Frontier Labs Seek Transparent Mandatory Framework for AI Model Approvals
Posts from X
2 posts
Post-Mythos, we no longer live in a world without AI regulation. The menu of options is now: (a) continue with the current "voluntary"-but-not-really regime under which some models are approved by the USG, but only if they're submitted for approval "voluntarily", and the USG's process is based on classified standards and opaque OR (b) agree on a transparent written framework under which certain specific models will be *required* to be approved As the recent excellent essay by @demishassabis and today's policy write-up from Chris Lehane show, the frontier labs strongly prefer option (b). This is not surprising! Clear written rules of the road will always be preferable to a regime where the cop can stop you for any reason or no reason. (Not having cops policing traffic is no longer an option at all - see above.) What do (or should) the labs want? I think it's something along the following lines: - If possible, legislation enacted by Congress, not a set of rules prescribed by an Executive Order - which can easily be amended or supplemented at a later date. - A *mandatory*, not voluntary approval process for "covered" models. There should be written criteria for what qualifies as a "covered" model subject to approval. Models not meeting these criteria should be permitted to be publicly released without undergoing any prior approval process. - A safe harbor from approval requirements for models that are not any more dangerous than open-source models that are already publicly available at the time of proposed model release. - Preemption of any state or local laws attempting to impose their own approval requirements on release of AI models. This should be a matter left to the federal government. - A set of specific grounds under which the USG could disapprove the model. Examples: cybersecurity risk, bio risk. No "catch-alls". We would not want a regime under which the USG could disapprove a model, e.g., because it is too "woke" or not "woke" enough, or because the USG has issues with one particular lab, or because of any other factors that don't have anything to do with catastrophic risk. - A requirement that model testing be performed, and approval be issued solely based on recommendations of, a team of technical safety experts. Maybe it's a "FINRA-like" body envisioned by Demis Hassabis; maybe it's CAISI. The decision should not be left in the hands of career politicians who don't understand safety from a technical perspective. - Transparency around the criteria and benchmarks the USG will use to test the models, and a requirement that a publicly available approval order be issued by the USG each time a new model has passed or failed, with an explanation of the benchmarking that was done, how the model passed/failed the benchmarks, and any conditions imposed on the approval. - A relatively short mandatory timeline (30 days?) for the USG completing its work and issuing the approval. - A clear path for expedited judicial review of USG disapproval decisions by the aggrieved frontier lab. But now, let's flip the table and consider all this from the perspective of the USG. Remember: these people are politicians, not AI safety experts. They are non-technical. They feel the weight of responsibility for steering the U.S. through the era of mass adoption of this technology. They are terrified of the public harm that AI could - in theory - cause, and are cognizant of the effect this would have on the polls. They are generally pro-technology. They do not want to cede ground to China. They sort of understand the technology by now, but they know that there are "unknown unknowns". They are on the phone with frontier lab CEOs probably on at least a weekly basis, and are told things like: "Well, our AI models pose huge cyber risks now, and bio/chemical are probably next, and there could be a huge impact on the labor market as well, and terrorists could probably use this technology to more easily make bombs or even fully autonomous weapons, but we don't even really know ourselves when this could happen - capabilities overhang, you see!" Given this, if you are the USG, what you want first and foremost for dealing with AI is *flexibility*. This means: - No rules of the road clearly written out, and definitely no legislation. - No transparency behind your decisions at all, because it is much easier to get things done and get catastrophic risks quashed if you can move quickly and outside the public eye. - No requirement to involve technical experts like CAISI in anything you do (although of course you will involve them to the extent you deem it necessary). You are making *political* decisions first and foremost about the impact of this technology, not technical decisions. - Maximum flexibility regarding your approval criteria. Today it's cyber, tomorrow it's bio, the day after tomorrow, it's whatever the day after tomorrow brings. - No judicial review (of the kind mandated by legislation in these specific circumstances). I suspect that this will continue to be a point of tension between the frontier labs and the USG for quite some time.
The admin says AI model review is voluntary. But it pressured Meta to join, told CAISI to stop publishing, and export-controlled Anthropic's models. It's voluntary on paper but not in practice. My latest is on Mythos and the new ad hoc licensing regime: https://exformation.williamrinehart.com/p/the-mythos-moment-and-the-reordering