Inspired by this Tweet by Rohit Krishnan https://x.com/krishnanrohit/status/1923097822385086555
One thing LLMs can teach us is that memorisation of facts is, in fact, a necessary part of reasoning and intelligent behaviour
There’s a very simple model under which memorization is important:
if good reasoning resembles test-time compute, judgement involves iterating through lots of potential hypotheses and validating them
the speed with which you validate each hypothesis is partly a function of the time it takes to fetch the data necessary to assess it
when that data is memorized (i.e. stored “in your brain’s weights”) access is really fast
when that data is stored externally (i.e. have to run a google search to find it) validation is OOMs slower
so you’ll be OOMs slower if you haven’t internalized the context of the problem
further, there are upper bounds on the slower process (because you get bored, or you fatigue, etc.) so this difference in speed may also end up being a difference in capability. it’s not just that it’s OOMs slower, it may not even get there
This doesn’t please me as I generally dislike memorization, but the logic seems plausible.
Epistemics vs Video Generation
Veo 3 released yesterday serves as another example of what’s surely coming in terms of being able to generate video that’s indistinguishable from reality. We will be coming off a many-decades period where we could just believe video as a source of truth: what a luxury that will have been, in hindsight!
Assuming it’s important to have sources of truth, I see the following options going forward:
we will have to just accept that our world has permanently become epistemically worse. sad!
we will have to trust real life more than we do external evidence. this seems very limiting: my real life friends don’t have much insight into what’s happening in Ukraine, for example
we will need to pair video with a sort of “chain of custody” whereby the video itself will not be worth anything (epistemically) unless it comes with someone, whose identity is established in some other way, vouching for it being true. this is, in a way, how we presently treat written evidence.
lacking a trusted source, but making up for it by combining multiple videos of the same thing. imagine multiple angles of the same event (as confirmed by hard-to-spoof video metadata) corroborating each other (perhaps this can even be AI-enabled, in a privacy-preserving way). hoping that this is easy to verify but hard to fake (hard to generate mutually consistent multiple angles of the same thing) though this might be wishful thinking since this seems like something that can be trained eventually, but maybe buys us a few years
the physics solution: from first principles, video had the property that it was much easier to verify than it was to fake (a la certain encryption). video is, reductively, just a time series of photons. perhaps we can build The Truth Camera, which captures lots of other particles (or maybe captures photons at a frame rate and resolution that is costly, and thus there isn’t much data of, and thus is hard to credibly fake). essentially perhaps we can snapshot more aspects of reality in such a way that we re-create the condition that made video credible: easy to verify, hard to generate.
nb: There will likely be a long period where video is fake but still believed, much like people still believe things that are written down over things that are spoken, despite the fact that anyone can freely write down anything at any point.