Pinned post
Great news! Governor Newsom has signed SB 813 into law, a Fathom-sponsored bill that establishes the first comprehensive framework for independent verification organizations (IVOs). 🧵
Fathom
Verified account This account is verified. Learn more@Fathom_org
Fathom is an independent nonprofit that finds, builds, and scales policy and technical innovations designed for the AI century.
279 Following 1.1K Followers 994 Posts
Latest posts 5
“A big-name push for AI audits”: @BrendanBordelon has the first look at our new Executive Council in @politico Decoded this morning.
“Fathom was the driving force behind California’s passage this year of SB 813, which requires the state to set standards for outside groups that evaluate AI safety risks. It now wants to spread that to other states, countries and Washington DC.”
“Fathom was the driving force behind California’s passage this year of SB 813, which requires the state to set standards for outside groups that evaluate AI safety risks. It now wants to spread that to other states, countries and Washington DC.”
We’re grateful to welcome these leaders to our Executive Council. Their experience will help us build AI governance policies that work in practice.
Big day for Fathom. We are announcing our Executive Council, which will help guide our strategic decisions. Some of them have been with us since the early days. I am honestly humbled that this group has chosen to give us their time, reputations, and efforts.
Bringing about meaningful AI oversight that works at the speed of this technology is a generational challenge that will take a big-tent approach, across sectors, partisan divides, and even international boundaries.
This council is made up of former governors of both parties, national security leaders and AI experts who know how complex policy needs to work in practice. Excited to have them on board.
fathom.org/about#executiv…View quoted post on X
Our CEO on what independent AI oversight needs to get right.
Labs, lawmakers in both parties, and even the President have said we need independent testing of AI. But as that testing gets underway, what does “good” actually look like for embedded/independent/third party evaluations?
As the architects of the IVO framework, we think a lot about this. Here's my take, the first in a series from @Fathom_org where we’ll go deeper. Give us your input:
fathomai.substack.com/p/independent-…View quoted post on X
AI is increasingly helping build the next generation of AI. For people to trust it, those in charge need to understand what’s being built.
AI now writes 75%+ of new code at Anthropic, OpenAI, and Google. Meanwhile, in open-source AI projects, the number of review comments per line of code has dropped by roughly half since early 2024, which means human review might not be keeping pace.
Our Senior Advisor @ghadfield, @ronbodkin, and @bahradx have an idea called comprehension audits. Independent auditors select important pieces of AI R&D work, and the teams responsible have to explain what they built, how they built it, and what they learned, kind of like defending a thesis.
The bar is practical: teams must demonstrate meaningful understanding of their own work, including its design, implementation, and results, without being expected to explain every aspect of model behavior.
This isn’t to punish individual researchers. It's to identify gaps in understanding. If a team can't demonstrate adequate understanding, the affected work and any dependent releases are held until the gaps are addressed.
Through a mix of targeted and random spot checks, independent auditors can help catch problems without having to review everything.
Having humans in the loop isn't enough if they don't understand what they're approving. Independent verification can help make sure that human oversight actually means something.
AI now writes 75%+ of new code at Anthropic, OpenAI, and Google. Meanwhile, in open-source AI projects, the number of review comments per line of code has dropped by roughly half since early 2024, which means human review might not be keeping pace.
Our Senior Advisor @ghadfield, @ronbodkin, and @bahradx have an idea called comprehension audits. Independent auditors select important pieces of AI R&D work, and the teams responsible have to explain what they built, how they built it, and what they learned, kind of like defending a thesis.
The bar is practical: teams must demonstrate meaningful understanding of their own work, including its design, implementation, and results, without being expected to explain every aspect of model behavior.
This isn’t to punish individual researchers. It's to identify gaps in understanding. If a team can't demonstrate adequate understanding, the affected work and any dependent releases are held until the gaps are addressed.
Through a mix of targeted and random spot checks, independent auditors can help catch problems without having to review everything.
Having humans in the loop isn't enough if they don't understand what they're approving. Independent verification can help make sure that human oversight actually means something.
Claude leads 26% of Anthropic's R&D and at least collaborates on 90%+. Dario Amodei says recursive self-improvement "is starting to happen across the industry."View quoted post on X
New preprint with @ghadfield and @bahradx on how to respond: arxiv.org/abs/2610.10064 (2/11)
