LessWrong dev & admin as of July 5th, 2022.
RobertM
Karma: 2,652
Against “argument from overhang risk”
Yeah, I agree that it’s too early to call it re: hitting a wall. I also just realized that releasing 4o for free might be some evidence in favor of 4.5/5 dropping soon-ish.
Vaguely feeling like OpenAI might be moving away from GPT-N+1 release model, for some combination of “political/frog-boiling” reasons and “scaling actually hitting a wall” reasons. Seems relevant to note, since in the worlds where they hadn’t been drip-feeding people incremental releases of slight improvements over the original GPT-4 capabilities, and instead just dropped GPT-5 (and it was as much of an improvement over 4 as 4 was over 3, or close), that might have prompted people to do an explicit orientation step. As it is, I expect less of that kind of orientation to happen. (Though maybe I’m speaking too soon and they will drop GPT-5 on us at some point, and it’ll still manage to be a step-function improvement over whatever the latest GPT-4* model is at that point.)
It’s not obvious to me why training LLMs on synthetic data produced by other LLMs wouldn’t work (up to a point). Under the model where LLMs are gradient-descending their way into learning algorithms that predict tokens that are generated by various expressions of causal structure in the universe, tokens produced by other LLMs don’t seem redundant with respect to the data used to train those LLMs. LLMs seem pretty different from most other things in the universe, including the data used to train them! It would surprise me if the algorithms that LLMs developed to predict non-LLM tokens were perfectly suited for predicting other LLM tokens “for free”.
EDIT: looks like habryka got there earlier and I didn’t see it.
https://www.lesswrong.com/posts/zXJfH7oZ62Xojnrqs/#sLay9Tv65zeXaQzR4
Intercom is indeed hidden on mobile (since it’d be pretty intrusive at that screen size).
Ah, does look like Zach beat me to the punch :)
I’m also still moderately confused, though I’m not that confused about labs not speaking up—if you’re playing politics, then not throwing the PM under the bus seems like a reasonable thing to do. Maybe there’s a way to thread the needle of truthfully rebutting the accusations without calling the PM out, but idk. Seems like it’d be difficult if you weren’t either writing your own press release or working with a very friendly journalist.
I hadn’t, but I just did and nothing in the article seems to be responsive to what I wrote.
Amusingly, not a single news source I found reporting on the subject has managed to link to the “plan” that the involved parties (countries, companies, etc) agreed to.
Nothing in that summary affirmatively indicates that companies agreed to submit their future models to pre-deployment testing by the UK AISI. One might even say that it seems carefully worded to avoid explicitly pinning the companies down like that.
EDIT: I believe I’ve found the “plan” that Politico (and other news sources) managed to fail to link to, maybe because it doesn’t seem to contain any affirmative commitments by the named companies to submit future models to pre-deployment testing by UK AISI.
I’ve seen a lot of takes (on Twitter) recently suggesting that OpenAI and Anthropic (and maybe some other companies) violated commitments they made to the UK’s AISI about granting them access for e.g. predeployment testing of frontier models. Is there any concrete evidence about what commitment was made, if any? The only thing I’ve seen so far is a pretty ambiguous statement by Rishi Sunak, who might have had some incentive to claim more success than was warranted at the time. If people are going to breathe down the necks of AGI labs about keeping to their commitments, they should be careful to only do it for commitments they’ve actually made, lest they weaken the relevant incentives. (This is not meant to endorse AGI labs behaving in ways which cause strategic ambiguity about what commitments they’ve made; that is also bad.)
Huh, that went somewhere other than where I was expecting. I thought you were going to say that ignoring letter-of-the-rule violations is fine when they’re not spirit-of-the-rule violations, as a way of communicating the actual boundaries.
Yeah, there needs to be something like a nonlinearity somewhere. (Or just preference inconsistency, which humans are known for, to say nothing of larger organizations.)
I’m not sure I personally endorse the model I’m proposing, but imagine a slightly less spherical AGI lab which has more than one incentive (profit maximization) driving its behavior. Maybe they care at least a little bit about not advancing the capabilities frontier as fast as possible. This can cause a preference ordering like:
don’t argmax capabilities, because there’s no open-source competition making it impossible to profit from current-gen models
argmax capabilities, since you need to stay ahead of open-source models nipping at your heels
don’t argmax capabilities; go bankrupt because open-source catches up to you (or gets “close enough” for enough of your customers)
ETA: But in practice most of my concerns around open-source AI development are elsewhere.
LW Frontpage Experiments! (aka “Take the wheel, Shoggoth!”)
I think there might be many local improvements, but I’m pretty uncertain about important factors like elasticity of “demand” (for robbery) with respect to how much of a medication is available on demand. i.e. how many fewer robberies do you get if you can get at most a single prescriptions’ worth of some kind of controlled substance (and not necessarily any specific one), compared to “none” (the current situation) or “whatever the pharmacy has in stock” (not actually sure if this was the previous situation—maybe they had time delay safes for storing medication that wasn’t filling a prescription, and just didn’t store the filled prescriptions in the safes as well)?
Headline claim: time delay safes are probably much too expensive in human time costs to justify their benefits.
The largest pharmacy chains in the US, accounting for more than 50% of the prescription drug market[1][2], have been rolling out time delay safes (to prevent theft)[3]. Although I haven’t confirmed that this is true across all chains and individual pharmacy locations, I believe these safes are used for all controlled substances. These safes open ~5-10 minutes after being prompted.
There were >41 million prescriptions dispensed for adderall in the US in 2021[4]. (Note that likely means ~12x fewer people were prescribed adderall that year.) Multiply that by 5 minutes and you get >200 million minutes, or >390 person-years, wasted. Now, surely some of that time is partially recaptured by e.g. people doing their shopping while waiting, or by various other substitution effects. But that’s also just adderall!
Seems quite unlikely that this is on the efficient frontier of crime-prevention mechanisms, but alas, the stores aren’t the ones (mostly) paying the costs imposed by their choices, here.
use spaces that your community already has (Lighthaven?), even if they’re not quite set up the right way for them
Not set up the right way would be an understatement, I think. Lighthaven doesn’t have an indoor space which can seat several hundred people, and trying to do it outdoors seems like it’d require solving maybe-intractable logistical problems (weather, acoustics, etc). (Also Lighthaven was booked, and it’s not obvious to me to what degree we’d want to subsidize the solstice celebration. It’d also require committing a year ahead of time, since most other suitable venues are booked up for the holidays quite far in advance.)
I don’t think there are other community venues that could host the solstice celebration for free, but there might be opportunities for cheaper (or free) venues outside the community (with various trade-offs).
Having said that, I would NOT describe this as asking “how could I have arrived at the same destination by a shorter route”. I would just describe it as asking “what did I learn here, really”.
I mean, yeah, they’re different things. If you can figure out how to get to the correct destination faster next time you’re trying to figure something out, that seems obviously useful.
Some related thoughts. I think the main issue here is actually making the claim of permanent shutdown & deletion credible. I can think of some ways to get around a few obvious issues, but others (including moral issues) remain, and in any case the current AGI labs don’t seem like the kinds of organizations which can make that kind of commitment in a way that’s both sufficiently credible and legible that the remaining probability mass on “this is actually just a test” wouldn’t tip the scales.
I am not covering training setups where we purposefully train an AI to be agentic and autonomous. I just think it’s not plausible that we just keep scaling up networks, run pretraining + light RLHF, and then produce a schemer.[2]
Like Ryan, I’m interested in how much of this claim is conditional on “just keep scaling up networks” being insufficient to produce relevantly-superhuman systems (i.e. systems capable of doing scientific R&D better and faster than humans, without humans in the intellectual part of the loop). If it’s “most of it”, then my guess is that accounts for a good chunk of the disagreement.
Curated. I liked that this post had a lot of object-level detail about a process that is usually opaque to outsiders, and that the “Lessons Learned” section was also grounded enough that someone reading this post might actually be able to skip “learning from experience”, at least for a few possible issues that might come up if one tried to do this sort of thing.
Yeah, “they’re following their stated release strategy for the reasons they said motivated that strategy” also seems likely to share some responsibility. (I might not think those reasons justify that release strategy, but that’s a different argument.)