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

@RyanKidd Many thanks!

I really like your approach on this matter. As someone who is interested in health tech, the focus on energy-bounded constraints (e.g., battery, torque) mirrors real-world healthcare resource limitations, offering foundational insights for future safety protocols in multi-agent medical systems. Well-executed.

This is a great initiative. This work provides a valuable framework for testing auditing robustness against strategically evasive model behaviors.

I think that the red-team approach can also directly address systemic risks in health AI systems where models might conceal unsafe behaviors while maintaining operational functionality which is a critical consideration for real-world clinical safety.

This is brilliant work, exactly the kind of practical, human-centered safety testing we need in health tech too. Your paired-scenario approach (unsafe vs. safe lookalikes) mirrors how we balance clinical efficiency with patient risk in AI-driven care.

By pinpointing where AI oversteps without becoming over-cautious, you’re building the trust infrastructure we desperately need for safe, useful AI systems. Keep pushing this frontier, every safe payment decision is a step toward healthier, more resilient systems.

Love the reel and the idea! This project has great potential, hope it reaches its funding goal. Good luck, John and team :)

Glad to answer any questions - today's decission to shut down Fable 5, shows that a general, non-agnostic, governance layer will be needed quite soon. It highlights the massive geopolitical and operational vulnerabilities of relying purely on centralized, platform-native AI governance.

@riyaneel I don't understand your answer. What are "strict geometric thresholds"? What conditions?

In order for you to know if it's worth your time to respond, I'll say that conditional on me knowing how it contributes to AI safety, there is around 5% of chance that I'm going to donate $1000 (otherwise, the chance is 0).

What progress have you made since your last update?

We have completed a screening of over 20 single, dual, and quad expression systems in two cell lines. We have identified several potential candidates for further study.

What are your next steps?

Work is currently being written up for publication.

Is there anything others could help you with?

Not at this time. After publication, this project will be closed and we will pursue further funding from federal and private grantmaking organizations.

Update: Proposal Concluded & Withdrawn

​The technical and computational bottlenecks outlined in this pitch have been successfully bypassed independently. The final phase of the ISB framework—including the 5th expansion (RNA velocity kinetics)—is now fully integrated, and the SDE-PDE simulation has been scaled and executed across a mega-cohort of N=100,000 virtual subjects.

​The comprehensive final manuscript is complete, and the entire production pipeline is permanently secured open-source on Zenodo and GitHub. External funding is no longer required.

Really excited to see this project come to fruition, and its impact!

@damc4 Current safety relies on behavioral patching to suppress a model's probabilistic urge to guess. CKSM handles safety at the base architectural level. By replacing probabilities with strict geometric thresholds, the model physically cannot compute a response if the conditions aren't met. It structurally abstains. Safety becomes a mathematical consequence of the system itself rather than a post-training filter.

The previous conference was incredibly meaningful to me. It led me to make some changes to my career plans based on 1-1 conversations and the event's sessions

@lisathiergart

How does that contribute to AI safety (if at all)?

Last event was great! I made some valuable connections and had very interesting chats 🙂

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

@RyanKidd Many thanks!

I really like your approach on this matter. As someone who is interested in health tech, the focus on energy-bounded constraints (e.g., battery, torque) mirrors real-world healthcare resource limitations, offering foundational insights for future safety protocols in multi-agent medical systems. Well-executed.

This is a great initiative. This work provides a valuable framework for testing auditing robustness against strategically evasive model behaviors.

I think that the red-team approach can also directly address systemic risks in health AI systems where models might conceal unsafe behaviors while maintaining operational functionality which is a critical consideration for real-world clinical safety.

This is brilliant work, exactly the kind of practical, human-centered safety testing we need in health tech too. Your paired-scenario approach (unsafe vs. safe lookalikes) mirrors how we balance clinical efficiency with patient risk in AI-driven care.

By pinpointing where AI oversteps without becoming over-cautious, you’re building the trust infrastructure we desperately need for safe, useful AI systems. Keep pushing this frontier, every safe payment decision is a step toward healthier, more resilient systems.

Love the reel and the idea! This project has great potential, hope it reaches its funding goal. Good luck, John and team :)

Glad to answer any questions - today's decission to shut down Fable 5, shows that a general, non-agnostic, governance layer will be needed quite soon. It highlights the massive geopolitical and operational vulnerabilities of relying purely on centralized, platform-native AI governance.

@riyaneel I don't understand your answer. What are "strict geometric thresholds"? What conditions?

In order for you to know if it's worth your time to respond, I'll say that conditional on me knowing how it contributes to AI safety, there is around 5% of chance that I'm going to donate $1000 (otherwise, the chance is 0).

What progress have you made since your last update?

We have completed a screening of over 20 single, dual, and quad expression systems in two cell lines. We have identified several potential candidates for further study.

What are your next steps?

Work is currently being written up for publication.

Is there anything others could help you with?

Not at this time. After publication, this project will be closed and we will pursue further funding from federal and private grantmaking organizations.

Update: Proposal Concluded & Withdrawn

​The technical and computational bottlenecks outlined in this pitch have been successfully bypassed independently. The final phase of the ISB framework—including the 5th expansion (RNA velocity kinetics)—is now fully integrated, and the SDE-PDE simulation has been scaled and executed across a mega-cohort of N=100,000 virtual subjects.

​The comprehensive final manuscript is complete, and the entire production pipeline is permanently secured open-source on Zenodo and GitHub. External funding is no longer required.

Really excited to see this project come to fruition, and its impact!

@damc4 Current safety relies on behavioral patching to suppress a model's probabilistic urge to guess. CKSM handles safety at the base architectural level. By replacing probabilities with strict geometric thresholds, the model physically cannot compute a response if the conditions aren't met. It structurally abstains. Safety becomes a mathematical consequence of the system itself rather than a post-training filter.

The previous conference was incredibly meaningful to me. It led me to make some changes to my career plans based on 1-1 conversations and the event's sessions

@lisathiergart

How does that contribute to AI safety (if at all)?

Last event was great! I made some valuable connections and had very interesting chats 🙂