= 27d567bc2d48d9f40e9ccc20ea370b15)
quartz comments on Q&A with new Executive Director of Singularity Institute - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (177)
How are you going to address the perceived and actual lack of rigor associated with SIAI?
There are essentially no academics who believe that high-quality research is happening at the Singularity Institute. This is likely to pose problems for your plan to work with professors to find research candidates. It is also likely to be an indicator of little high-quality work happening at the Institute.
In his recent Summit presentation, Eliezer states that "most things you need to know to build Friendly AI are rigorous understanding of AGI rather than Friendly parts per se". This suggests that researchers in AI and machine learning should be able to appreciate high-quality work done by SIAI. However, this is not happening, and the publications listed on the SIAI page--including TDT--are mostly high-level arguments that don't meet this standard. How do you plan to change this?
David Chalmers has said that the decision theory work is a major advance (along with various other philosophers), although he is frustrated that it hasn't been communicated more actively to the academic decision theory and philosophy communities. A number of current and former academics, including David, Stephen Omohundro, James Miller (above), and Nick Bostrom have reported that work at SIAI has been very helpful for their own research and writing in related topics.
Evan Williams, now a professor of philosophy at Purdue cites, in his dissertation, three inspirations leading to the work: John Stuart Mill's "On Liberty," John Rawls' "Theory of Justice," and Eliezer Yudkowsky's "Creating Friendly AI" (2001), discussed at greater length than the others. Nick Beckstead, a Rutgers (#2 philosophy program) philosophy PhD student who works on existential risks and population ethics reported large benefits to his academic work from discussions with SIAI staff.
These folk are a minority, and SIAI is not well integrated with academia (no PhDs on staff, publishing, etc), but also not negligible.
I think that work in this area has been disproportionately done by Eliezer Yudkowsky, and to a lesser extent Marcello Herreshoff. Eliezer has been heavily occupied with Overcoming BIas, Less Wrong, and his book for the last several years, in part to recruit a more substantial team for this. He also is reluctant to release work that he thinks is relevant to building AGI. Problems in recruiting and the policies of secrecy seem like the big issues here.
Eliezer's investment into OB/LW apparently hasn't returned even a single full-time FAI researcher for SIAI after several years (although a few people are almost certainly doing more and better FAI-related research than if the Sequences didn't happen). Has this met SIAI's initial expectations? Do you guys think we're at the beginning of a snowball effect, or has OB/LW pretty much done as much as it can, as far as creating/recruiting FAI researchers is concerned? What are your current expectations for the book in this regard?
I have noticed increasing numbers of very talented math and CS folk expressing interest or taking actions showing significant commitment. A number of them are currently doing things like PhD programs in AI. However, there hasn't been much of a core FAI team and research program to assimilate people into. Current plans are for Eliezer to switch back to full time AI after his book, with intake of more folk into that research program. Given the mix of people in the extended SIAI community, I am pretty confident that with abundant funding a team of pretty competent researchers (with at least some indicators like PhDs from the top AI/CS programs, 1 in 100,000 or better performance on mathematics contests, etc) could be mustered over time, based on people I already know.
I am less confident that a team can be assembled with so much world-class talent that it is a large fraction of the quality-adjusted human capital applied to AGI, without big gains in recruiting (e.g. success with the rationality book or communication on AI safety issues, better staff to drive recruiting, a more attractive and established team to integrate newcomers, relevant celebrity endorsements, etc). The Manhattan Project had 21 then- or future Nobel laureates. AI, and certainly FAI, are currently getting a much, much smaller share of world scientific talent than nukes did, so that it's easier for a small team to loom large, but it seems to me like there is still a lot of ground to be covered to recruit a credibly strong FAI team.
Thanks. You didn't answer my questions directly, but it sounds like things are proceeding more or less according to expectations. I have a couple of followup questions.
At what level of talent do you think an attempt to build an FAI would start to do more (expected) good than harm? For simplicity, feel free to ignore the opportunity cost of spending financial and human resources on this project, and just consider the potential direct harmful effects, like accidentally creating an UFAI while experimenting to better understand AGI, or building a would-be FAI that turns out to be an UFAI due to a philosophical, theoretical or programming error, or leaking AGI advances that will allow others to build an UFAI, or starting an AGI arms race.
I have a serious concern that if SIAI ever manages to obtain abundant funding and a team of "pretty competent researchers" (or even "world-class talent", since I'm not convinced that even a team of world-class talent trying to build an FAI will do more good than harm), it will proceed with an FAI project without adequate analysis of the costs and benefits of doing so, or without continuously reevaluating the decision in light of new information. Do you think this concern is reasonable?
If so, I think it would help a lot if SIAI got into the habit of making its strategic thinking more transparent. It could post answers to questions like the ones I asked in the grandparent comment without having to be prompted. It could publish the reasons behind every major strategic decision, and the metrics it keeps to evaluate its initiatives. (One way to do this, if such strategic thinking often occurs or is presented at board meetings, would be to publish the meeting minutes, as I suggested in another comment.)
I'm not sure that scientific talent is the relevant variable here. More talented folk are more likely to achieve both positive and negative outcomes. I would place more weight on epistemic rationality, motivations (personality, background checks), institutional setup and culture, the strategy of first trying to get test the tractability of robust FAI theory and then advancing FAI before code (with emphasis on the more-FAI-less-AGI problems first), and similar variables.
Certainly it's a reasonable concern from a distance. Folk do try to estimate and reduce the risks you mentioned, and to investigate alternative non-FAI interventions. My personal sense is that these efforts have been reasonable but need to be bolstered along with the FAI research team. If it looks like a credible (to me) team may be assembled my plan would be (and has been) to monitor and influence team composition, culture, and exposure to information. In other words, I'd like to select folk ready to reevaluate as well as to make progress, and to work hard to build that culture as researchers join up.
I can't speak for everyone, but I am happy to see SIAI become more transparent in various ways. The publication of the strategic plan is part of that, and I believe Luke is keen (with encouragement from others) to increase communication and transparency in other ways.
This one would be a decision for the board, but I'll give my personal take again. Personally, I like the recorded GiveWell meetings and see the virtues of transparency in being more credible to observers, and in providing external incentives. However, I would also worry that signalling issues with a diverse external audience can hinder accurate discussion of important topics, e.g. frank discussions of the strengths and weaknesses of potential Summit speakers, partners, and potential hires that could cause hurt feelings and damage valuable relationships. Because of this problem I would be more wholehearted in supporting other forms of transparency, e.g. more frequent and detailed reporting on activities, financial transparency, the strategic plan, things like Luke's Q&A, etc. But I wouldn't be surprised if this happens too.
Let's assume that all the other variables are already optimized for to minimize the risk of creating an UFAI. It seems to me that the the relationship between the ability level of the FAI team and probabilities of the possible outcomes must then look something like this:
This chart isn't meant to communicate my actual estimates of the probabilities and crossover points, but just the overall shapes of the curves. Do you disagree with them? (If you want to draw your own version, click here and then click on "Modify This Chart".)
Has anyone posted SIAI's estimates of those risks?
That seems reasonable, and given that I'm more interested in the "strategic" as opposed to "tactical" reasoning within SIAI, I'd be happy for it to be communicated through some other means.
I like this chart.
If we condition on having all other variables optimized, I'd expect a team to adopt very high standards of proof, and recognize limits to its own capabilities, biases, etc. One of the primary purposes of organizing a small FAI team is to create a team that can actually stop and abandon a line of research/design (Eliezer calls this "halt, melt, and catch fire") that cannot be shown to be safe (given limited human ability, incentives and bias). If that works (and it's a separate target in team construction rather than a guarantee, but you specified optimized non-talent variables) then I would expect a big shift of probability from "UFAI" to "null."
What I'm afraid of is that a design will be shown to be safe, and then it turns out that the proof is wrong, or the formalization of the notion of "safety" used by the proof is wrong. This kind of thing happens a lot in cryptography, if you replace "safety" with "security". These mistakes are still occurring today, even after decades of research into how to do such proofs and what the relevant formalizations are. From where I'm sitting, proving an AGI design Friendly seems even more difficult and error-prone than proving a crypto scheme secure, probably by a large margin, and there is no decades of time to refine the proof techniques and formalizations. There's good recent review of the history of provable security, titled Provable Security in the Real World, which might help you understand where I'm coming from.
Your comment has finally convinced me to study some practical crypto because it seems to have fruitful analogies to FAI. It's especially awesome that one of the references in the linked article is "An Attack Against SSH2 Protocol" by W. Dai.
Correct me if I'm wrong, but it doesn't seem as though "proofs" of algorithm correctness fail as frequently as "proofs" of cryptosystem unbreakableness.
Where does your intuition that friendliness proofs are on the order of reliability of cryptosystem proofs come from?
Thanks for clarifying.
I agree.
Could you elaborate on the ability axis. Could you name some people that you perceive to be of world class ability in their field. Could you further explain if you believe that there are people who are sufficiently above that class.
For example, what about Terence Tao? What about the current SIAI team?
Basically it ensures that all serious discussion and decision making is made prior to any meeting in informal conversations so that the meeting sounds good. Such a record should be considered a work of fiction regardless of whether it is a video transcript or a typed document. (Only to the extent that the subject of the meeting matters - harmless or irrelevant things wouldn't change.)
That's more like it!
I'll take this opportunity to mention that I'm against publishing SIAI's board meeting minutes. First, for the reasons Carl gave above. Second, because then we'd have to invest a lot of time explaining the logic behind each decision, or else face waves of criticism for decisions that appear arbitrary when one merely publishes the decision and not the argument.
However, I'm definitely making big effort to improve SIAI transparency. Our new website (under development) has a page devoted to transparency, where you'll be able to find our strategic plan, our 990s, and probably other links. I'm also publishing the monthly progress reports, and recently co-wrote 'Intelligence Explosion: Evidence and Import', which for the first time (excepting Chalmers) summarizes many of our key pieces of reasoning with the clarity of mainstream academic form. We're also developing an annual report, and I'm working toward developing some other documents that will make SIAI strategy more transparent. But all this takes time, especially when starting from pretty close to 0 on transparency, and having lots of other problems to fix, too.
Are the arguments not made during the board meetings? Or do you guys talk ahead of time and just formalize the decisions during the board meetings?
In any case, I think you should invest more time explaining the logic behind your decisions, and not just make the decisions themselves more transparent. If publishing board meeting minutes is not the best way to do that, then please think about some other way of doing it. I'll list some of the benefits of doing this, in case you haven't thought of some of them:
The arguments are strewn across dozens of conversations in and out of board meetings (mostly out).
As for finding other ways to explain the logic behind our decisions, I agree, and I'm working on it. One qualification I would add, however, is that I predict more benefit to my strategic thinking from one hour with Paul Christiano and one hour with Nick Bostrom than from spending four hours to write up my strategic thinking on subject X and publishing it so that passersby can comment on it. It takes a lot of effort to be so well-informed about these issues that one can offer valuable strategic advice. But for some X we have already spent those many productive hours with Christiano and Bostrom and so on, and it's a good marginal investment to write up our strategic thinking on X.
This reminds me a bit of Eliezer's excuse when he was resisting calls for him to publish his TDT ideas on LW:
I suggest you may be similarly overestimating the difficulty of explaining your strategic ideas/problems to a sufficiently large audience to get useful feedback. Why not just explain them the same way that you would explain to Christiano and Bostrom? If some among the LW community don't understand, they can ask questions and others could fill them in.
The decision theory discussions on LW generated significant progress, but perhaps more importantly created a pool of people with strong interest in the topic (some of whom ended up becoming your research associates). Don't you think the same thing could happen with Singularity strategies?
Minutes can be much more general than (video) transcripts.
I would be surprised if the optimal solution isn't a third alternative and is instead total secrecy or manipulable complete transcription.
I believe that the SIAI has has been very successful in using OB/LW to not only rise awareness of risks from AI but to lend credence to the idea. From the very beginning I admired that feat.
Eliezer Yudkowsky's homepage is a perfect example of its type. Just imagine he would have concentrated solely on spreading the idea of risks from AI and the necessity of a friendliness theory. Without any background relating to business or an academic degree, to many people he would appear to be yet another crackpot spreading prophecies of doom. But someone who is apparently well-versed in probability theory, who studied cognitive biases and tries to refine the art of rationality? Someone like that can't possible be deluded enough to hold some complex beliefs that are completely unfounded, there must be more to it.
That's probably the biggest public relations stunt in the history of marketing extraordinary ideas.
Certainly, by many metrics LW can be considered wildly successful, and my comment wasn't meant to be a criticism of Eliezer or SIAI. But if SIAI was intending to build an FAI using its own team of FAI researchers, then at least so far LW has failed to recruit them any such researchers. I'm trying to figure out if this was the expected outcome, and if not, how updating on it has changed SIAI's plans. (Or to remind them to update in case they forgot to do so.)
Most of your analysis seems right, but the last sentence seems likely to be off. There have been a lot of clever PR stunts in history.
Most of them have not been targeting smart and educated nonconformists. Eliezer successfully changed people's mind by installing a way of thinking (a framework of heuristics, concepts and ideas) that is fine-tuned to non-obviously culminate in one inevitable conclusion, that you want to contribute money to his charity because it is rational to do so.
Take a look at the sequences in the light of the Singularity Institute. Even the Quantum Sequence helps to hit a point home that is indispensable to convince people, who would otherwise be skeptical, that it is rational to take risks from AI seriously. The Sequences promulgate that logical implications of general beliefs you already have do not cost you extra probability and that it would be logically rude to demand some knowably unobtainable evidence.
A true masterpiece.
I have informally been probing smart people I meet whether they're aware of LW. The answers have been surprisingly high number of 'Yes'. I expect this is already making impact on, at the very least, a less risky distribution of funding sources, and probably a good increase in funding once some of them (as many are in startups) will hit paydirt.
Sooner or later he will have to present some results. As the advent of AGI is moving closer people will start to panic and demand hard evidence that the SIAI is worth their money. Even someone who has published a lot of material on rationality and a popular fanfic will run out of credit and people will stop taking his word for it.
I believe that high-quality research is happening at the Singularity Institute.
James Miller, Associate Professor of Economics, Smith College.
PhD, University of Chicago.
To distinguish the above from the statement "I like the Singularity Institute", could you be specific about what research activities you have observed in sufficient detail to confidently describe as "high-quality"?
ETA: Not a hint of sarcasm or snark intended, I'm sincerely curious.
I'm currently writing a book on the Singularity and have consequently become extremely familiar with the organization's work. I have gone through most of EY's writings and have an extremely high opinion of them. His research on AI plays a big part in my book. I have also been ending my game theory classes with "rationality shorts" in which I present some of EY's material from the sequences.
I also have a high opinion of Carl Shulman's (an SI employee) writings including “How Hard is Artificial Intelligence? The Evolutionary Argument and Observation Selection Effects." (Co-authored with Bostrom) and Shulman's paper on AGI and arms races.
Luke discussed this a while back here.
I agree that this is an important question.
This is my favorite of the questions so far.
A clarifying question. By 'rigor', do you mean the kind of rigor that is required to publish in journals like Risk Analysis or Minds and Machines, or do you mean something else by 'rigor'?
I mean the kind of precise, mathematical analysis that would be required to publish at conferences like NIPS or in the Journal of Philosophical Logic. This entails development of technical results that are sufficiently clear and modular that other researchers can use them in their own work. In 15 years, I want to see a textbook on the mathematics of FAI that I can put on my bookshelf next to Pearl's Causality, Sipser's Introduction to the Theory of Computation and MacKay's Information Theory, Inference, and Learning Algorithms. This is not going to happen if research of sufficient quality doesn't start soon.
My day brightened imagining that!
Thanks for clarifying.
Addendum: Since the people who upvoted the question were in the same position as you with respect to its interpretation, it would be good to not only address my intended meaning, but all major modes of interpretation.
I can't speak for the original questioner, but take for example the latest post by Holden Karnofsky from GiveWell. I would like to see a response by the SIAI that applies the same amount of mathematical rigor to show that it actually is the rational choice from the point of view of charitable giving.
A potential donor might currently get the impression that the SIAI has written a lot of rather colloquial posts on rationality than rigorous papers on the nature of AGI, not to mention friendly AI. In contrast, GiveWell appears to concentrate on their main objective, the evaluation of charities. In doing so they are being strictly technical, an appraoch that introduces a high degree of focus by tabooing colloquial language and thereby reducing ambiguity, while allowing others to review their work.
Some of the currently available papers might, in a less favorably academic context, be viewed as some amount of handwaving mixed with speculations.