riceissa

Karma: 452

riceissa 12 Sep 2019 3:24 UTC
3 points

on: Conversation with Paul Christiano

like, my views aren’t that internally coherent. My suspicion is others’ views are even less internally coherent.

I would appreciate hearing more concretely what it means to be internally coherent/incoherent (e.g. are there examples of contradictory statements a single person is making?).

riceissa 12 Sep 2019 3:19 UTC
6 points

on: Conversation with Paul Christiano

It’s lots of saving throws, you know? And you multiply the saving throws together and things look better. And they interact better than that because– well, in one way worse because it’s correlated: If you’re incompetent, you’re more likely to fail to solve the problem and more likely to fail to coordinate not to destroy the world. In some other sense, it’s better than interacting multiplicatively because weakness in one area compensates for strength in the other. I think there are a bunch of saving throws that could independently make things good, but then in reality you have to have a little bit here and a little bit here and a little bit here, if that makes sense.

I don’t understand this part. Translating to math, I think it’s saying something like, if $p_{i}$ is the probability that saving throw $i$ works, then the probability that at least one of them works is $1 - \prod_{i} (1 - p_{i})$ (assuming the saving throws are independent), which is higher the more saving throws there are; but due to correlation, the saving throws are not independent, so we effectively have fewer saving throws. I don’t understand what “weakness in one area compensates for strength in the other” or “a little bit here and a little bit here and a little bit here” mean.

riceissa 12 Sep 2019 2:26 UTC
6 points

on: Conversation with Paul Christiano

Christiano cares more about making aligned AIs that are competitive with unaligned AIs, whereas MIRI is more willing to settle for an AI with very narrow capabilities.

Looking at the transcript, it seems like “AI with very narrow capabilities” is referring to the “copy-paste a strawberry” example. It seems to me that the point of the strawberry example (see Eliezer’s posts 1, 2, and Dario Amodei’s comment here) is that by creating an AGI that can copy and paste a strawberry, we necessarily solve most of the alignment problem. So it isn’t the case that MIRI is aiming for an AI with very narrow capabilities (even task AGI is supposed to perform pivotal acts).

riceissa 11 Sep 2019 4:06 UTC
1 point

in reply to: Wei_Dai's comment on: Counterfactual Oracles = online supervised learning with random selection of training episodes
Thanks! I think I understand this now.

I will say some things that occurred to me while thinking more about this, and hope that someone will correct me if I get something wrong.
- “Human imitation” is sometimes used to refer to the outward behavior of the system (e.g. “imitation learning”, and in posts like “Just Imitate Humans?”), and sometimes to refer to the model of the human inside the system (e.g. here when you say “the human imitation is telling the system what to do”).
- A system that is more capable than a human can still be a “human imitation”, because “human imitation” is being used in the sense of “modeling humans inside the system” instead of “has the outward behavior of a human”.
- There is a distinction between the counterfactual training procedure vs the resulting system. “Counterfactual oracle” (singular) seems to be used to refer to the resulting system, and Paul calls this “the system” in his “Human-in-the-counterfactual-loop” post. “Counterfactual oracles” (plural) is used both as a plural version of the resulting system and also as a label for the general training procedure. “Human-in-the-counterfactual-loop”, “counterfactual human oversight”, and “counterfactual oversight” all refer to the training procedure (but only when the procedure uses a model of the human).

riceissa 11 Sep 2019 1:50 UTC
1 point

on: Counterfactual Oracles = online supervised learning with random selection of training episodes
Paul Christiano does have a blog post titled Counterfactual oversight vs. training data, which talks about the same thing as this post except that he uses the term “counterfactual oversight”, which is just Counterfactual Oracles applied to human imitation (which he proposes to use to “oversee” some larger AI system).

I am having trouble parsing/understanding this part.
- The linked post by Paul doesn’t seem to talk about human imitation. Is there a separate post/comment somewhere that connects counterfactual oversight to human imitation, or is the connection to human imitation somehow implicit in the linked post?
- The linked post by Paul seems to be talking about counterfactual oversight as a way to train the counterfactual oracle, but I’m parsing your sentence as saying that there is then a further step where the counterfactual oracle is used to oversee a larger AI system (i.e. the “oversee” in your sentence is different from the “oversight” in “counterfactual oversight”). Is this right?

riceissa 11 Sep 2019 1:20 UTC
1 point

on: Iterated Distillation and Amplification
I noticed that I have two distinct “mental pictures” for what the overseer is, depending on how the Distill procedure works (i.e. depending on the narrow technique used in the Distill procedure).
1. For imitation learning and narrow inverse reinforcement learning: a “passive” overseer that just gets used as a template/target for imitation.
2. For narrow reinforcement learning and in discussions about approval-directed agents: an “active” overseer that rates actions or provides rewards.
I wonder if this way of thinking about the overseer is okay/correct, or if I’m missing something (e.g. maybe even in case (1), the overseer has a more active role than I can make out). Assuming this way of thinking about the overseer is okay, it seems like for case (1), the term “overseer” has connotations that extend beyond the role played by the overseer (i.e. it doesn’t really provide any oversight since it is passive).

riceissa 30 Aug 2019 5:47 UTC
4 points

on: Iterated Distillation and Amplification
Based on discussion between Vladimir Slepnev and Paul in this thread, it seems like statements in this post (“we assume that A[0] can acquire nearly human-level capabilities through this process”, “Given an aligned agent H we can use narrow safe learning techniques to train a much faster agent A which behaves as H would have behaved”) that the first stage of IDA will produce nearly-human-level assistants are misleading. In the same thread, Paul says that he “will probably correct it”, but as far as I can tell, neither the Medium post nor the version of the post in this sequence (which was published after the discussion) has been corrected.

riceissa 30 Aug 2019 5:32 UTC
2 points

in reply to: rmoehn's comment on: Iterated Distillation and Amplification
I had this same thought, but my understanding (which is not solid) is that in the first iteration, since A is random, H can just ignore A and go with its own output (if my assistants are unhelpful, I can just try to perform the task all on my own). So Amplify(H, A) becomes H, which means A ← Distill(Amplify(H, A)) is basically A ← Distill(H), exactly as you suggested.

riceissa 30 Aug 2019 5:28 UTC
2 points

on: Paul’s research agenda FAQ
I’m still confused about the difference between HCH and the amplification step of IDA. Initially I thought that the difference is that with HCH, the assistants are other copies of the human, whereas in IDA the assistants are the distilled agents from the previous step (whose capabilities will be sub-human in early stages of IDA and super-human in later stages). However, this FAQ says “HCHs should not be visualized as having humans in the box.”

My next guess is that while HCH allows the recursion for spawning new assistants to be arbitrarily deep, the amplification step of IDA only allows a single level of spawning (i.e. the human can spawn new assistants, but the assistants themselves cannot make new assistants). Ajeya Cotra’s post on IDA talks about the human making calls to the assistant, but not about the assistants making further calls to other assistants, so it seems plausible to me that the recursive nature of HCH is the difference. Can someone who understands HCH/IDA confirm that this is a difference and/or name other differences?

riceissa 19 Aug 2019 22:22 UTC
5 points

on: GreaterWrong Arbital Viewer
The page https://arbital.greaterwrong.com/p/AI_safety_mindset/ is blank in the GreaterWrong version, but has content in the obormot.net version.

Inversion of theorems into definitions when generalizing

riceissa

4 Aug 2019 17:44 UTC

24 points

3 comments5 min readLW link

riceissa 28 Jul 2019 21:06 UTC
12 points

in reply to: Wei_Dai's comment on: AALWA: Ask any LessWronger anything
I was surprised to see, both on your website and the white paper, that you are part of Mercatoria/ICTP (although your level of involvement isn’t clear based on public information). My surprise is mainly because you have a couple of comments on LessWrong that discuss why you have declined to join MIRI as a research associate. You have also (to my knowledge) never joined any other rationality-community or effective altruism-related organization in any capacity.

My questions are:
1. What are the reasons you decided to join or sign on as a co-author for Mercatoria/ICTP?
2. More generally, how do you decide which organizations to associate with? Have you considered joining other organizations, starting your own organization, or recruiting contract workers/volunteers to work on things you consider important?

riceissa 1 Jul 2019 5:26 UTC
7 points

on: What’s the most “stuck” you’ve been with an argument, that eventually got resolved?
“Sam Harris and the Is–Ought Gap” might be one example.

riceissa 29 Jun 2019 3:34 UTC
3 points

on: GreaterWrong Arbital Viewer
Some of the math seems to be garbled, e.g. this page (compare with the version on obormot.net) and this page (compare with the version on obormot.net).

riceissa 7 Jun 2019 3:20 UTC
10 points

on: Arbital scrape
The scrape seems to be missing the Solomonoff induction dialogue, which was available at https://arbital.com/p/solomonoff_induction/?l=1hh (at the moment I just get an error).

ETA (2019-06-28): The new version of the scrape has this page, and can be viewed on the GreaterWrong version.

riceissa 28 May 2019 4:31 UTC
15 points

on: Was CFAR always intended to be a distinct organization from MIRI?
Going by public documents, it seems like the intention was to have a separate organization pretty early on (although I can’t say about the very beginning since I wasn’t involved).

The Articles of Incorporation for CFAR (called Feynman Foundation at the time) are dated July 18, 2011, and 501(c)(3) status was approved on July 26, 2011.

MIRI’s 2011 strategic plan (updated August 2011; unclear when it was first written) says “Encourage a new organization to begin rationality instruction similar to what Singularity Institute did in 2011 with Rationality Minicamp and Rationality Boot Camp.”

The Minicamp took place in May–June 2011, and it seems like the Boot Camp took place in June–August 2011, so it looks like by the time the first workshop finished and the second was in progress, the plan was already to start the new organization. However, it’s still possible that there was no plan for a separate organization before or during the first workshop.

MIRI’s December 2011 progress report also talks about plans to create the “Rationality Org”.

I also wrote a timeline of CFAR a while back, which has more links.

riceissa 20 May 2019 9:12 UTC
2 points

in reply to: Said Achmiz's comment on: “One Man’s Modus Ponens Is Another Man’s Modus Tollens”
The examples in the buckets error post have “modus delens” as the correct response. To take the diet example from the post, A = “diet worth being on”, B = “zero studies suggesting health risks”. Adam has $A ⟹ B$ stored in his brain, and Betty presents $\neg B$ , so Adam’s brain computes $\neg A$ . The “protecting epistemology” move is to instead adamantly believe A (“I need to stay motivated!”) which ends up rejecting what Betty said. But the desired response is to instead deny B but also accept A, and hence to deny the implication $A ⟹ B$ .

So in these buckets error examples, modus ponens corresponds to the “automatic” reasoning, modus tollens corresponds to the “flinching away from the truth” move, and modus delens corresponds to the “rational” move that avoids the buckets error.

I explained this more in a comment on the post.

I can’t comment as to the relative frequency of this response and how often it’s correct (this sort of question seems difficult to answer).

riceissa 19 May 2019 15:44 UTC
5 points

in reply to: Zvi's comment on: Announcement: AI alignment prize round 4 winners
Was this post ever published? (Looking at Zvi’s posts since January, I don’t see anything that looks like it.)

riceissa 14 May 2019 20:04 UTC
15 points

on: Flashcards for AI Safety?
I’ve made around 250 Anki cards about AI safety. I haven’t prioritized sharing my cards because I think finding a specific card useful requires someone to have read the source material generating the card (e.g. if I made the card based on a blog post, one would need to read that exact blog post to get value out of reviewing the card; see learn before you memorize). Since there are many AI safety blog posts and I don’t have the sense that lots of Anki users read any particular blog post, it seems to me that the value generated from sharing a set of cards about a blog post isn’t high enough to overcome the annoyance cost of polishing, packaging, and uploading the cards.

More generally, from a consumer perspective, I think people tend to be pretty bad at making good Anki cards (I’m often embarrassed at the cards I’ve created several months ago!), which makes it unexciting for me to spend a lot of effort trying to collaborate with others on making cards (because I expect to receive poorly-made cards in return for the cards I provide). I think collaborative card-making can be done though, e.g. Michael Nielsen and Andy Matuschak’s quantum computing guide comes with pre-made cards that I think are pretty good.

Different people also have different goals/interests so even given a single source material, the specifics one wants to Ankify can be different. For example, someone who wants to understand the technical details of logical induction will want to Ankify the common objects used (market, pricing, trader, valuation feature, etc.), the theorems and proof techniques, and so forth, whereas someone who just wants a high-level overview and the “so what” of logical induction can get away with Ankifying much less detail.

Something I’ve noticed is that many AI safety posts aren’t very good at explaining things (not enough concrete examples, not enough emphasis on common misconceptions and edge cases, not enough effort to answer what I think of as “obvious” questions); this fact is often revealed by the comments people make in response to a post. This makes it hard to make Anki cards because one doesn’t really understand the content of the post, at least not well enough to confidently generate Anki cards (one of the benefits of being an Anki user is having a greater sensitivity to when one does not understand something; see “illusion of explanatory depth” and related terms). There are other problems like conflicting usage of terminology (e.g. multiple definitions of “benign”, “aligned”, “corrigible”) and the fact that some of the debates are ongoing/some of the knowledge is still being worked out.

For “What would be a good strategy for generating useful flashcards?”: I try to read a post or a series of posts and once I feel that I understand the basic idea, I will usually reread it to add cards about the basic terms and ask myself simple questions. Some example cards for iterated amplification:
- what kind of training does the Distill step use?
- in the pseudocode, what step gets repeated/iterated?
- how do we get A[0]?
- write A[1] in terms of H and A[0]
- when Paul says IDA is going to be competitive with traditional RL agents in terms of time and resource costs, what exactly does he mean?
- advantages of A[0] over H
- symbolic expression for the overseer
- why should the amplified system (of human + multiple copies of the AI) be expected to perform better than the human alone?

riceissa 5 May 2019 20:18 UTC
18 points

on: How much do major foundations grant per hour of staff time?
You might have in mind a Facebook post that Vipul Naik wrote in 2017.

riceissa

In­ver­sion of the­o­rems into defi­ni­tions when generalizing

Inversion of theorems into definitions when generalizing