Back in December 2013, Jonah Sinick and I launched Cognito Mentoring, an advising service for intellectually curious students. Our goal was to improve the quality of learning, productivity, and life choices of the student population at large, and we chose to focus on intellectually curious students because of their greater potential as well as our greater ability to relate with that population. We began by offering free personalized advising. Jonah announced the launch in a LessWrong post, hoping to attract the attention of LessWrong's intellectually curious readership.

Since then, we feel we've done a fair amount, with a lot of help from LessWrong. We've published a few dozen blog posts and have an information wiki. Slightly under a hundred people contacted us asking us for advice (many from LessWrong), and we had substantive interactions with over 50 of them. As our reviews from students and parents suggest, we've made a good impression and have had a positive impact on many of the people we've advised. We're proud of what we've accomplished and grateful for the support and constructive criticism we've received on LessWrong.

However, what we've learned in the last few months has led us to the conclusion that Cognito Mentoring is not ripe for being a full-time work opportunity for the two of us.

For the last few months, we've eschewed regular jobs and instead done contract work that provides us the flexibility to work on Cognito Mentoring, eating into our savings somewhat to cover the cost of living differences. This is a temporary arrangement and is not sustainable. We therefore intend to scale back our work on Cognito Mentoring to "maintenance mode" so that people can continue to benefit from the resources we've already collected, with minimal additional effort on our part, freeing us up to take regular jobs with more demanding time requirements.

We might revive Cognito Mentoring as a part-time or full-time endeavor in the future if there are significant changes to our beliefs about the traction, impact, and long-run financial viability of Cognito Mentoring. Part of the purpose of "maintenance mode" will be to leave open the possibility of such a revival if the idea does indeed have potential.

In this post, I discuss some of the factors that led us to change our view, the conditions under which we might revive Cognito Mentoring, and more details about how "maintenance mode" for Cognito Mentoring will look.

Reason #1: Downward update on social value

We do think that the work we've done on Cognito Mentoring so far has generated social value, and the continued presence of the website will add more value over time. However, our view has shifted in the direction of lower marginal social value from working on Cognito Mentoring full-time, relative to simply keeping the website live and doing occasional work to improve it. Specifically:

• It's quite possible that the lowest-hanging fruit with respect to the advisees who would be most receptive to our advice has already been plucked. We received the bulk of our advisees through LessWrong within the month after our initial posting. Other places where we've posted about our service have led to fewer advisees (more here).
• Of our website content, only a small fraction of the content gets significant traction (see our list of popular pages), so honing and promoting our best content might be a better strategy for improving social value than trying to create a comprehensive resource. This can be done while in maintenance mode, and does not require full-time effort on our part.

What might lead us to change our minds: If we continue to be contacted by large numbers of potentially high-impact people, or we get evidence that the advising we've already done has had significantly greater impact than we think it did, we'll update our social value upward.

Reason #2: Downward update on long-run financial viability

We have enough cash to go on for a few more months. But for Cognito Mentoring to be something that we work full time on, we need an eventual steady source of income from it. Around mid-March 2014, we came to the realization that charging advisees is not a viable revenue source, as Jonah described at the end of his post about how Cognito Mentoring can do the most good (see also this comment by Luke Muehlhauser and Jonah's response to it below the comment). At that point, we decided to focus more on our informational content and on looking for philanthropic funding.

Our effort at looking into philanthropic funding did give us a few leads, and some of them could plausibly result in us getting small grants. However, none of the leads we got pointed to potential steady long-term income sources. In other words, we don't think philanthropic funding is a viable long-term revenue model for Cognito Mentoring.

Our (anticipated) difficulty in getting philanthropic funding arises from two somewhat different reasons.

1. What we're doing is somewhat new and does not fit the standard mold of educational grants. Educational foundations tend to give grants for fairly specific activities, and what we're doing does not seem to fit those.
2. We haven't demonstrated significant traction or impact yet (even though we've had a reasonable amount of per capita impact, the total number of people we've influenced so far is relatively small). This circles back to Reason #1: funders' reluctance to fund us may in part stem from their belief that we won't have much social value, given our lack of traction so far. Insofar as funders' judgment carries some information value, this should also strengthen Reason #1.

What might lead us to change our minds: If we are contacted by a funder who is willing to bankroll us for over a year and also offer a convincing reason for why he/she thinks bankrolling us is a good idea (so that we're convinced that our funding can be sustained beyond a year) we'll change our minds.

Reason #3: Acquisition of knowledge and skills

One of the reasons we've been able to have an impact through Cognito Mentoring so far is that both Jonah and I have knowledge of many diverse topics related to the questions that our advisees have posed to us. But our knowledge is still woefully inadequate in a number of areas. In particular, many advisees have asked us questions in the realms of technology, entrepreneurship, and the job environment, and while we have pointed them to resources on these, firsthand experience, or close secondhand experience, would help us more effectively guide advisees. We intend to take jobs related to computer technology (in fields such as programming or data science), and these jobs might be at startups or put us in close contact with startups. This will better position us to return to mentoring later if we choose to resume it part-time or full-time.

Knowledge and skills we acquire working in the technology sector could also help us design better interfaces or websites that can more directly address the needs of our audience. So far, we've thought of ourselves as content-oriented people, so we've used standard off-the-shelf software such as WordPress (for our main website and blog) and MediaWiki (for our information wiki). Part of the reason is that we wanted to focus on content creation rather than interface design, but part of the reason we've stuck to these is that we didn't think we could design interfaces. Once we've acquired more programming and design experience, we might be more open to the idea of designing interfaces and software that can meet particular needs of our target audience.We might design an interface that helps people study more effectively, make better life decisions, or share reviews of courses and colleges, in a manner similar to softwares or websites such as Anki or Beeminder or Goodreads. There might also be potential for a more effective online resource that teaches programming than those in existence (e.g. Codecademy). It's not clear right now whether there exists a useful opportunity of this sort that we are particularly well-suited to, but with more coding experience, we'll at least be able to implement an idea of this sort if we decide it has promise.

Reason #4: Letting it brew in the background can give us a better idea of the potential

If we continue to gradually add content to the wiki, and continue to get links and traffic to it from other sources, it's likely that the traffic will grow slowly and steadily. The extent of organic growth will help us figure out how much promise Cognito Mentoring has. If our wiki gets to the point of steadily receiving thousands of pageviews a day, we will reconsider reviving Cognito Mentoring as a part-time or full-time endeavor. If, on the other hand, traffic remains at approximately the current level (about a hundred pageviews a day, once we exclude spikes arising from links from LessWrong and Marginal Revolution) then the idea is probably not worth revisiting, and we'll leave it in maintenance mode.

In addition, by maintaining contact with the people we've advised, we can get more insight into the sort of impact we've had, whether it is significant over the long term, and how it can be improved. This again can tell us whether our impact is sufficiently large as to make Cognito Mentoring worth reviving.

What "maintenance mode" entails

1. We'll continue to have contact information available, but will scale back on personalized advising: People are welcome to contact us with questions and suggestions about content, but we will not generally offer detailed personalized responses or do research specific to individuals who contact us. We'll attempt to point people to relevant content we've already written, or to other resources we're already aware of that can address their concerns.
2. The information wiki will remain live, and we will continue to make occasional improvements, but we won't have a time schedule of when particular improvements have to be implemented by.
3. Existing blog posts will remain, but we probably won't be making many new blog posts. New blog posts will happen only if one of us has an idea that really seems worth sharing and for which the Cognito Mentoring blog is an ideal forum.
4. We'll continue our administrative roles in the communities of existing Cognito Mentoring advisees
5. We'll continue periodically reviewing the progress of people we've advised so far: This will help us get a better sense of how valuable our work has been, and can be useful should we choose to revive Cognito Mentoring.
6. We'll continue to correspond with advisees we have so far (time permitting), though we'll give more priority to advisees who continue to maintain contact of their own accord and those whose activities seem to have higher impact potential.
7. We'll try to get our best content linked from other sources, such as about.com: Sources like about.com are targeted at the general population. We can try to get linked to from there as an additional resource for the more intellectually curious population that's outside the core focus of about.com.
8. We'll link more extensively to other sources that people can use: For instance, we can more emphatically point to 80,000 Hours for people who are interested in career advising in relation to effective altruist pursuits. We can point to about.com and College Confidential for more general information about mainstream institutions. We already make a number of recommendations on our website, but as we stop working actively, it becomes all the more important that people who come to us are appropriately redirected to other sources that can help them.

Conclusion and summary (TL;DR)

We (qua Cognito Mentoring) are grateful to LessWrong for being welcoming of our posts, offering constructive criticism, and providing us with some advisees we've enjoyed working with. We think that the work we've done has value, but don't think that there's enough marginal value from full-time work on Cognito Mentoring. We think we can do more good for ourselves and the world by switching Cognito Mentoring to maintenance mode and freeing our time currently spent on Cognito Mentoring for other pursuits. The material that we have already produced will continue to remain in the public domain and we hope that people will benefit from it. We may revisit our "maintenance mode" decision if new evidence changes our view regarding traction, impact, and long-run financial viability.

Consider a simple game: You are shown a random-looking 512-bit string h. You may then press one of two buttons, labeled '0' and '1'. No matter which button you press, you will then be shown a 256-bit string s such that SHA512(s) = h. In addition, if you pressed '1' you are given 1$.

This game seems pretty simple, right? s and h are irrelevant, and you should simply press '1' all the time (I'm assuming you value recieving money). Well, let's see how AIXI and AIXI-tl fare at this game.

Let's say the machine already played the game many times. Its memory is h0, b0, r0, s0, h1, ..., b_(n-1), r_(n-1), s_(n-1), h_n, where the list is in chronological order, inputs are unbolded while decisions are bolded, and r_i is the reward signal. It is always the case that r_i=b_i and h_i=SHA512(s_i).

First let's look at AIXI. It searches for models that compress and extrapolate this history up to the limit of its planning horizon. One class of such models is this: there is a list s0, ..., s_N of random 256-bits strings and a list b0, ..., b_N of (possibly compressible) bits. The history is SHA512(s0), b0, b0, s0, ..., b_(N-1), s_(N-1), SHA512(s_N), b_N. Here s_i for i<n must match the s_i in its memory, and s_n must be the almost certainly unique value with SHA512(s_n) = h_n. While I don't have a proof, it intuitively seems like this class of models will dominate the machines probability mass, and repeated arg-max should lead to the action of outputting 1. It wouldn't always do this due to exploration/exploitation considerations and due to the incentive to minimize K (b0, ... b_N) built into its prior, but it should do it most of the time. So AIXI seems good.

Now let's consider AIXI-tl. It picks outputs by having provably correct programs assign lower bounds to the expected utility, and picking the one with the best lower bound, where expected utility is as measured by AIXI. This would include accepting the analysis I just made with AIXI if that analysis can be made provably accurate. Here lies a problem: the agent has seen h_n but hasn't seen s_n. Therefore, it can't be certain that there is an s_n with SHA512(s_n)=h_n. Therefore, it can't be certain that the models used for AIXI actually works (this isn't a problem for AIXI since it has infinite computational power and can always determine that there is such an s_n).

There is an ad hoc fix for this for AIXI-tl: Take the same model as before, but h_n is a random string rather than being SHA512(s_n). This seems at first to work okay. However, it adds n, the current time, as an input to the model, which adds to its complexity. Now other models dependent on the current time also need to be considered. For instance, what if h_n was a random string, and in addition r_n=not(b_n). This models seems more complicated, but maybe in the programming language used for the Solomonoff prior it is shorter. The key point is that the agent won't be able to update past that initial prior no matter how many times it plays the game.  In that case AIXI-tl may consistently prefer 0, assuming there aren't any other models it considers that make its behavior even more complicated.

The key problem is that AIXI-tl is handling logical uncertainty badly. The reasonable thing to do upon seeing h_n is assuming that it, like all h_i before it, is a hash of some 256-bit string. Instead, it finds itself unable to prove this fact and is forced into assuming the present h_n is special. This makes it assume the present time is special and makes it incapable of learning from experience.

tl;dr Mechanism design studies how to design incentives for fun and profit. A puzzle about whether or not to paint a room is posed. A modeling framework is introduced, with lots of corresponding notation.

Mechanism design is a framework for constructing institutions for group interactions, giving us a language for the design of everything from voting systems to school admissions to auctions to crowdsourcing. Think of it as the engineering side of game theory, building algorithms for strategic agents. In game theory, the primary goal is to answer the question, “Given agents who can take some actions that will lead to some payoffs, what do we expect to happen when the agents strategically interact?” In other words, game theory describes the outcomes of fixed scenarios. In contrast, mechanism design flips the question around and asks, “Given some goals, what payoffs should agents be assigned for the right outcome to occur when agents strategically interact?” The rules of the game are ours to choose, and, within some design constraints, we want to find the best possible ones for a situation.

Although many people, even high-profile theorists, doubt the usefulness of game theory, its application in mechanism design is one of the major success stories of modern economics. Spectrum license auctions designed by economists paved the way for modern cell-phone networks and garnered billions in revenue for the US and European governments. Tech companies like Google and Microsoft employ theorists to improve advertising auctions. Economists like Al Roth and computer scientists like Tuomas Sandholm have been instrumental in establishing kidney exchanges to facilitate organ transplants, while others have been active in the redesign of public school admissions in Boston, Chicago, and New Orleans.

The objective of this post is to introduce all the pieces of a mechanism design problem, providing the setup for actual conclusions later on. I assume you have some basic familiarity with game theory, at the level of understanding the concept of a dominant strategies and Nash equilbria. Take a look at Yvain’s Game Theory Intro if you’d like to brush up. 

Cat, who has volunteered extensively at CFAR (and taught at CFAR), will be visiting many cities in Europe over the coming months.

She is awesome.  

Also, the list of cities that she is visiting will probably be determined in the next few days.

If you'd like to have her visit your LW meetup group, share some of our classes with your meetup, and generally bring connections back and forth... comment below, or PM her (or me)!  I suspect this can be a lot of fun, and useful as well.  Offers of couch space and similar are also appreciated.

For now this is just for Europe, probably, but no harm in touching base from other cities as well; it's possible she'll visit elsewhere later.

I'm not sure this is something that can be consciously done, but in this post I want to prime you to consider whether something you do is really, totally, completely wacky and absurd. 

We have trained ourselves a lot to notice when we are wrong. We trained ourselves even more to notice when we are confused and to tell word confusion from substance confusion. 

But here is the tale of what happened to me today, and I don't think it qualifies as any of those:

I had a serious motivational problem yesterday, and got absolutely nothing done. So today I thought I should do things in a different manner, so as to decrease probability of two bad days in a row. One of the most effective things for me is going into the LW Study Hall (the password is in the group's description when you click this link). A very nice place to work that I recommend for everyone to check out, and do one or two pomo's every now and then.  

And I did, I gave myself ten minutes observing others working, and I noticed something remarkable: The property of the LW chat that causes me to be motivated is "Presence of long haired people". Yes.  Presence of people with a long hair. For weeks I had been trying to work out why it was efficient sometimes and not others. The most obvious initial alternative was that when there was a woman, I would feel more driven. I assumed that was the case. But I started getting false negatives and false positives. Today I finally came to terms with the fact. I am motivated by the presence of people whose hair goes to their shoulders. Women or Men. 

Now why did I not notice this before? Seems to me that basically it was such a far fetched hypothesis that I simply had no prior for it. In vain hopes of being rational, I would read about how we fear the twinge of starting, how to beat procrastination, and how to get things done, and valid as those were and are, they would never have given me a complete picture of the unbelievable things my brain thinks behind my back.

Maybe there is something similar taking place in your mind. Even if there isn't, just update with me on the fact that this is true at least for someone, and how there may be millions of other tiny absurd facts controlling people's actions way beyond the scope of imagination of any economist or psychologist.

I have now one more piece of understanding about what is it like to be me, about how to tame and steer my future behavior, and specially one more thing to tell people in awkward silence moments to break the ice and face the absurdity of reality. 

For obvious reasons, if you have long hair, I'd like to make an even stronger case for you to try to work and do pomos at the LW study hall. It's not only yourself that you'll be helping!