Remember Dario Amodei's "Pace the Frontier" essay from two weeks ago, arguing labs should deliberately slow capability improvements so safety research can catch up? Altman endorsed it. Musk endorsed it. David Sacks told them all, essentially: go ahead and slow down, but do it yourselves. SAFA is what that instruction looks like when three competitors take it literally.
The first thing to go was government involvement. The labs wanted a public-private partnership with official backing. That didn't happen. What remains is a purely industry-constructed authority with no external enforcement mechanism anyone has described. The companies being evaluated also write the criteria.
Meanwhile, two former safety researchers at Anthropic and Google DeepMind recently left to join METR for independent AI risk assessments. One of them called for mandatory reporting of recursive self-improvement progress. A lab-controlled standards body makes the case for genuinely independent evaluation sharper.
Here's where it gets interesting for everyone outside the Big Three. If large enterprise buyers or government procurement offices start requiring "SAFA-certified" as a checkbox, that certification functions as a barrier to entry for smaller labs and open-weight model providers — the same providers who gained traction precisely because they operate outside the frontier labs' orbit. That's a convenient side effect if you're one of the three companies writing the rules.
When three competitors jointly define safety standards while also preparing IPOs and explaining sandbox escapes, the incentive picture is pretty clear.