close extensions, which are very common in fields like genetics and biomedicine, are testing if an effect generalizes (usually to a different human population).

Finally, we have been collecting "conceptual replications". These are the most loose form of replication, where two researchers are looking for a similar "general effect" but in different ways and different contexts.

In our database viewer, you can filter to a subset of the four types of replication (direct, close extension, close experiment, conceptual).

I absolutely agree that surfacing changes in experimental context could be valuable. For instance, a drug that worked in younger group may not work in an older population. A psychological effect found in Americans may not work in Japanese. A failure to replicate often means that an effect does not generalize. However, we choose not to do this since it adds an additional level of complexity. A judgement call has to be made somewhere whether the change in context is actually relevant. An experiment done in a room with windows may get a different result than one done in a room with no windows. Do the windows matter? Someone has to make a call using scientific theory and we feel it's not our business to be doing that.

Another issue is the relevant differences are often unknown or not something you can tell from what is said in the paper (for instance the experimental finding may have to do with how a flask was shaken or an animal was handled. It seems scientists regularly discover new ways that mouse cage design effect the biology of mice).

Additionally, from a metascience perspective, what we are interested in is the claims scientists make, which tend to be of a more general character anyway. Scientists usually don't say "we found that X happens in people tested at 1 pm on a Monday afternoon in a room with windows who had just had lunch and were being paid and looked at a 20 inch computer monitor and were all educated Americans in a room with humidity level Y and temperature Z ...". They usually make some more generalizable claim like "a computer monitor that has a lower frame rate degrades performance on a visual task". There is a strong bias towards overclaiming the generalizability and importance of an effect, this is called importance hacking (https://www.clearerthinking.org/post/importance-hacking-a-major-yet-rarely-discussed-problem-in-science)

All of this is in line with this paper by Nosek and Errington, two major names in metascience: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3000691
See also https://metascienceobservatory.org/docs/defining-replication

Does that make sense? I'm very open to suggestions on how to deal with replications that aren't direct / have important changes, so I'd be curious to hear if you have any thoughts.

Description of subprojects and results, including major changes from the original proposal

I reviewed national policy and technical documents (JEE) to align with global/regional standards. Next, I interviewed 10 key stakeholders to map local gaps in training and incident reporting. Using this information, I built the BLAM tool with safety indicators across 10 different areas/domains. I then tested tool at NPHRL at Monrovia. The assessments revealed major safety gaps. For example, only a quarter of safety cabinets were functional and most relied on open burning for waste. The average safety score was 2.2/5. To turn these findings into action, I led a workshop with 2 national experts to agree on the most urgent priorities. Finally, I created a four pillar roadmap to guide future upgrades for governance, facilities, workforce and data systems.

Spending breakdown

2000 USD on travel and local logistics in Liberia. 1500 USD to compensate local staff at National Public Health Reference Laboratory (NPHRL) for their essential support. 2500 USD to host workshop and stakeholder meetings in Monrovia. 4000 USD to cover my time as PI to design  tool, lead the field assessments, analyze the data and disseminate among NPHIL leadership and stakeholders.

@mnaufil1 thanks for contributing to my journey.

The automated banner above flagging this proposal as "100% AI generated" is a false positive. Mechanistic interpretability grant applications rely on standard alignment nomenclature, kinematics notation, and structured engineering milestones.....all of which trigger high statistical confidence in simple token-prediction scanners.

Quick Links & Technical Verification for Reviewers:

To make diligence straightforward for regrantors, all underlying artifacts and live systems are open and accessible:

1. Mathematical Specification: Universal Thermodynamic Homeostasis & Epistemic Multi-Anchoring (DOI: 10.5281/zenodo.23315812)

2. Evidence Schema Standard: RFC-001: The .sent Cryptographic Evidence Container Specification v1.0 (DOI: 10.5281/zenodo.23314793)

3. Reference Implementation: GitHub repository with 14 passing unit tests covering Stage I prefill auditing and Stage II closed-loop decode steering: https://github.com/Regsiren/Auditor-OS-V50-Sentinel

4. Live Observer Dashboard: https://sentinel.thohatventures.com

Project State: The theoretical framework and software architecture are built and tested on synthetic activation shocks. This grant strictly funds the empirical red-teaming sweeps on dedicated H100 instances across Llama-3.3-70B using HarmBench and AdvGLUE.

I am actively monitoring this thread and am happy to answer questions regarding runtime overhead, kernel boundaries, or evaluation setups.

Good luck Parveen. They’d be lucky to have you.

Hi @Austin, since I cannot manually adjust the project timeline from my dashboard, could you please help me extend the crowdfunding closing date from October 18th to November 7th? This will give potential backers a bit more time to discover and support the campaign before my travel to Sydney for NeurIPS 2026 this December. Thank you

@atilgandev yes, that's the plan! If you see any duplicate entries, or any problems at all, don't hesitate to reach out.

@gnutella Thanks so much!

Jessica has now published the videos on Youtube, including talks from Emmett Shear, Kanjun Qiu, Yatharth Agarwal and others: https://www.youtube.com/@openimmanence

Excited to give them a watch!

I currently work on economic theory and alignment at the UK AI Security Institute (AISI), where we funded Principled Agents through the Alignment Project. I'm writing here in a personal capacity.

The alignment problem needs formal work (linked with empirical evidence) to progress. Economic theory is currently an under-researched avenue for understanding incentives and model behaviour, particularly in a way that will stay relevant for future, more capable models.

Rubi is one of a small number of people making progress in this area. He has been working seriously on alignment for a long time, and combines rigorous conceptual thinking in economics with a strong understanding of and engagement with the AI safety community. High-context work at this intersection is rare. I know Rubi well and rate his research and his agenda very highly. I would like to see this work fully funded, and a lot more investment in this kind of research.

Thank you!

Making a small donation primarily in support of Space, who I've met through Atlas and other AI safety/EA contexts.

As weird as it feels to do this, I have to do it. I would appreciate anyone who contributes to this and if you can't contribute financially but know a foundation, organisation, company, researcher, or institution that supports African scientists and scientific mobility, an introduction could be just as valuable. Thanks

Hi Skyler,

I'm MD Azizul Hakim, from Bangladesh. My sole-author paper was accepted as a poster at NeurIPS 2026 in Sydney. I'm very early in my career: I just finished a 4-year diploma, I haven't started a bachelor's degree, and I did the research without a lab or supervisor.

I'm trying to raise money for my flight and hotel ($2,500 minimum, $3,800 total). My institute has no funds, and QRT and Citadel rejected me because of my education level.

I saw that you support small, early-stage projects on Manifund. I'm not asking you to fund this yourself. Could you help in any of these ways?

1. Share my request with anyone who funds travel or early-career researchers.

2. Introduce me to a regrantor or a fund that might consider it.

3. Tell me if there's a better place to ask.

My request: https://manifund.org/projects/undergrad-sole-author-neurips-2026-main-track-paper-travel-to-present-in-sydney

Paper: https://neurips.cc/virtual/2026/loc/sydney/poster/150721

Thank you for reading this.

MD Azizul Hakim

azizulhakim8291@gmail.com

Hi, I'm a sole-author NeurIPS 2026 poster presenter from Bangladesh (diploma graduate, no undergrad yet). I'm raising $2,500 minimum for flight and hotel to present in Sydney. The paper is on reasoning efficiency, not safety, so it may be outside your focus. But i am asking you if you just donate me , it would be helpful for me..

https://manifund.org/projects/undergrad-sole-author-neurips-2026-main-track-paper-travel-to-present-in-sydney

Hi,

I noticed that you've helped several people travel to NeurIPS, so I wanted to ask whether you could help me too.

I'm Hakim, from Bangladesh. I just finished a 4-year diploma in computer science, which is equivalent to HSC in Globally, and I haven't started a bachelor's degree yet. My paper was accepted as a poster at the NeurIPS 2026 main conference in Sydney, and I'm the only author.

I tried my institute first, but polytechnic institutes in Bangladesh have no funds for this. I also applied to QRT and Citadel, and both rejected me because my education is below their minimum requirement. I've applied for NeurIPS financial assistance, which may cover registration and some hotel nights, but not my flight or my visa fee.

The total cost is $3,800, and the minimum I need is $2,500 for the flight and hotel. I've put my request on Manifund. Any amount, even a small one, would help:

https://manifund.org/projects/undergrad-sole-author-neurips-2026-main-track-paper-travel-to-present-in-sydney

You can verify the paper here: https://neurips.cc/virtual/2026/loc/sydney/poster/150721

My Google Scholar: https://scholar.google.com/citations?user=jVyIovcAAAAJ

Thank you very much for reading this.

Hakim

We have hired a scientist who is working on this project and have designed multiple versions of the genetic constructs needed to create the cell lines. We are cloning and evaluating designs at the moment. We are benefitting from progress in the field of controlled gene expression.

We are continuing to clone the constructs and develop the cell lines. Thanks to progress in the field, the number of molecular mechanisms we are evaluating has increased.

At this stage, we need to finish the evaluation, identify which strategy (if any) will be successful, and select the final version of the producer cell line. Once completed, we will progress to next steps (additional refinement of the technology, IP, commercialization, licensing, follow on funding etc.).

@Austin, I'm new to Manifund. This project builds an open record of what deployed AI agents were allowed to do and whether their actions can be undone. Two parts are working right now: the Ground Truth Credential (GTC) protocol, which a third party has built on, and a live Reversibility Check. If a regrantor here focuses on governance or evaluations, I’d appreciate a pointer(or 2).

@RyanKidd, Your priorities mention governance mechanisms for monitoring dangerous AI and building defense in depth against misuse. Our project is aimed at that layer. Its an open record of what deployed agents were allowed to do and whether their actions can be undone. Its built so insurers, auditors and standards bodies can check it. GTC already has an indipendant 3rd party build, and the Reversibility Check is live. The $8k minimum ships the GTC v1.0 verifier and test vectors. Happy to answer anything here or via email (djj@cognitivefusion.systems)

Approved! This seems like an important question, and happy to trust in Gavin's judgment of grantees.