(Also posted on LessWrong.)
Treaties to pause training or otherwise restrict dangerous uses of AI would require difficult verification schemes. Plan A's proposal is that network taps inside data centers would record copies of traffic (or hashes of it), and later recomputations on random samples could check whether sampled outputs match declared workloads, as part of a broader verification scheme.
I'm an AI engineer at a company building tools for the construction industry. Much of our work is with data centers. The architects, engineers and construction managers on our team work every week with data center developers, their design engineers, and the general contractors building their sites.
As of now, the company does nothing at all involving AI verification, and I'm writing this in my personal capacity to try to learn: Can we be useful (e.g. helping with research) in these efforts? I do not speak for the company, but my hope is that if I learn how we can be useful, I could use my influence to make a case for this becoming a company-endorsed effort (albeit a side project).
We have a few features:
This is all built via AI engineers like me working closely with domain experts like architects, electrical engineers, construction management experts, etc.
Some of the writing on this that I've looked at over the last few days are:
(But: There's a lot! A lot of work has gone into this and I just started thinking about it so despite reading those, I still haven't internalized much of it.)
The tap-based approaches require physical changes to what goes into a new data center as well as retrofitting existing ones:
Plan A's retrofit scenario assumes most of the equipment is powered down during the work and includes a one-month retrofit target:
> retrofitting 40% of the US compute into inference units of 4 server racks (or 4K H100e) with their existing stock of network taps in 1 month
I didn't find a labor or installation-sequencing estimate behind that one-month timeline.
Amodo's SITREP explicitly calls for an installation playbook and says there has been little work on how to inspect verification systems at scale. Those sound like areas where construction expertise might help.
Any work we do on this would draw on our people's general professional expertise, on public reference designs, and on the developers and contractors we work with. I think it's possible that (with effort) our team (and maybe product) could help with:
I don't know what form our help would take, e.g.:
Our company will not shift its focus to these problems. For now, I’m looking for people working on verification who see a specific use for this expertise. If there’s a fit, I could explore whether we can help. I’m not sure whether the teams pushing verification forward already have access to the kinds of experts I work with.
We don't directly build or operate data centers ourselves, but maybe this isn't too far from Plan A's Get Involved page requesting help from "datacenter builders and operators who know what is operationally feasible."
If you're working on these problems and think a small portion of our time could be useful to you, please comment or reach out to me. I'd also appreciate introductions to people working on installation and inspection.
