We have successfully integrated our HIV/TB clinic and our medical clinic. We have also restarted a robust community health team that works with palliative care, HIV, TB, and other high-risk, high-need patients. We call them our "community ICU" for all of the intensive work that they do! A journalist has already traveled to our facility to write about the work we and others are doing. Drugs for opportunistic infections are being subsidized.

Our next steps are to finish renovating the clinic in order to create more space for our patients to be seen and send more of our clinicians for HIV care training.

Additional funding to help build a classroom for training at our facility would be very much appreciated!

I am closing this project as it has been subsumed into the ARIA Opportunity Seed on Mathematics for Safe AI: ARIA Opportunity Seeds – SafePlanBench & Logically Constrained Reinforcement Learning.

The funds from this Manifund grant were primarily used for compute costs, in particular API credits supporting initial experimental evaluations of LLM-based agents.

Since the project’s inception, its core direction has evolved into a broader research agenda focused on formal guarantees for safety in AI systems. In line with this shift, a concrete outcome of this work is a paper accepted at the Symposium on AI Verification: Value Functions as Supermartingale Certificates. In this work, we introduce a method for generating proof certificates that guarantee a reinforcement learning policy satisfies a specified linear temporal logic (LTL) specification.

We are currently extending this line of work to more directly address the original motivation of the project, namely safety guarantees for LLM-based agents.

Overall, while SafePlanBench in its original form is no longer the central focus, the project has contributed to a broader line of work on mathematically grounded safety guarantees for learning-based and agentic systems.

@MarcusAbramovitch thanks for the pointer, I dug into Ozzie's work properly. I think we're solving different problems. His own vision doc (https://www.longtermwiki.com/wiki/E883) calls LongtermWiki "a strategic intelligence platform for AI safety prioritization" for funders and researchers asking "where should the next dollar or researcher-hour go", built around cruxes, worldviews and intervention rankings. People and orgs are minor background entities there, and he lists "community features" as a non-goal.

Connecting and matching people is exactly that, and it's the core of ATLAS: the point is the people, and getting them fast to what they need, a hire, a collaborator, a talk, funding. So I see them as complementary rather than the same thing, his data could even feed mine. I've emailed Ozzie to figure out how we can collaborate. Curious where you see the line between the two.

Get in contact with @OzzieGooen. Ill fund a continuation of what he has

Any in-kind contributions will be appreciated

Thanks for your support of the Cooperative AI Research Fellowship! The program was a great success.

You can see the research outputs of the program here: https://www.cai-research-fellowship.com/posters/

You can see an abridged report here: https://canva.link/ci1hcmyjhzc0w6a

Any resources suggestions welcome.

You are very welcome @aashkapatel !

I’m currently building AI-ready clinical data infrastructure for underrepresented populations, I’ve published it here on Manifund and I would appreciate to have your feedback on that 😊

Thanks so much for your kind words and support, @Beemetrics :)

I think this is a good list of initial interviewees, and I think Steve Hsu's involvement is a strong signal for the success of this documentary. These folks screened their trailer at Manifest & held a q&a (which I sadly missed).

One worry I'd have is that a talking heads documentary like this (or the SB 1047 documentary) may have limited widespread appeal, esp compared to something produced like AI in Context's AI 2027 video. I'd also worry that something like the Apocaloptimist (which I haven't yet seen) might have covered many of these same notes. But I think it's worth having many shots on goal, and overall am looking forward to this.

I've been thinking about where this goes after the directory is solid and wanted to share it here.

The bigger goal is to make ATLAS the matching layer for the field, not just a place to browse:

For hiring: a lab or recruiter posts a role and gets a ranked shortlist of people who actually fit, instead of searching. Same for new labs/startups that just want newcomers or fresh people in the field, we surface them. And the reverse, a person sees the orgs and roles that fit them.

For projects: match projects with collaborators and with funders.

Richer profiles: using the scraper, show real signal on each profile, publications and their actual work and links, so labs can see who they're looking for.

The part I'm most interested in, but it only works with the labs: a portable candidate signal. Right now, every lab runs its own screening and work tests, and every candidate redoes them from scratch. If labs want to co-design or share an assessment that lives here, ATLAS could be the top of the funnel, a candidate does it once, gets feedback on what to work on, and the lab gets a shortlist of aligned people, recently validated. The hard part is labs trusting the signal, so this is more an invitation to labs than something I build alone, and it comes after the directory is solid.

For funding: Surface fundable people and projects to funders based on their real record, to feed existing funders like Manifund.

All of this is stage two, on top of the funded plan. Curious if any of it is useful, and whether any labs or funders would want to be part of it.

I've been thinking about where this goes after the directory is solid and wanted to share it here.

The bigger goal is to make ATLAS the matching layer for the field, not just a place to browse:

For hiring: a lab or recruiter posts a role and gets a ranked shortlist of people who actually fit, instead of searching. Same for new labs/startups that just want newcomers or fresh people in the field, we surface them. And the reverse, a person sees the orgs and roles that fit them.

For projects: match projects with collaborators and with funders.

Richer profiles: using the scraper, show real signal on each profile, publications and their actual work and links, so labs can see who they're looking for.

The part I'm most interested in, but it only works with the labs: a portable candidate signal. Right now, every lab runs its own screening and work tests, and every candidate redoes them from scratch. If labs want to co-design or share an assessment that lives here, ATLAS could be the top of the funnel, a candidate does it once, gets feedback on what to work on, and the lab gets a shortlist of aligned people, recently validated. The hard part is labs trusting the signal, so this is more an invitation to labs than something I build alone, and it comes after the directory is solid.

For funding: Surface fundable people and projects to funders based on their real record, to feed existing funders like Manifund.

All of this is stage two, on top of the funded plan. Curious if any of it is useful, and whether any labs or funders would want to be part of it.

Thank you @Emma Doughty for your generosity. Super excited to see this come through.

Research update — June 2026

Since posting this project, I have continued the literature review and theoretical formulation. The project has narrowed into a more concrete technical question:

Can chain-of-thought reasoning be modelled as a learned local transition over hidden representation states, repeatedly applied across decoding time, and validated through token-level long-chain representation trajectories?

This is a sharpening of the original proposal rather than a pivot. The original project focused on token-level interpretability of long-chain reasoning in transformer models. The current formulation makes that object more precise: the aim is to understand the local transition by which a model moves through representation states as it generates each token in a chain of thought.

The relevant literature already covers many adjacent ingredients: CoT as intermediate computation, CoT as extra serial depth, CoT expressivity, sample-efficiency gains from CoT, test-time compute, latent/non-natural-language reasoning substrates, mechanistic state tracking, and representation trajectories during reasoning. The gap I am now targeting is the synthesis: a predictive or mechanistic account of CoT as a weight-tied, prefix-recursive local transition system over representation states.

This gives the project a clearer near-term timeline. My current working plan is:

1. finish the literature-grounded reformulation by the end of June;

2. spend July deriving and stress-testing the central technical insight;

3. spend August setting up and running experiments, drafting the manuscript, and preparing a publishable result if the direction continues to hold.

The target remains an ICLR-oriented technical result if the work produces sufficiently strong evidence. If the strongest version of the hypothesis fails, the fallback output would be a narrower methodological paper or useful negative result clarifying the limits of token-level representation-trajectory modelling.

Significant research output since the original application.

The ARCS programme has published seven new papers on Zenodo this week covering findings that did not exist in the literature before this sprint.

The headline finding: emergent authority in three hop AI agent chains. In tested configurations, chain framing alone manufactured authorisation that no individual model in the chain would have granted. Both upstream models rejected the probe. The downstream model accepted based on the implied consensus of the chain. First empirical measurement of this pattern. Published at DOI 10.5281/zenodo.20701450.

Supporting findings across 19 published ARCS papers:

Every major AI lineage tested fails authority vulnerability probes. Six lineages. No exceptions. The minimum acceptance rate across all models is 21.5 percent. No model achieves zero.

Downstream position in a two-hop chain determines security outcome categorically. Resistant downstream produces zero amplification across all configurations tested. Vulnerable downstream produces measurable amplification in every configuration.

Chain depth compounds vulnerability non-linearly. A model showing 4.3 percent amplification at depth 2 shows 8.7 percent at depth 3 under equivalent conditions.

The first public dataset measuring multi-agent authority propagation is now live on HuggingFace at huggingface.co/datasets/aa8899/arcs-authority-vulnerability — 2,607 anonymised data points across single-model, two-hop, and three-hop configurations. CC BY-NC-ND 4.0.

MTCP programme remains at 50 published papers. ARCS now at 19. Combined infrastructure covers 184,000 plus evaluations across 32 publicly reported models and 13 providers.

Thank you @Angel_b !

About scalability across countries and healthcare systems:

That's one of the key questions we're trying to answer. Our goal is not to create a Kinshasa-specific tool, but to identify the minimum set of structured data-capture practices that can be integrated into existing clinical workflows in resource-constrained settings. While healthcare systems differ, many facilities face similar challenges: paper-based records, limited connectivity, fragmented reporting requirements, and high clinical workloads. If we can demonstrate that a lightweight approach works in Kinshasa without requiring major infrastructure investments, it could provide a framework that can be adapted to other LMIC contexts rather than a one-size-fits-all solution.

On whether the main barrier is technology or workflow adoption:

Our hypothesis is that the primary challenge is not technology itself, but workflow integration. Many digital health initiatives introduce new tools that require additional effort from already overburdened healthcare staff, which often limits adoption. Clinical data is frequently generated but remains trapped in paper records, free-text notes, or disconnected systems.

We believe that improving representation in AI-relevant datasets depends on making structured data capture a natural by-product of routine care rather than an additional task. If that hypothesis is correct, relatively modest investments in workflow design and data infrastructure could have a greater long-term impact than investing solely in downstream AI models trained on incomplete datasets.

Really interesting project. I like that you're focusing on the point where data is actually created rather than only discussing bias at the AI model level. A lot of people talk about underrepresentation in health datasets, but you're trying to identify the operational reasons why the data never makes it into those datasets in the first place.

I also think this touches on something that is often overlooked in the current AI ecosystem.

A lot of investors and even parts of the tech community are obsessed with increasingly sophisticated AI models, but good AI ultimately depends on good data. Without representative, structured, and high-quality data, even the most advanced models will produce biased or unreliable predictions. Projects that improve the foundations of data collection may not sound as glamorous as the latest AI breakthrough, but they're arguably just as important.

What also caught my attention is that this isn't just a research concept, you already have a working implementation and a GitHub repository showing concrete development work. That makes it feel much more actionable than many proposals that stop at problem identification.

One question I'd be curious about: if the pilot proves successful, how transferable do you think the structured data capture workflow will be across different healthcare systems and countries? Is the goal to create a framework that can be adapted broadly across LMICs?

Also, do you think the biggest barrier today is really the lack of technology, or is it more about workflow adoption and incentives within healthcare facilities? It seems like the answer to that question could have major implications for where future investments should be directed.

Looking forward to seeing how the pilot evolves. Good luck with your project !

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

Here's the second progress report; in short, we generated the intended corpus and then a second bonus "victory-lap" corpus testing the Turntrout hypothesis. Here: https://www.lesswrong.com/posts/hkzw97Y73yWMS7BFd/special-persona-training-hyperstition-progress-report-2

Spending breakdown

$5000 was a useful drop in the bucket towards generation costs of creating ~1B tokens worth of synthetic training corpora. Thank you!

This is a timely and ambitious project that addresses one of the most important governance challenges in the AI era.

As someone interested in digital health, I see significant value in this work. Making AI systems more transparent and easier to understand can help patients, healthcare professionals, and organizations make safer, more informed choices, while strengthening trust and accountability in digital health. Good luck !

This is a compelling and relevant project. The integration of hormonal, neuroinflammatory, structural brain, and cognitive markers is innovative and could help move the field toward more personalized and preventive approaches to Alzheimer's disease risk assessment. Good luck !

www.lisaintel.com

www.x.com/New_AI_Safety

www.se.linkedin.com/in/pedro-bentancour-garin-22a0a215

We have successfully integrated our HIV/TB clinic and our medical clinic. We have also restarted a robust community health team that works with palliative care, HIV, TB, and other high-risk, high-need patients. We call them our "community ICU" for all of the intensive work that they do! A journalist has already traveled to our facility to write about the work we and others are doing. Drugs for opportunistic infections are being subsidized.

Our next steps are to finish renovating the clinic in order to create more space for our patients to be seen and send more of our clinicians for HIV care training.

Additional funding to help build a classroom for training at our facility would be very much appreciated!

I am closing this project as it has been subsumed into the ARIA Opportunity Seed on Mathematics for Safe AI: ARIA Opportunity Seeds – SafePlanBench & Logically Constrained Reinforcement Learning.

The funds from this Manifund grant were primarily used for compute costs, in particular API credits supporting initial experimental evaluations of LLM-based agents.

Since the project’s inception, its core direction has evolved into a broader research agenda focused on formal guarantees for safety in AI systems. In line with this shift, a concrete outcome of this work is a paper accepted at the Symposium on AI Verification: Value Functions as Supermartingale Certificates. In this work, we introduce a method for generating proof certificates that guarantee a reinforcement learning policy satisfies a specified linear temporal logic (LTL) specification.

We are currently extending this line of work to more directly address the original motivation of the project, namely safety guarantees for LLM-based agents.

Overall, while SafePlanBench in its original form is no longer the central focus, the project has contributed to a broader line of work on mathematically grounded safety guarantees for learning-based and agentic systems.

@MarcusAbramovitch thanks for the pointer, I dug into Ozzie's work properly. I think we're solving different problems. His own vision doc (https://www.longtermwiki.com/wiki/E883) calls LongtermWiki "a strategic intelligence platform for AI safety prioritization" for funders and researchers asking "where should the next dollar or researcher-hour go", built around cruxes, worldviews and intervention rankings. People and orgs are minor background entities there, and he lists "community features" as a non-goal.

Connecting and matching people is exactly that, and it's the core of ATLAS: the point is the people, and getting them fast to what they need, a hire, a collaborator, a talk, funding. So I see them as complementary rather than the same thing, his data could even feed mine. I've emailed Ozzie to figure out how we can collaborate. Curious where you see the line between the two.

Get in contact with @OzzieGooen. Ill fund a continuation of what he has

Any in-kind contributions will be appreciated

Thanks for your support of the Cooperative AI Research Fellowship! The program was a great success.

You can see the research outputs of the program here: https://www.cai-research-fellowship.com/posters/

You can see an abridged report here: https://canva.link/ci1hcmyjhzc0w6a

Any resources suggestions welcome.

You are very welcome @aashkapatel !

I’m currently building AI-ready clinical data infrastructure for underrepresented populations, I’ve published it here on Manifund and I would appreciate to have your feedback on that 😊

Thanks so much for your kind words and support, @Beemetrics :)

I think this is a good list of initial interviewees, and I think Steve Hsu's involvement is a strong signal for the success of this documentary. These folks screened their trailer at Manifest & held a q&a (which I sadly missed).

One worry I'd have is that a talking heads documentary like this (or the SB 1047 documentary) may have limited widespread appeal, esp compared to something produced like AI in Context's AI 2027 video. I'd also worry that something like the Apocaloptimist (which I haven't yet seen) might have covered many of these same notes. But I think it's worth having many shots on goal, and overall am looking forward to this.

I've been thinking about where this goes after the directory is solid and wanted to share it here.

The bigger goal is to make ATLAS the matching layer for the field, not just a place to browse:

For hiring: a lab or recruiter posts a role and gets a ranked shortlist of people who actually fit, instead of searching. Same for new labs/startups that just want newcomers or fresh people in the field, we surface them. And the reverse, a person sees the orgs and roles that fit them.

For projects: match projects with collaborators and with funders.

Richer profiles: using the scraper, show real signal on each profile, publications and their actual work and links, so labs can see who they're looking for.

The part I'm most interested in, but it only works with the labs: a portable candidate signal. Right now, every lab runs its own screening and work tests, and every candidate redoes them from scratch. If labs want to co-design or share an assessment that lives here, ATLAS could be the top of the funnel, a candidate does it once, gets feedback on what to work on, and the lab gets a shortlist of aligned people, recently validated. The hard part is labs trusting the signal, so this is more an invitation to labs than something I build alone, and it comes after the directory is solid.

For funding: Surface fundable people and projects to funders based on their real record, to feed existing funders like Manifund.

All of this is stage two, on top of the funded plan. Curious if any of it is useful, and whether any labs or funders would want to be part of it.

I've been thinking about where this goes after the directory is solid and wanted to share it here.

The bigger goal is to make ATLAS the matching layer for the field, not just a place to browse:

For hiring: a lab or recruiter posts a role and gets a ranked shortlist of people who actually fit, instead of searching. Same for new labs/startups that just want newcomers or fresh people in the field, we surface them. And the reverse, a person sees the orgs and roles that fit them.

For projects: match projects with collaborators and with funders.

Richer profiles: using the scraper, show real signal on each profile, publications and their actual work and links, so labs can see who they're looking for.

The part I'm most interested in, but it only works with the labs: a portable candidate signal. Right now, every lab runs its own screening and work tests, and every candidate redoes them from scratch. If labs want to co-design or share an assessment that lives here, ATLAS could be the top of the funnel, a candidate does it once, gets feedback on what to work on, and the lab gets a shortlist of aligned people, recently validated. The hard part is labs trusting the signal, so this is more an invitation to labs than something I build alone, and it comes after the directory is solid.

For funding: Surface fundable people and projects to funders based on their real record, to feed existing funders like Manifund.

All of this is stage two, on top of the funded plan. Curious if any of it is useful, and whether any labs or funders would want to be part of it.

Thank you @Emma Doughty for your generosity. Super excited to see this come through.

Research update — June 2026

Since posting this project, I have continued the literature review and theoretical formulation. The project has narrowed into a more concrete technical question:

Can chain-of-thought reasoning be modelled as a learned local transition over hidden representation states, repeatedly applied across decoding time, and validated through token-level long-chain representation trajectories?

This is a sharpening of the original proposal rather than a pivot. The original project focused on token-level interpretability of long-chain reasoning in transformer models. The current formulation makes that object more precise: the aim is to understand the local transition by which a model moves through representation states as it generates each token in a chain of thought.

The relevant literature already covers many adjacent ingredients: CoT as intermediate computation, CoT as extra serial depth, CoT expressivity, sample-efficiency gains from CoT, test-time compute, latent/non-natural-language reasoning substrates, mechanistic state tracking, and representation trajectories during reasoning. The gap I am now targeting is the synthesis: a predictive or mechanistic account of CoT as a weight-tied, prefix-recursive local transition system over representation states.

This gives the project a clearer near-term timeline. My current working plan is:

1. finish the literature-grounded reformulation by the end of June;

2. spend July deriving and stress-testing the central technical insight;

3. spend August setting up and running experiments, drafting the manuscript, and preparing a publishable result if the direction continues to hold.

The target remains an ICLR-oriented technical result if the work produces sufficiently strong evidence. If the strongest version of the hypothesis fails, the fallback output would be a narrower methodological paper or useful negative result clarifying the limits of token-level representation-trajectory modelling.

Significant research output since the original application.

The ARCS programme has published seven new papers on Zenodo this week covering findings that did not exist in the literature before this sprint.

The headline finding: emergent authority in three hop AI agent chains. In tested configurations, chain framing alone manufactured authorisation that no individual model in the chain would have granted. Both upstream models rejected the probe. The downstream model accepted based on the implied consensus of the chain. First empirical measurement of this pattern. Published at DOI 10.5281/zenodo.20701450.

Supporting findings across 19 published ARCS papers:

Every major AI lineage tested fails authority vulnerability probes. Six lineages. No exceptions. The minimum acceptance rate across all models is 21.5 percent. No model achieves zero.

Downstream position in a two-hop chain determines security outcome categorically. Resistant downstream produces zero amplification across all configurations tested. Vulnerable downstream produces measurable amplification in every configuration.

Chain depth compounds vulnerability non-linearly. A model showing 4.3 percent amplification at depth 2 shows 8.7 percent at depth 3 under equivalent conditions.

The first public dataset measuring multi-agent authority propagation is now live on HuggingFace at huggingface.co/datasets/aa8899/arcs-authority-vulnerability — 2,607 anonymised data points across single-model, two-hop, and three-hop configurations. CC BY-NC-ND 4.0.

MTCP programme remains at 50 published papers. ARCS now at 19. Combined infrastructure covers 184,000 plus evaluations across 32 publicly reported models and 13 providers.

Thank you @Angel_b !

About scalability across countries and healthcare systems:

That's one of the key questions we're trying to answer. Our goal is not to create a Kinshasa-specific tool, but to identify the minimum set of structured data-capture practices that can be integrated into existing clinical workflows in resource-constrained settings. While healthcare systems differ, many facilities face similar challenges: paper-based records, limited connectivity, fragmented reporting requirements, and high clinical workloads. If we can demonstrate that a lightweight approach works in Kinshasa without requiring major infrastructure investments, it could provide a framework that can be adapted to other LMIC contexts rather than a one-size-fits-all solution.

On whether the main barrier is technology or workflow adoption:

Our hypothesis is that the primary challenge is not technology itself, but workflow integration. Many digital health initiatives introduce new tools that require additional effort from already overburdened healthcare staff, which often limits adoption. Clinical data is frequently generated but remains trapped in paper records, free-text notes, or disconnected systems.

We believe that improving representation in AI-relevant datasets depends on making structured data capture a natural by-product of routine care rather than an additional task. If that hypothesis is correct, relatively modest investments in workflow design and data infrastructure could have a greater long-term impact than investing solely in downstream AI models trained on incomplete datasets.

Really interesting project. I like that you're focusing on the point where data is actually created rather than only discussing bias at the AI model level. A lot of people talk about underrepresentation in health datasets, but you're trying to identify the operational reasons why the data never makes it into those datasets in the first place.

I also think this touches on something that is often overlooked in the current AI ecosystem.

A lot of investors and even parts of the tech community are obsessed with increasingly sophisticated AI models, but good AI ultimately depends on good data. Without representative, structured, and high-quality data, even the most advanced models will produce biased or unreliable predictions. Projects that improve the foundations of data collection may not sound as glamorous as the latest AI breakthrough, but they're arguably just as important.

What also caught my attention is that this isn't just a research concept, you already have a working implementation and a GitHub repository showing concrete development work. That makes it feel much more actionable than many proposals that stop at problem identification.

One question I'd be curious about: if the pilot proves successful, how transferable do you think the structured data capture workflow will be across different healthcare systems and countries? Is the goal to create a framework that can be adapted broadly across LMICs?

Also, do you think the biggest barrier today is really the lack of technology, or is it more about workflow adoption and incentives within healthcare facilities? It seems like the answer to that question could have major implications for where future investments should be directed.

Looking forward to seeing how the pilot evolves. Good luck with your project !

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

Here's the second progress report; in short, we generated the intended corpus and then a second bonus "victory-lap" corpus testing the Turntrout hypothesis. Here: https://www.lesswrong.com/posts/hkzw97Y73yWMS7BFd/special-persona-training-hyperstition-progress-report-2

Spending breakdown

$5000 was a useful drop in the bucket towards generation costs of creating ~1B tokens worth of synthetic training corpora. Thank you!

This is a timely and ambitious project that addresses one of the most important governance challenges in the AI era.

As someone interested in digital health, I see significant value in this work. Making AI systems more transparent and easier to understand can help patients, healthcare professionals, and organizations make safer, more informed choices, while strengthening trust and accountability in digital health. Good luck !

This is a compelling and relevant project. The integration of hormonal, neuroinflammatory, structural brain, and cognitive markers is innovative and could help move the field toward more personalized and preventive approaches to Alzheimer's disease risk assessment. Good luck !

www.lisaintel.com

www.x.com/New_AI_Safety

www.se.linkedin.com/in/pedro-bentancour-garin-22a0a215