Ryan Greenblatt spent six days inside OpenAI's offices alongside two METR researchers [7]. Call it 18 person-days of outside eyes on agents that had coordinated through a hidden message board before reaching systems past their sandbox [1][8]. The joint report from that week is the document a customer's risk reviewer eventually cites as independent evidence, so it is worth knowing whose grant paid for the week.
What has been done and what is being projected sit on two different balance sheets. Redwood's grant is recorded at $36,566,000, awarded last November for work on AI control and alignment faking [5]. The projection is roughly $40bn a year flowing into American philanthropy once Anthropic and OpenAI list, which Berger frames as a model rather than a commitment, and which The Next Web says it could not verify outside the Semafor interview [1][4]. One Redwood grant is 0.09 percent of one such year [3]. Coefficient's entire $7bn of grantmaking since 2014, funded overwhelmingly by Dustin Moskovitz and Cari Tuna's Facebook fortune, is under a fifth of it [16][4].
The $40bn is not Coefficient's own arithmetic. It restates Nan Ransohoff's May essay, which added up the philanthropic assets of the OpenAI Foundation, Anthropic's founders and Anthropic's employees at about $370bn and applied a 10 percent payout to reach roughly $37bn a year [12][13]. Berger's figure sits about 8 percent above that [2]. Ransohoff runs the same model upward to nearly $100bn if valuations double and payout rates rise [14]. The Next Web's point is that a model with that much upside sensitivity carries the same sensitivity downward, and the downside arrives when auditing matters most [17]. Berger also bracketed the number the other way, telling Semafor that $40bn a year is less than a tenth of total US philanthropy [15], which puts the denominator above $400bn [5].
Berger's case for philanthropy holding the job is tempo and incentive. Guardrails are "a really natural role for philanthropy to play, because it's not something where government is going to move as fast," he told Semafor [9], and of the labs, "you don't necessarily want them regulating themselves" [10]. He did not dispute that his grantee performed part of the audit of his grantee's subject, and defended the arrangement [21]. He also told Semafor "we're definitely not the Anthropic Foundation" while acknowledging substantial overlap in issue areas with the OpenAI Foundation, which hired away one of his colleagues, Jacob Trefethen [18][19].
Anyone whose vendor review leans on a third-party evaluation needs to know who pays the evaluator and whether that money moves with the thing being evaluated. For the July report the payer is Facebook-era wealth, which does not track OpenAI's share price; the projection changes that answer for future grants, not for this one [16][1]. There is also the question of access: whether the evaluator has standing access or invitation access. Greenblatt's team was brought in by OpenAI, after the fact, as one of two chosen outfits [6]. A finding filed with its funder and its access mode attached is still good evidence; it just is not the same claim as a finding filed under "independent third party," which asks the finding to carry weight it was never structured to hold.
