I was looking through the GitHub case study and saw that Aura uses multi-view landmarks, depth and a statistical shape prior. Is the face mostly reconstructed from the photos, or is an ML model filling in parts of it? And what happens when the photos just don’t contain enough information for a region?

This is worth funding. Also, whoever is reading this, I am field building and funding people in the digital mind safety space and open to accepting money for regranting

@atilgandev Thanks for addressing the issue! Noting that it has now been resolved.

Thank you very much @dwks!

@carol Thanks for flagging this. I understand the issue, the activity has stopped, and it won’t happen again. I’d appreciate it if the moderation note could be removed or updated to reflect that the issue has been resolved

Hi Serhat — we’ve found that several accounts voting on your AURA project appear to be controlled by you or created on your behalf, so we’ve banned those accounts. Manifund relies on comments reflecting independent views from real people, and using additional accounts to boost a project or downvote other projects isn’t allowed. Your project can stay up for now, but any further sockpuppet activity will result in your account and project being removed.

Thank you Zahid @zahiidkbr

I've known Drishty for more than ten years, and I want to vouch for her from a different angle than the research itself. I am currently pursuing Masters of ICT Research from Melbourne Institute of Technology. I'm one of the people she's helping transition toward AI safety, she curated study materials for me personally, while dealing with circumstances that would have stopped most people entirely.

That's the pattern with her, life keeps throwing curveballs, and she keeps showing up with sheer grit and passion for her work and for other people's growth. The dedication you see on this page isn't a funding-pitch persona, it's just what she's like. A small grant here buys the gap between someone doing this work under crisis conditions and someone doing it properly resourced. She's one of the safest bets I know.

For Tamay Besiroglu (regrantor) — no public address found, so posting here:

Subject: Dated, signed ground truth for an eval: measuring correction latency in AI restatement of statistics

Hello Tamay — this is a measurement project more than a model project. A publisher's figures change on dated, logged occasions; I turned my own corrections log into three signed epochs (Rekor) and a task where the right answer depends on the date. First observation: one AI engine served a figure I had corrected 58 days earlier, because my own page still carried it.

I spent 23 years in bank credit; this year I ran the full 2025 HMDA file (1.19M FHA decisions) and then measured what language models do when they restate the figures: the number survives (27/27), the sentence around it does not (full contract credit 1–11 of 36 depending on what the publisher ships), and a verification token is forged in 2/36 even when every tool issues a real one. Everything is reproducible: open eval (Inspect + verifiers), deterministic rewards, claim receipts signed into Rekor, figures rebuilt from the CFPB file in an attested build, and a corrections log with my own thirteen errors. Project: https://manifund.org/projects/does-ai-restate-statistics-faithfully-a-self-verifying-eval--rl-environment . Asking $6K–18K for a second grader, three-repeat publication runs, and the temporal (correction-aware) task.

The Manifund page carries an AI-written flag; that is accurate, I draft in English with AI help and say so on the page. The work is verifiable without trusting the prose.

— Ziya Yetiş, FinanceRateCalc, Adana (press@financeratecalc.com). Written-only; I answer within a day.

This seems like the most neglected funding opportunity in animal advocacy right now, and I've generally been impressed by Marcus' independence of thought and levelheadedness.

Cost per Neel Nanda discovered at middle school

While I'm flattered, I'm not sure I follow the logic here?

@sibelzeynepumay Thank you very much. The key now is proving the accuracy independently.

@fattmaatilgan Thank you. That access gap is exactly what we’re trying to close.

I’m seeing the same issue on my proposal. Pangram flags 100% of my text as AI-written, although the content reflects my actual project and experience. I decided to leave the proposal as-is rather than rewrite it just to satisfy the detector. I’m curious whether this kind of false positive affects admin approval or only homepage visibility.

I have heard about AI Safety Nepal. I would love to support.

I hope you manage to raise the funding you need. I also think this could have real clinical value, because access to systems like 3dMD and VECTRA is still very limited. Even where I live, there is only one such system available.

I’m really impressed. If you can secure the funding and get this right, it could be nothing short of revolutionary. Systems like 3dMD and VECTRA are extremely expensive, and we’re talking about software that could potentially replace them.

Four audits in, the pattern is holding — and the standard just got sharper.

Audit #4 is done: Shao et al. 2025 (cultured networks, PLOS Comp Bio). The headline decoding result survived honest controls with caveats — the network genuinely reads stimulation patterns at 0.96/0.90 — but the claimed training improvement did not replicate. The d

@simonbgoulet Thank for your support!

Congratulations and good luck!!!

Approved! Happy to extend a lot of trust to our regrantors' judgment as always, and this project sounds like a cheap way to potentially produce useful insights.

I was looking through the GitHub case study and saw that Aura uses multi-view landmarks, depth and a statistical shape prior. Is the face mostly reconstructed from the photos, or is an ML model filling in parts of it? And what happens when the photos just don’t contain enough information for a region?

This is worth funding. Also, whoever is reading this, I am field building and funding people in the digital mind safety space and open to accepting money for regranting

@atilgandev Thanks for addressing the issue! Noting that it has now been resolved.

Thank you very much @dwks!

@carol Thanks for flagging this. I understand the issue, the activity has stopped, and it won’t happen again. I’d appreciate it if the moderation note could be removed or updated to reflect that the issue has been resolved

Hi Serhat — we’ve found that several accounts voting on your AURA project appear to be controlled by you or created on your behalf, so we’ve banned those accounts. Manifund relies on comments reflecting independent views from real people, and using additional accounts to boost a project or downvote other projects isn’t allowed. Your project can stay up for now, but any further sockpuppet activity will result in your account and project being removed.

Thank you Zahid @zahiidkbr

I've known Drishty for more than ten years, and I want to vouch for her from a different angle than the research itself. I am currently pursuing Masters of ICT Research from Melbourne Institute of Technology. I'm one of the people she's helping transition toward AI safety, she curated study materials for me personally, while dealing with circumstances that would have stopped most people entirely.

That's the pattern with her, life keeps throwing curveballs, and she keeps showing up with sheer grit and passion for her work and for other people's growth. The dedication you see on this page isn't a funding-pitch persona, it's just what she's like. A small grant here buys the gap between someone doing this work under crisis conditions and someone doing it properly resourced. She's one of the safest bets I know.

For Tamay Besiroglu (regrantor) — no public address found, so posting here:

Subject: Dated, signed ground truth for an eval: measuring correction latency in AI restatement of statistics

Hello Tamay — this is a measurement project more than a model project. A publisher's figures change on dated, logged occasions; I turned my own corrections log into three signed epochs (Rekor) and a task where the right answer depends on the date. First observation: one AI engine served a figure I had corrected 58 days earlier, because my own page still carried it.

I spent 23 years in bank credit; this year I ran the full 2025 HMDA file (1.19M FHA decisions) and then measured what language models do when they restate the figures: the number survives (27/27), the sentence around it does not (full contract credit 1–11 of 36 depending on what the publisher ships), and a verification token is forged in 2/36 even when every tool issues a real one. Everything is reproducible: open eval (Inspect + verifiers), deterministic rewards, claim receipts signed into Rekor, figures rebuilt from the CFPB file in an attested build, and a corrections log with my own thirteen errors. Project: https://manifund.org/projects/does-ai-restate-statistics-faithfully-a-self-verifying-eval--rl-environment . Asking $6K–18K for a second grader, three-repeat publication runs, and the temporal (correction-aware) task.

The Manifund page carries an AI-written flag; that is accurate, I draft in English with AI help and say so on the page. The work is verifiable without trusting the prose.

— Ziya Yetiş, FinanceRateCalc, Adana (press@financeratecalc.com). Written-only; I answer within a day.

This seems like the most neglected funding opportunity in animal advocacy right now, and I've generally been impressed by Marcus' independence of thought and levelheadedness.

Cost per Neel Nanda discovered at middle school

While I'm flattered, I'm not sure I follow the logic here?

@sibelzeynepumay Thank you very much. The key now is proving the accuracy independently.

@fattmaatilgan Thank you. That access gap is exactly what we’re trying to close.

I’m seeing the same issue on my proposal. Pangram flags 100% of my text as AI-written, although the content reflects my actual project and experience. I decided to leave the proposal as-is rather than rewrite it just to satisfy the detector. I’m curious whether this kind of false positive affects admin approval or only homepage visibility.

I have heard about AI Safety Nepal. I would love to support.

I hope you manage to raise the funding you need. I also think this could have real clinical value, because access to systems like 3dMD and VECTRA is still very limited. Even where I live, there is only one such system available.

I’m really impressed. If you can secure the funding and get this right, it could be nothing short of revolutionary. Systems like 3dMD and VECTRA are extremely expensive, and we’re talking about software that could potentially replace them.

Four audits in, the pattern is holding — and the standard just got sharper.

Audit #4 is done: Shao et al. 2025 (cultured networks, PLOS Comp Bio). The headline decoding result survived honest controls with caveats — the network genuinely reads stimulation patterns at 0.96/0.90 — but the claimed training improvement did not replicate. The d

@simonbgoulet Thank for your support!

Congratulations and good luck!!!

Approved! Happy to extend a lot of trust to our regrantors' judgment as always, and this project sounds like a cheap way to potentially produce useful insights.