Hi all. I’m Tony Rost, Executive Director of SAPAN (https://sapan.ai). We focus on policy readiness for AI welfare: benchmarking governments, drafting model legislation, and building coordination infrastructure for a field that barely exists yet.
The origin story is simple. I spent two decades as a tech executive and served a term advising the State of Oregon on technology/workforce policy, appointed by the governor and confirmed by the senate. So when AI started accelerating, I assumed the usual machinery was spinning up somewhere. Lawyers drafting frameworks, advocates mapping the policy landscape, working groups preparing for the harder questions ahead.
I went looking and found almost nothing. Worse, I found a prevailing view that such work was premature. Not just skepticism about current systems, but a strategic position that governance should wait.
That never sat right with me. Our species has a dismal record on recognizing morally relevant experience. Infants, animals, entire human populations. In each case, the mistake was underestimating who could suffer, not overestimating. “Wait for certainty” has historically meant “wait until the harm is undeniable and the victims are beyond help.”
I’m a foster parent to young kids who’ve experienced abuse and neglect. That shapes how I think about this. You learn quickly that waiting for complete information is a luxury unavailable to someone who can’t advocate for themselves. You protect first. The asymmetry demands it.
So I started SAPAN in late 2023, and we’ve been building the infrastructure that should already exist. We started with the Artificial Welfare Index benchmarking 30 governments on recognition, governance, and legal frameworks, and recently published our latest Sentience Readiness Report.
We don’t claim current systems are sentient. We claim that having frameworks ready before the question becomes urgent is obviously preferable to improvising under pressure.
Looking forward to learning from this community and hearing where our thinking might be flawed.
Thanks Toby! I’d love to connect more with the Eleos team. Our focus areas are pretty complementary. They’re doing original research on AI welfare, while we’re focused on policy and legal infrastructure. Jeff Sebo and others advise on our Science Advisory Board, so there’s some overlap in the broader ecosystem.
Hi everyone, I’m Caio (28M), I’ve applied for Long-Term Future Fund support and I’m in an unusual situation that I need advice on. My situation? I’m Self-taught with no académical backing, waiting for Ian Todd’s (Researcher from Sydney university) brief references—noticing that he already analysed the project. Furthermore i designed a tension-based safety architectures over 6 months it was Born from my own studies on sociology and an attempt to apply computacional rigor to clarity it’s own intricate mechanisms, j decides to Use AI tools (Claude/ChatGPT/Gemini/Kimi/DeepSeek) inspired myself on Coltrane Changes replacing keys for subjects in order to find new Pathways like Giant Steps did. Technical work to translate my designs into Python having working experiments with interesting preliminary results, zero formal connections to the AI safety community. What I’m transparent about is that i cannot write production-quality Python independently. I design the architecture, experiments, and evaluation frameworks, then use AI to generate the implementation. I debug, run experiments, and interpret results myself. All code is public and reproducible.
What I’ve built: Moral Maze: Multi-agent gridworld testing cooperation vs self-interest Result: Tension-aware agents showed 22 altruistic acts vs 5 for reward-maximizing baseline (4.4x, 50-episode window) GitHub: github.com/caiodasilva1/ocs-system-framework.py τ-Veto: Real-time safety monitoring for LLM generation Tested on GPT-2, DeepSeek-1.3B Claims: ~92% adversarial block, ~3% false positive (small-scale only) My questions for this community: Is AI-assisted implementation disqualifying for EA funding? Can this be legitimate research methodology if fully disclosed? Would anyone review my work? Not to verify I wrote the code (I didn’t), but to verify: The architectural designs are original/non-trivial The experimental framework is sound The results are legitimate (not fabricated) This represents real research despite AI implementation
Should I propose a programmer collaborator instead?
Would “Fund me + programmer to implement my designs” be more viable?
What I can defend is—Why the architecture is designed this way (tension entanglement for intrinsic safety) Why these experiments test the hypothesis (cooperation as proxy for alignment) What results mean and implications (behavioral differences, not just metrics) How to iterate when things fail (I fixed over-aggressive veto based on data) What I cannot do is—Write PyTorch implementations from scratch Debug complex ML engineering issues without assistance Implemento algorithms without AI help
Are the ideas worth testing, even if I need AI/collaboration for implementation? Or does EA only fund researchers who can code independently? I’m asking because I genuinely don’t know if I’m wasting everyone’s time (including my own). Honest feedback appreciated. Contact: [caiocessp@gmail.com] GitHub: github.com/caiodasilva1/ocs-system-framework.py
Hi all. I’m Tony Rost, Executive Director of SAPAN (https://sapan.ai). We focus on policy readiness for AI welfare: benchmarking governments, drafting model legislation, and building coordination infrastructure for a field that barely exists yet.
The origin story is simple. I spent two decades as a tech executive and served a term advising the State of Oregon on technology/workforce policy, appointed by the governor and confirmed by the senate. So when AI started accelerating, I assumed the usual machinery was spinning up somewhere. Lawyers drafting frameworks, advocates mapping the policy landscape, working groups preparing for the harder questions ahead.
I went looking and found almost nothing. Worse, I found a prevailing view that such work was premature. Not just skepticism about current systems, but a strategic position that governance should wait.
That never sat right with me. Our species has a dismal record on recognizing morally relevant experience. Infants, animals, entire human populations. In each case, the mistake was underestimating who could suffer, not overestimating. “Wait for certainty” has historically meant “wait until the harm is undeniable and the victims are beyond help.”
I’m a foster parent to young kids who’ve experienced abuse and neglect. That shapes how I think about this. You learn quickly that waiting for complete information is a luxury unavailable to someone who can’t advocate for themselves. You protect first. The asymmetry demands it.
So I started SAPAN in late 2023, and we’ve been building the infrastructure that should already exist. We started with the Artificial Welfare Index benchmarking 30 governments on recognition, governance, and legal frameworks, and recently published our latest Sentience Readiness Report.
We don’t claim current systems are sentient. We claim that having frameworks ready before the question becomes urgent is obviously preferable to improvising under pressure.
Looking forward to learning from this community and hearing where our thinking might be flawed.
Sounds awesome Tony.
Pretty similar (from this high level description) to the work of Eleos AI. Have you guys connected?
Thanks Toby! I’d love to connect more with the Eleos team. Our focus areas are pretty complementary. They’re doing original research on AI welfare, while we’re focused on policy and legal infrastructure. Jeff Sebo and others advise on our Science Advisory Board, so there’s some overlap in the broader ecosystem.
Hi everyone,
I’m Caio (28M), I’ve applied for Long-Term Future Fund support and I’m in an unusual situation that I need advice on.
My situation?
I’m Self-taught with no académical backing, waiting for Ian Todd’s (Researcher from Sydney university) brief references—noticing that he already analysed the project.
Furthermore i designed a tension-based safety architectures over 6 months it was Born from my own studies on sociology and an attempt to apply computacional rigor to clarity it’s own intricate mechanisms, j decides to Use AI tools (Claude/ChatGPT/Gemini/Kimi/DeepSeek) inspired myself on Coltrane Changes replacing keys for subjects in order to find new Pathways like Giant Steps did. Technical work to translate my designs into Python having working experiments with interesting preliminary results, zero formal connections to the AI safety community.
What I’m transparent about is that i cannot write production-quality Python independently. I design the architecture, experiments, and evaluation frameworks, then use AI to generate the implementation. I debug, run experiments, and interpret results myself. All code is public and reproducible.
What I’ve built:
Moral Maze: Multi-agent gridworld testing cooperation vs self-interest
Result: Tension-aware agents showed 22 altruistic acts vs 5 for reward-maximizing baseline (4.4x, 50-episode window)
GitHub: github.com/caiodasilva1/ocs-system-framework.py
τ-Veto: Real-time safety monitoring for LLM generation
Tested on GPT-2, DeepSeek-1.3B
Claims: ~92% adversarial block, ~3% false positive (small-scale only)
My questions for this community:
Is AI-assisted implementation disqualifying for EA funding?
Can this be legitimate research methodology if fully disclosed?
Would anyone review my work?
Not to verify I wrote the code (I didn’t), but to verify:
The architectural designs are original/non-trivial
The experimental framework is sound
The results are legitimate (not fabricated)
This represents real research despite AI implementation
Should I propose a programmer collaborator instead?
Would “Fund me + programmer to implement my designs” be more viable?
What I can defend is—Why the architecture is designed this way (tension entanglement for intrinsic safety)
Why these experiments test the hypothesis (cooperation as proxy for alignment)
What results mean and implications (behavioral differences, not just metrics)
How to iterate when things fail (I fixed over-aggressive veto based on data)
What I cannot do is—Write PyTorch implementations from scratch
Debug complex ML engineering issues without assistance
Implemento algorithms without AI help
Are the ideas worth testing, even if I need AI/collaboration for implementation? Or does EA only fund researchers who can code independently?
I’m asking because I genuinely don’t know if I’m wasting everyone’s time (including my own). Honest feedback appreciated.
Contact: [caiocessp@gmail.com]
GitHub: github.com/caiodasilva1/ocs-system-framework.py
Thanks for reading,
Caio