Senior Research Scientist at NTT Research, Physics & Informatics Lab. jessriedel.com , jessriedel[at]gmail[dot]com
Jess_Riedel
I’m also a little surprised you think that modeling when we will have systems using similar compute as the human brain is very helpful for modeling when economic growth rates will change. (Like, for sure someone should be doing it, but I’m surprised you’re concentrating on it much.) As you note, the history of automation is one of smooth adoption. And, as I think Eliezer said (roughly), there don’t seem to be many cases where new tech was predicted based on when some low-level metric would exceed the analogous metric in a biological system. The key threshold for recursive feedback loops (*especially* compute-driven ones) is how well they perform on the relevant tasks, not all tasks. And the way in which machines perform tasks usually looks very different than how biological systems do it (bird vs. airplanes, etc.).
If you think that compute is the key bottleneck/driver, then I would expect you to be strongly interested in what the automation of the semiconductor industry would look like.
I like this post a lot but I will disobey Rapoport’s rules and dive straight into criticism.
Historically, many AI researchers believed that creating general AI would be more about coming up with the right theories of intelligence, but over and over again, researchers eventually found that impressive results only came after the price of computing fell far enough that simple, “blind” techniques began working (Sutton 2019).
I think this is a poor way to describe a reasonable underlying point. Heavier-than-air flying machines were pursued for centuries, but airplanes appeared almost instantly (on a historic scale) after the development of engines with sufficient power density. Nonetheless, it would be confusing to say “flying is more about engine power than the right theories of flight”. Both are required. Indeed, although the Wright brothers were enabled by the arrival of powerful engines, they beat out other would-be inventors (Ader, Maxim, and Langley) who emphasized engine power over flight theory. So a better version of your claim has to be something like “compute quantity drives algorithmic ability; if we independently vary compute (e.g., imagine an exogenous shock) then algorithms follow along”, which (I think) is what you arguing further in the post.
But this also doesn’t seem right. As you observe, algorithmic progress has been comparable to compute progress (both within and outside of AI). You list three “main explanations” for where algorithmic progress ultimately comes from and observe that only two of them explain the similar rates of progress in algorithms and compute. But both of these draw a causal path from compute to algorithms without considering the (to-me-very-natural) explanation that some third thing is driving them both at a similar rate. There are a lot of options for this third thing! Researcher-to-researcher communication timescales, the growth rate of the economy, the individual learning rate of humans, new tech adoption speed, etc. It’s plausible to me that compute and algorithms are currently improving more or less as fast as they can, given their human intermediaries through one or all of these mechanisms.
The causal structure is key here, because the whole idea is to try and figure out when economic growth rates change, and the distinction I’m trying to draw becomes important exactly around the time that you are interested in: when the AI itself is substantially contributing to its own improvement. Because then those contributions could be flowing through at least three broad intermediaries: algorithms (the AI is writing its own code better), compute (the AI improves silicon lithography), or the wider economy (the AI creates useful products that generate money which can be poured into more compute and human researchers).
Of course, even if AI performance is, in principle, predictable as a function of scale, we lack data on how AIs are currently improving on the vast majority of tasks in the economy, hindering our ability to predict when AI will be widely deployed. While we hope this data will eventually become available, for now, if we want to predict important AI capabilities, we are forced to think about this problem from a more theoretical point of view.
Humans have been automating mechanical task for many centuries, and information-processing tasks for many decades. Moore’s law, the growth rate of the thing (compute) that you ague drives everything else, has been stated explicitly for almost 58 years (and presumably applicable for at least a few decade before that). Why are you drawing a distinction between all the information processing that happened in the past and “AI”, which you seem to be taking as a basket of things that have mostly not had a chance to be applied yet (so no data)?
If compute is the central driving force behind AI, and transformative AI (TAI) comes out of something looking like our current paradigm of deep learning, there appear to be a small set of natural parameters that can be used to estimate the arrival of TAI. These parameters are:
The total training compute required to train TAI
The average rate of growth in spending on the largest training runs, which plausibly hits a maximum value at some significant fraction of GWP
The average rate of increase in price-performance for computing hardware
The average rate of growth in algorithmic progress
This list is missing the crucial parameters that would translate the others into what we agree is most notable: economic growth. I think needs to be discussed much more in section 4 for it to be a useful summary/invitation to the models you mention.
I agree it’s important to keep the weaker fraud protection on debit cards in mind. However, for the use I mentioned above, you can just lock the debit card and only unlock it when you have a cash flow problem. (Btw, if you don’t use your IB debit card, you should lock it even if you aren’t using it.) Debit card liability is capped at $50 and $500 if you report fraudulent transactions within 2 days and 60 days, respectively.
That said, I have most of my net worth elsewhere, so I’m less worried about tail risks than you would reasonably be if you’re mostly invested through IB.
If you have non-qualified investments and just keep money in a savings account in case of unexpected large expenses or interruptions to your income, it may be better to instead move the money in the savings account to Interactive Brokers and invest it. Crucially, you can get a debit card from Interactive Brokers that allows you to spend on margin (borrow) at a low rate (~5%, much less than credit cards) using your investments there as collateral. That way you keep essentially all your money invested (presumably earning more than the savings account) while still having access to liquidity when you need it.
Just to be clear: we mostly don’t argue for the desirability or likelihood of lock-in, just its technological feasibility. Am I correctly interpreting your comment to be cautionary, questioning the desirability of lock-in given the apparent difficulty of doing so while maintaining sufficiently flexibility to handle unforeseen philosophical arguments?
AGI and Lock-In
If the Federal government is just buying, on the open market, an amount of coal comparable to how much would have been sold without government action, then it’s going to drive up the price of coal and increase the total amount of coal extracted. How much extra coal gets extracted depends on the supply and demand curves, and the amount of coal actually burned will almost certainly be less than in the world where the government didn’t act, but it does mean the environmental benefits of this plan will be significantly muted.
Paul Graham writes that Noora Health is doing something like this.
https://twitter.com/Jess_Riedel/status/1389599895502278659
https://opensea.io/assets/0x495f947276749ce646f68ac8c248420045cb7b5e/96773753706640817147890456629920587151705670001482122310561805592519359070209
Regarding your 4 criteria, I think they don’t really delineate how to make the sort of judgment calls we’re discussing here, so it really seems like it should be about a 5th criterion that does delineate that.
Sorry I was unclear. Those were just 4 desiderata that the criteria need to satisfy; the desiderata weren’t intended to fully specify the criteria.
If a small group of researchers at MIRI were trying to do work on verification but not getting much traction in the academic community, my intuition is that their papers would reliably meet your criteria.
Certainly possible, but I think this would partly be because MIRI would explicitly talk in their paper about the (putative) connection to TAI safety, which makes it a lot easier for me see. (Alternative interpretation: it would be tricking me, a non-expert, into thinking there was more of a substantive connection to TAI safety than actually is there.) I am trying not to penalize researchers for failing to talk explicitly about TAI, but I am limited.
I think it’s more likely the database has inconsistencies of the kind you’re pointing at from CHAI, Open AI, and (as you’ve mentioned) DeepMind, since these organizations have self-described (partial) safety focus while still doing lots of research non-safety and near-term-safety research. When confronted with such inconsistencies, I will lean heavily toward not including any of them since this seems like the only feasible choice given my resources. In other words, I select your final option: “The hypothetical MIRI work shouldn’t have made the cut”.I definitely agree that you shouldn’t just include every paper on robustness or verification, but perhaps at least early work that led to an important/productive/TAI-relevant line should be included
Here I understand you to be suggesting that we use a notability criterion that can make up for the connection to TAI safety being less direct. I am very open to this suggestion, and indeed I think an ideal database would use criteria like this. (It would make the database more useful to both researchers and donors.) My chief concern is just that I have no way to do this right now because I am not in a position to judge the notability. Even after looking at the abstracts of the work by Raghunathan et al. and Wong & Kolter, I, as a layman, am unable to tell that they are quite notable.
Now, I could certainly infer notability by (1) talking to people like you and/or (2) looking at a citation trail. (Note that a citation count is insufficient because I’d need to know it’s well cited by TAI safety papers specifically.) But this is just not at all feasible for me to do for a bunch of papers, much less every paper that initially looked equally promising to my untrained eyes. This database is a personal side project, not my day job. So I really need some expert collaborators or, at the least, some experts who are willing to judge batches of papers based on a some fixed set of criteria.
Sure, sure, we tried doing both of these. But they were just taking way too long in terms of new papers surfaced per hour worked. (Hence me asking for things that are more efficient than looking at reference lists from review articles and emailing the orgs.) Following the correct (promising) citation trail also relies more heavily on technical expertise, which neither Angelica nor I have.
I would love to have some collaborators with expertise in the field to assist on the next version. As mentioned, I think it would make a good side project for a grad student, so feel to nudge yours to contact us!
for instance if you think Wong and Cohen should be dropped then about half of the DeepMind papers should be too since they’re on almost identical topics and some are even follow-ups to the Wong paper).
Yea, I’m saying I would drop most of those too.
I think focusing on motivation rather than results can also lead to problems, and perhaps contributes to organization bias (by relying on branding to asses motivation).
I agree this can contribute to organizational bias.
I do agree that counterfactual impact is a good metric, i.e. you should be less excited about a paper that was likely to soon happen anyways; maybe that’s what you’re saying? But that doesn’t have much to do with motivation.
Just to be clear: I’m using “motivation” here in the technical sense of “What distinguishes this topic for further examination out of the space of all possible topics?”, i.e., is the topic unusually likely to lead to TAI safety results down the line?” (It’s not anything to do with the author’s altruism or whatever.)
I think what would best advance this conversation would be for you to propose alternative practical inclusion criteria which could be contrasted the ones we’ve given.
Here’s how is how I arrived at ours. The initial desiderata are:
-
Criteria are not based on the importance/quality of the paper. (Too hard for us to assess.)
-
Papers that are explicitly about TAI safety are included.
-
Papers are not automatically included merely for being relevant to TAI safety. (There are way too many.)
-
Criteria don’t exclude papers merely for failure to mention TAI safety explicitly. (We want to find and support researchers working in institutions where that would be considered too weird.)
(The only desiderata that we could potentially drop are #2 or #4. #1 and #3 are absolutely crucial for keeping the workload manageable.)
So besides papers explicitly about TAI safety, what else can we include given the fact that we can’t include everything relevant to safety? Papers that TAI safety researchers are unusually likely (relative to other researchers) to want to read, and papers that TAI safety donors will want to fund. To me, that means the papers that are building toward TAI safety results more than most papers are. That’s what I’m trying to get across by “motivated”.
Perhaps that is still too vague. I’m very in your alternative suggestions!
-
Thanks Jacob. That last link is broken for me, but I think you mean this?
You sort of acknowledge this already, but one bias in this list is that it’s very tilted towards large organizations like DeepMind, CHAI, etc.
Well, it’s biased toward safety organizations, not large organizations. (Indeed, it seems to be biased toward small safety organizations over larges ones since they tend to reply to our emails!) We get good coverage of small orgs like Ought, but you’re right we don’t have a way to easily track individual unaffiliated safety researchers and it’s not fair.
I look forward to a glorious future where this database is so well known that all safety authors naturally send us a link to their work when its released, but for now the best way we have of finding papers is (1) asking safety organizations for what they’ve produced and (2) taking references from review articles. If you can suggest another option for getting more comprehensive coverage per hour of work we’d be very interested to hear it (seriously!).
For what it’s worth, the papers by Hendrycks are very borderline based on our inclusion criteria, and in fact I think if I were classifying it today I think I would not include it. (Not because it’s not high quality work, but just because I think it still happens in a world where no research is motivated by the safety of transformative AI; maybe that’s wrong?) For now I’ve added the papers you mention by Hendrycks, Wong, and Cohen to the database, but my guess is they get dropped for being too near-term-motivated when they get reviewed next year.
More generally, let me mention that we do want to recognize great work, but our higher priority is to (1) recognize work that is particularly relevant to TAI safety and (2) help donors assess safety organizations.
Thanks again! I’m adding your 2019 review to the list.
TAI Safety Bibliographic Database
Jaime gave a great thorough explanation. My catch-phrase version: This is not a holistic Bayesian prediction. The confidence intervals come from bootstrapping (re-sampling) a fixed dataset, not summing over all possible future trajectories for reality.
I was curious about the origins of this concept in the EA community since I think it’s correct, insightful, and I personally had first noticed it in conversation among people at Open Phil. On Twitter, @alter_ego_42 pointed out the existence of the Credal Resilience page in the “EA concepts” section of this website. That page cites
Skyrms, Brian. 1977. Resiliency, propensities, and causal necessity. The journal of philosophy 74(11): 704-713. [PDF]
which is the earliest thorough academic reference to this idea that I know of. With apologies to Greg, this seems like the appropriate place to post a couple comments on that paper so others don’t have to trudge through it.
I didn’t find Skyrms’s critique of frequentism at the beginning, or his pseudo-formalization of resilency on page 705 (see for instance the criticism “Some Remarks on the Concept of Resiliency” by Patrick Suppes in the very next article, pages 713-714), to be very insightful, so I recommend the time-pressed reader concentrate on
The bottom of p. 705 (“The concept of probabilistic resiliency is nicely illustrated...”) to the top of p. 708 (”… well confirmed to its degree of instantial resiliency, as specified above..”).
The middle of p. 712 (“The concept of resiliency has connections with...”) to p. 713 (the end).
Skyrms quotes Savage (1954) as musing about the possibility of introducing “second-order probabilities”. This is grounded in a relative-frequency intuition: when I say that there is a (first-order) probability p of X occurring but that I am uncertain, what I really mean is something like that there is some objective physical process that generates X with (second-order) probability q, but I am uncertain about the details of that process (i.e., about what q is), so my value of p is obtained by integrating over some pdf f (q).
There is, naturally, a Bayesian version of the same idea: We shouldn’t concern ourselves with a hypothetical giant (second-order) ensemble of models, each of which generates a hypothetical (first-order) ensemble of individual trials. Resilience about probabilities is best measured by our bets on how future evidence would change those probabilities, just as probabilities is best measured by our bets on future outcomes.
(Unfortunately, and unlike the case for standard credences, there seems to be multiple possible formulations depending on which sorts of evidence we are supposing: what I expect to learn in the actual future, what I could learn if I thought about it hard, what a superforecaster would say in my shoes, etc.)
Were there a lot of new unknown or underappreciated facts in this book? From the summary, it sounds mostly like a reinterpretation of the standard history, which hinges on questions of historical determinism.
Consider changing the visual format a bit to better distinguish this forum from LW. They are almost indistinguishable right now, especially once you scroll down just a bit and the logo disappears.
Could you explain your first sentence? What risks are you talking about?
Also, how does one lottery up further if all the block sizes are $100k? Diving it up into multiple blocks doesn’t really work.
I listed this example in my comment, it was incorrect by an order of magnitude, and it was a retrodiction. “I didn’t look up the data on Google beforehand” does not make it a prediction.