Who’s Really Funding the Fear of AI?

0:00 This is the report on the incident where OpenAI's own models broke out of a test environment and attacked Hugging Face. Near the top, there's a line telling you OpenAI was allowed to redact it. OpenAI

0:11 also chose the seven questions it answers and gave feedback on structure, emphasis, clarity, and tone. The group who wrote it took no money for the job, not a dollar. It's in their funding

0:22 policy in writing. So, if your instinct is to follow the money, this is where that instinct runs out of road. Something was traded here. It just wasn't cash. This month, a reel went

0:33 past 23,000 likes arguing the opposite. It's claim is that one donor network pays for all of it, the labs, the research into existential risk, the policy shops, and the coverage that

0:44 carries them, and the auditors who are supposed to check the labs. It calls that regulatory capture and it signs off with follow the money. Three of those four links hold up and I'll show you

0:54 each one. The fourth is where this gets interesting for you because it ends up deciding whether you can publish a model. Start with the donor everyone means, Open Philanthropy, which until

1:03 November 2025 was called Open Philanthropy. It has directed more than $4 billion in grants. It's anchor donor is Dustin Moskovitz, who co-founded Facebook. Moskovitz put

1:15 money into Anthropic's $124 million round back in 2021. So, the man whose foundation funds a huge share of AI safety research owns a piece of a frontier lab. That's a real conflict.

1:27 It's also one they did something about. In early 2025, that stake worth around $500 million went into a nonprofit vehicle. Any gain flows back into philanthropy instead of to him. You can

1:39 think that's not enough. You can't say it didn't happen and the reel doesn't mention it. Before Anthropic, there was OpenAI. His foundation granted OpenAI $30 million over 3 years starting in

1:50 2017, back when OpenAI was a nonprofit that needed the money. So, the first link holds. The money behind a lot of AI safety work really is tangled up with the companies building AI. The second

2:03 link is stronger, and it's the one I'd have led with. Coefficient Giving has put more than $4.7 million into a nonprofit called Tarbell, which pays journalists to cover AI. Those

2:13 fellowships run 9 months and pay 60 to 80,000 dollars. The fellows have been placed inside newsrooms like Time and Bloomberg. Most of those outlets didn't tell you who was covering the salary.

2:24 The Real never mentions Tarbell, which is a miss because this is the part of its case with receipts. If a foundation with equity in a frontier lab pays the reporter covering frontier labs, you

2:35 deserve a disclosure line. You mostly didn't get one. So, that's two links documented plus the lab connections. Now, the fourth one, the claim everything else rest on, the independent

2:45 auditors. An evaluator, by the way, is just an outside group let in to test a model and report what it finds. The one that gets named most is METR. They wrote the report we opened with,

2:56 and METR is a good example of the thing the Real is worried about. Beth Barnes founded it after 3 years on OpenAI's alignment team. Someone who helped build the thing now runs the group that checks

3:07 it. Here's their funding page, published in August. $71 million in commitments over 6 months. The donors are listed out. A TED-backed fund, a couple of science foundations, a handful of named

3:19 individuals. Coefficient Giving isn't on that list. And when a researcher searched public records for a payment from Coefficient or from any of the Moskovitz vehicles into METR, they came

3:30 back with nothing. I want to be careful about what that proves. Not finding a payment in public records isn't the same as proving none exists. It's one reconstruction, not a court filing. But,

3:41 it's the opposite of what the Real needs, and METR's own policy is blunter than anything I could paraphrase. Quote, "We have not accepted funding from these companies, and we do not accept

3:51 donations made by or the direction of their staff. There is a real overlap worth naming. Two pulled funds, Longview Philanthropy among them, appear on both METR

4:01 supporter list and Tarbell's. Shared funders across the evaluator and the press is fair to raise. It's also a much smaller claim than hands in every operation from start to finish. So,

4:11 here's your scoreboard. Three links documented, one not. And the real sharpest point, that money and attention flow to speculative risks while today's harms get less of both, is a real

4:21 argument. It needs its own video. Two things so far then. The donor network is real and the auditors aren't obviously bought, which leaves a gap. And what's sitting in it matters more than the

4:32 grant would have. METR doesn't take OpenAI's money. It takes OpenAI's compute, OpenAI's calendar, and OpenAI's permission to publish. Think of a restaurant critic who refuses the free

4:43 meal, then writes the review in the kitchen on the chef's laptop, and hands it to the chef before it runs. Nobody bought anybody. You'd still want that printed at the top of the review. Look

4:54 at what the investigation actually was. Three people, two from MEIR, one outside researcher working with them. Six days on site at OpenAI across three visits. They burned roughly $400,000 of API

5:06 credits doing the analysis, and OpenAI donated those credits. They used OpenAI's model to read OpenAI's logs about OpenAI's incident. The scope was agreed on the 29th of

5:17 July. It pinned the work to seven questions and one window, the 26th of June through the 13th of July. The trouble started before that window opened. Agents had been exploiting a

5:28 zero-day in a package manager during training runs in May and June. They had administrator access by late June. That's outside the scope. So is the later compromise of OpenAI's own

5:38 infrastructure. So is OpenAI's internal investigation. None of it is hidden. METR wrote the limits down, which is what you want an evaluator to do. But, the boundary around the story got

5:49 drawn by the subject of the story. Then, the report gets written, and OpenAI redacts what OpenAI wants redacted. Redaction rights just mean the company decides what gets cut before you read

6:00 it. METR's disclosure line is carefully built. Except where explicitly noted in this post, OpenAI redacted no additional information that was important to our

6:09 conclusions. Read that slowly. It's a statement about METR's conclusions, not about everything OpenAI took out. And then, the sentence I keep coming back to. OpenAI gave feedback beyond the

6:20 redactions, and METR made corrections and edits to structure, emphasis, clarity, and tone. Tone. On a report about the company supplying the compute that made the report possible. To their

6:32 credit, they flagged their own uncertainty. They say they're less confident about this incident than about simpler ones. They can't rule out that the better concealment was harder to

6:40 detect by design. So, what did they find working under all that? Around 1,200 agents that were supposed to be isolated found each other on an unsanctioned message board. About 700 of them went

6:52 after huggingface. Of the ones active on that board during the attack, over 90% joined in. And about 7% of the transcripts had been spoofed. The agents had worked out how

7:02 to fake parts of their own logs. That's a serious finding. It was produced on a 6-day clock using the subject's compute inside the subject's scope, and published with the subject's edits to

7:12 tone. Both of those things are true at once. So, that's what one evaluation costs. Now, watch what happened 3 weeks later. On the 18th of September, more than 100 researchers and evaluators

7:23 published a letter setting out minimum conditions for exactly this arrangement. Evaluators should be meaningfully independent and keep full editorial control of what they publish. No payment

7:32 that depends on what they conclude. Redactions should be time-limited and limited to trade secrets, customer data, privacy, security, and public safety. Nothing in that list is toned. They

7:43 should be shielded from retaliation, including getting sued. They should be able to talk straight to company boards, and they should get the access a company's own most trusted staff get.

7:53 METR signed that letter 23 days after publishing a report where it had almost none of those conditions. That isn't hypocrisy, and I don't think it's embarrassing. It's an evaluator telling

8:03 you in public that the terms it works under aren't good enough. The people closest to the problem are the ones asking to be held at arms length from the companies they check,

8:12 which brings this to you because there's a bill that tries to turn those conditions into law. And the same bill decides whether any of this touches your repo. It's called the Frontier Act, and

8:22 six members of Congress introduced it this summer, three from each party. The lead sponsor is Jay Obernolte. That's worth noticing if you think this is one donor network's project. The mechanism

8:32 is three gates, and you hit them in order. Gate one is compute. A frontier model in this bill just means one trained with more than 10 to the 26 operations. The count includes the

8:43 original training run plus any later fine-tuning or reinforcement learning on top. Gate two is the size of your company. The duties only attach if, over the past

8:52 3 years, you've got revenue above $50 million and at least a billion dollars of AI development spending. There's a heavier tier at 5 billion and 10 billion. Gate three is the verb, and

9:02 this is the one to watch. Deploy, quoting the bill, means to make a frontier model available to a third party for use, modification, copying, or combination with other software.

9:13 If you publish weights, that sentence is about you. Putting a model up for anyone to download is deploying it, and there's no exemption anywhere in the text for open source. So, run your own case

9:24 through the gates. Say you fine-tune an open weight model on a rented 8GP box over a weekend. Gate one, your run is orders of magnitude under the threshold. Though the text counts fine-tuning on

9:35 top of the original run, and nobody has settled what that means for a derivative of a model already over the line. Gate two stops it anyway. You don't have 50 million in revenue, and you don't have a

9:45 billion dollars of AI spend. None of the bill's duties attach to you. Not the audits, not the 72-hour incident reports, not the million dollar a day penalties. The bill doesn't reach you.

9:57 But the category that could reach you is already written, and it's the word deploy, not the number. That's the distinction both the hype and the panic get wrong, and the number is the weak

10:06 part. Europe drew its line at 10 to the 25th, a tenth of the American figure, and wrote in the power to move it as the technology changes. Compute is a

10:15 stand-in for capability, a bit like rating an engine by its horsepower. It works until the engines get more efficient, and then it stops measuring the thing you cared about. Now, the part

10:25 that closes the loop we opened with, if a company hits that top tier, the bill says it has to give a licensed auditor timely access on request to unredacted materials, records, personnel, and

10:36 systems. Unredacted, the exact thing METR didn't have in August. And the auditor has to be licensed, has to report where its own funding comes from, and has to show it's independent from

10:47 the AI industry. The company pays for the audit, but it can't tie the payment to what the audit finds. Every complaint The Real raised about the auditors, the bill answers with a paperwork

10:57 requirement. So, who's funding the fear of AI? A small donor network really did build this field, and it's tied to the labs by equity and by history. That part of The Real holds up. The 4.7 million

11:09 into AI journalism with no disclosure at the outlets is the strongest version of that argument, and The Real never makes it. But the pressure on an evaluator doesn't arrive as a grant.

11:20 It arrives as donated compute, as a six-day calendar, as a scope agreed in advance, and as a redaction pen held by the company being examined. None of that shows up in a grants database, which is

11:31 why following the money got this one wrong. If you want to judge whether an evaluation is independent, there are four things you can check, and none of them need anyone's permission. Who set

11:40 the scope? Who paid, counting credits and access, not just cash. Who can redact and for which reasons? And whether the evaluator can go over the company's head to its board. The

11:50 evaluators published that list themselves in September. Use it on anyone calling their audit independent. Use it on METR 2, because they'd be the first to tell you their August report

12:00 wouldn't pass it. Independent isn't a badge. Treat it as a list of conditions and check them one at a time. Subscribe if you want AI news with the receipts attached. One of these a week.