Time start 13:35:06

For another exercise in speed writing, I wanted to share a few book reviews.

These are fairly well known, however there is a chance you haven't read all of them - in which case, this might be helpful.

Good and Real - Gary Drescher ★★★★★

This is one of my favourite books ever. Goes over a lot of philosophy, while showing a lot of clear thinking and meta-thinking. Number one replacement for Eliezer's meta-philosophy, if it had not existed. The writing style and language is somewhat obscure, but this book is too brilliant to be spoiled by that. The biggest takeaway is the analysis of ethics of non-causal consequences of our choices, which is something that actually has changed how I act in my life, and I have not seen any similar argument in other sources that would do the same. This book changed my intuitions so much that I now pay $100 in counterfactual mugging without second thought.

59 Seconds - Richard Wiseman ★★★

A collection of various tips and tricks, directly based on studies. The strength of the book is that it gives easy but detailed descriptions of lots of studies, and that makes it very fun to read. Can be read just to check out the various psychology results in an entertaining format. The quality of the advice is disputable, and it is mostly the kind of advice that only applies to small things and does not change much in what you do even if you somehow manage to use it. But I still liked this book, and it managed to avoid saying anything very stupid while saying a lot of things. It counts for something.

What You Can Change and What You Can't - Martin Seligman ★★★

It is a heartwarming to see that the author puts his best effort towards figuring out what psychology treatments work, and which don't, as well as builiding more general models of how people work that can predict what treatments have a chance in the first place. Not all of the content is necessarily your best guess, after updating on new results (the book is quite old). However if you are starting out, this book will serve excellently as your prior, on which you can update after checking out the new results. And also in some cases, it is amazing that the author was right about them 20 years ago, and mainstream psychology is STILL not caught up (like the whole bullshit "go back to your childhood to fix your problems" approach, which is in wide use today and not bothered at all by such things as "checking facts").

Thinking, Fast and Slow - Daniel Kahneman ★★★★★

A classic, and I want to mention it just in case. It is too valuable not to read. Period. It turns out some of the studies the author used for his claims have been later found not to replicate. However the details of those results is not (at least for me) a selling point of this book. The biggest thing is the author's mental toolbox for self-analysis and analysis of biases, as well concepts that he created to describe the mechanisms of intuitive judgement. Learn to think like the author, and you are 10 years ahead in your study of rationality.

Crucial Conversations - Al Switzler, Joseph Grenny, Kerry Patterson, Ron McMillan ★★★★

I have almost dropped this book. When I saw the style, it reminded me so much of the crappy self-help books without actual content. But fortunately I have read on a litte more, and it turns out that even while the style is the same in the whole book and it has litte content for the amount of text you read, it is still an excellent book. How is that possible? Simple: it only tells you a few things, but the things it tells you are actually important and they work and they are amazing when you put them into practice. Also on the concept and analysis side, there is precious little but who cares as long as there are some things that are "keepers". The authors spend most of the book hammering the same point over and over, which is "conversation safety". And it is still a good book: if you get this one simple point than you have learned more than you might from reading 10 other books.

How to Fail at Almost Everything and Still Win Big - Scott Adams ★★★

I don't agree with much of the stuff that is in this book, but that's not the point here. The author says what he thinks, and also he himself encourages you to pass it through your own filters. Around one third of the book, I thought it was obviously true; another one third, I had strong evidence that told me the author made a mistake or got confused about something; and the remaining one third gave me new ideas, or points of view that I could use to produce more ideas for my own use. This felt kind of like having a conversation with any intelligent person you might know, who has different ideas from you. It was a healthy ratio of agreement and disagreement, such that leads to progress for both people. Except of course in this case the author did not benefit, but I did.

Time end: 14:01:54

Total time to write this post: 26 minutes 48 seconds

Average writing speed: 31.2 words/minute, 169 characters/minute

The same data calculated for my previous speed-writing post: 30.1 words/minute, 167 characters/minute

From the Google Research blog:

We believe that AI technologies are likely to be overwhelmingly useful and beneficial for humanity. But part of being a responsible steward of any new technology is thinking through potential challenges and how best to address any associated risks. So today we’re publishing a technical paper, Concrete Problems in AI Safety, a collaboration among scientists at Google, OpenAI, Stanford and Berkeley.

While possible AI safety risks have received a lot of public attention, most previous discussion has been very hypothetical and speculative. We believe it’s essential to ground concerns in real machine learning research, and to start developing practical approaches for engineering AI systems that operate safely and reliably.

We’ve outlined five problems we think will be very important as we apply AI in more general circumstances. These are all forward thinking, long-term research questions -- minor issues today, but important to address for future systems:

• Avoiding Negative Side Effects: How can we ensure that an AI system will not disturb its environment in negative ways while pursuing its goals, e.g. a cleaning robot knocking over a vase because it can clean faster by doing so?
• Avoiding Reward Hacking: How can we avoid gaming of the reward function? For example, we don’t want this cleaning robot simply covering over messes with materials it can’t see through.
• Scalable Oversight: How can we efficiently ensure that a given AI system respects aspects of the objective that are too expensive to be frequently evaluated during training? For example, if an AI system gets human feedback as it performs a task, it needs to use that feedback efficiently because asking too often would be annoying.
• Safe Exploration: How do we ensure that an AI system doesn’t make exploratory moves with very negative repercussions? For example, maybe a cleaning robot should experiment with mopping strategies, but clearly it shouldn’t try putting a wet mop in an electrical outlet.
• Robustness to Distributional Shift: How do we ensure that an AI system recognizes, and behaves robustly, when it’s in an environment very different from its training environment? For example, heuristics learned for a factory workfloor may not be safe enough for an office.

We go into more technical detail in the paper. The machine learning research community has already thought quite a bit about most of these problems and many related issues, but we think there’s a lot more work to be done.

We believe in rigorous, open, cross-institution work on how to build machine learning systems that work as intended. We’re eager to continue our collaborations with other research groups to make positive progress on AI.

These are my notes and observations after attending the Safety in Artificial Intelligence (SafArtInt) conference, which was co-hosted by the White House Office of Science and Technology Policy and Carnegie Mellon University on June 27 and 28. This isn't an organized summary of the content of the conference; rather, it's a selection of points which are relevant to the control problem. As a result, it suffers from selection bias: it looks like superintelligence and control-problem-relevant issues were discussed frequently, when in reality those issues were discussed less and I didn't write much about the more mundane parts.

SafArtInt has been the third out of a planned series of four conferences. The purpose of the conference series was twofold: the OSTP wanted to get other parts of the government moving on AI issues, and they also wanted to inform public opinion.

The other three conferences are about near term legal, social, and economic issues of AI. SafArtInt was about near term safety and reliability in AI systems. It was effectively the brainchild of Dr. Ed Felten, the deputy U.S. chief technology officer for the White House, who came up with the idea for it last year. CMU is a top computer science university and many of their own researchers attended, as well as some students. There were also researchers from other universities, some people from private sector AI including both Silicon Valley and government contracting, government researchers and policymakers from groups such as DARPA and NASA, a few people from the military/DoD, and a few control problem researchers. As far as I could tell, everyone except a few university researchers were from the U.S., although I did not meet many people. There were about 70-100 people watching the presentations at any given time, and I had conversations with about twelve of the people who were not affiliated with existential risk organizations, as well as of course all of those who were affiliated. The conference was split with a few presentations on the 27th and the majority of presentations on the 28th. Not everyone was there for both days.

Felten believes that neither "robot apocalypses" nor "mass unemployment" are likely. It soon became apparent that the majority of others present at the conference felt the same way with regard to superintelligence. The general intention among researchers and policymakers at the conference could be summarized as follows: we need to make sure that the AI systems we develop in the near future will not be responsible for any accidents, because if accidents do happen then they will spark public fears about AI, which would lead to a dearth of funding for AI research and an inability to realize the corresponding social and economic benefits. Of course, that doesn't change the fact that they strongly care about safety in its own right and have significant pragmatic needs for robust and reliable AI systems.

Most of the talks were about verification and reliability in modern day AI systems. So they were concerned with AI systems that would give poor results or be unreliable in the narrow domains where they are being applied in the near future. They mostly focused on "safety-critical" systems, where failure of an AI program would result in serious negative consequences: automated vehicles were a common topic of interest, as well as the use of AI in healthcare systems. A recurring theme was that we have to be more rigorous in demonstrating safety and do actual hazard analyses on AI systems, and another was that we need the AI safety field to succeed in ways that the cybersecurity field has failed. Another general belief was that long term AI safety, such as concerns about the ability of humans to control AIs, was not a serious issue.

On average, the presentations were moderately technical. They were mostly focused on machine learning systems, although there was significant discussion of cybersecurity techniques.

The first talk was given by Eric Horvitz of Microsoft. He discussed some approaches for pushing into new directions in AI safety. Instead of merely trying to reduce the errors spotted according to one model, we should look out for "unknown unknowns" by stacking models and looking at problems which appear on any of them, a theme which would be presented by other researchers as well in later presentations. He discussed optimization under uncertain parameters, sensitivity analysis to uncertain parameters, and 'wireheading' or short-circuiting of reinforcement learning systems (which he believes can be guarded against by using 'reflective analysis'). Finally, he brought up the concerns about superintelligence, which sparked amused reactions in the audience. He said that scientists should address concerns about superintelligence, which he aptly described as the 'elephant in the room', noting that it was the reason that some people were at the conference. He said that scientists will have to engage with public concerns, while also noting that there were experts who were worried about superintelligence and that there would have to be engagement with the experts' concerns. He did not comment on whether he believed that these concerns were reasonable or not.

An issue which came up in the Q&A afterwards was that we need to deal with mis-structured utility functions in AI, because it is often the case that the specific tradeoffs and utilities which humans claim to value often lead to results which the humans don't like. So we need to have structural uncertainty about our utility models. The difficulty of finding good objective functions for AIs would eventually be discussed in many other presentations as well.

The next talk was given by Andrew Moore of Carnegie Mellon University, who claimed that his talk represented the consensus of computer scientists at the school. He claimed that the stakes of AI safety were very high - namely, that AI has the capability to save many people's lives in the near future, but if there are any accidents involving AI then public fears could lead to freezes in AI research and development. He highlighted the public's irrational tendencies wherein a single accident could cause people to overlook and ignore hundreds of invisible lives saved. He specifically mentioned a 12-24 month timeframe for these issues.

Moore said that verification of AI system safety will be difficult due to the combinatorial explosion of AI behaviors. He talked about meta-machine-learning as a solution to this, something which is being investigated under the direction of Lawrence Schuette at the Office of Naval Research. Moore also said that military AI systems require high verification standards and that development timelines for these systems are long. He talked about two different approaches to AI safety, stochastic testing and theorem proving - the process of doing the latter often leads to the discovery of unsafe edge cases.

He also discussed AI ethics, giving an example 'trolley problem' where AI cars would have to choose whether to hit a deer in order to provide a slightly higher probability of survival for the human driver. He said that we would need hash-defined constants to tell vehicle AIs how many deer a human is worth. He also said that we would need to find compromises in death-pleasantry tradeoffs, for instance where the safety of self-driving cars depends on the speed and routes on which they are driven. He compared the issue to civil engineering where engineers have to operate with an assumption about how much money they would spend to save a human life.

He concluded by saying that we need policymakers, company executives, scientists, and startups to all be involved in AI safety. He said that the research community stands to gain or lose together, and that there is a shared responsibility among researchers and developers to avoid triggering another AI winter through unsafe AI designs.

The next presentation was by Richard Mallah of the Future of Life Institute, who was there to represent "Medium Term AI Safety". He pointed out the explicit/implicit distinction between different modeling techniques in AI systems, as well as the explicit/implicit distinction between different AI actuation techniques. He talked about the difficulty of value specification and the concept of instrumental subgoals as an important issue in the case of complex AIs which are beyond human understanding. He said that even a slight misalignment of AI values with regard to human values along one parameter could lead to a strongly negative outcome, because machine learning parameters don't strictly correspond to the things that humans care about.

Mallah stated that open-world discovery leads to self-discovery, which can lead to reward hacking or a loss of control. He underscored the importance of causal accounting, which is distinguishing causation from correlation in AI systems. He said that we should extend machine learning verification to self-modification. Finally, he talked about introducing non-self-centered ontology to AI systems and bounding their behavior.

The audience was generally quiet and respectful during Richard's talk. I sensed that at least a few of them labelled him as part of the 'superintelligence out-group' and dismissed him accordingly, but I did not learn what most people's thoughts or reactions were. In the next panel featuring three speakers, he wasn't the recipient of any questions regarding his presentation or ideas.

Tom Mitchell from CMU gave the next talk. He talked about both making AI systems safer, and using AI to make other systems safer. He said that risks to humanity from other kinds of issues besides AI were the "big deals of 2016" and that we should make sure that the potential of AIs to solve these problems is realized. He wanted to focus on the detection and remediation of all failures in AI systems. He said that it is a novel issue that learning systems defy standard pre-testing ("as Richard mentioned") and also brought up the purposeful use of AI for dangerous things.

Some interesting points were raised in the panel. Andrew did not have a direct response to the implications of AI ethics being determined by the predominantly white people of the US/UK where most AIs are being developed. He said that ethics in AIs will have to be decided by society, regulators, manufacturers, and human rights organizations in conjunction. He also said that our cost functions for AIs will have to get more and more complicated as AIs get better, and he said that he wants to separate unintended failures from superintelligence type scenarios. On trolley problems in self driving cars and similar issues, he said "it's got to be complicated and messy."

Dario Amodei of Google Deepbrain, who co-authored the paper on concrete problems in AI safety, gave the next talk. He said that the public focus is too much on AGI/ASI and wants more focus on concrete/empirical approaches. He discussed the same problems that pose issues in advanced general AI, including flawed objective functions and reward hacking. He said that he sees long term concerns about AGI/ASI as "extreme versions of accident risk" and that he thinks it's too early to work directly on them, but he believes that if you want to deal with them then the best way to do it is to start with safety in current systems. Mostly he summarized the Google paper in his talk.

In her presentation, Claire Le Goues of CMU said "before we talk about Skynet we should focus on problems that we already have." She mostly talked about analogies between software bugs and AI safety, the similarities and differences between the two and what we can learn from software debugging to help with AI safety.

Robert Rahmer of IARPA discussed CAUSE, a cyberintelligence forecasting program which promises to help predict cyber attacks. It is a program which is still being put together.

In the panel of the above three, autonomous weapons were discussed, but no clear policy stances were presented.

John Launchbury gave a talk on DARPA research and the big picture of AI development. He pointed out that DARPA work leads to commercial applications and that progress in AI comes from sustained government investment. He classified AI capabilities into "describing," "predicting," and "explaining" in order of increasing difficulty, and he pointed out that old fashioned "describing" still plays a large role in AI verification. He said that "explaining" AIs would need transparent decisionmaking and probabilistic programming (the latter would also be discussed by others at the conference).

The next talk came from Jason Gaverick Matheny, the director of IARPA. Matheny talked about four requirements in current and future AI systems: verification, validation, security, and control. He wanted "auditability" in AI systems as a weaker form of explainability. He talked about the importance of "corner cases" for national intelligence purposes, the low probability, high stakes situations where we have limited data - these are situations where we have significant need for analysis but where the traditional machine learning approach doesn't work because of its overwhelming focus on data. Another aspect of national defense is that it has a slower decision tempo, longer timelines, and longer-viewing optics about future events.

He said that assessing local progress in machine learning development would be important for global security and that we therefore need benchmarks to measure progress in AIs. He ended with a concrete invitation for research proposals from anyone (educated or not), for both large scale research and for smaller studies ("seedlings") that could take us "from disbelief to doubt".

The difference in timescales between different groups was something I noticed later on, after hearing someone from the DoD describe their agency as having a longer timeframe than the Homeland Security Agency, and someone from the White House describe their work as being crisis reactionary.

The next presentation was from Andrew Grotto, senior director of cybersecurity policy at the National Security Council. He drew a close parallel from the issue of genetically modified crops in Europe in the 1990's to modern day artificial intelligence. He pointed out that Europe utterly failed to achieve widespread cultivation of GMO crops as a result of public backlash. He said that the widespread economic and health benefits of GMO crops were ignored by the public, who instead focused on a few health incidents which undermined trust in the government and crop producers. He had three key points: that risk frameworks matter, that you should never assume that the benefits of new technology will be widely perceived by the public, and that we're all in this together with regard to funding, research progress and public perception.

In the Q&A between Launchbury, Matheny, and Grotto after Grotto's presentation, it was mentioned that the economic interests of farmers worried about displacement also played a role in populist rejection of GMOs, and that a similar dynamic could play out with regard to automation causing structural unemployment. Grotto was also asked what to do about bad publicity which seeks to sink progress in order to avoid risks. He said that meetings like SafArtInt and open public dialogue were good.

One person asked what Launchbury wanted to do about AI arms races with multiple countries trying to "get there" and whether he thinks we should go "slow and secure" or "fast and risky" in AI development, a question which provoked laughter in the audience. He said we should go "fast and secure" and wasn't concerned. He said that secure designs for the Internet once existed, but the one which took off was the one which was open and flexible.

Another person asked how we could avoid discounting outliers in our models, referencing Matheny's point that we need to include corner cases. Matheny affirmed that data quality is a limiting factor to many of our machine learning capabilities. At IARPA, we generally try to include outliers until they are sure that they are erroneous, said Matheny.

Another presentation came from Tom Dietterich, president of the Association for the Advancement of Artificial Intelligence. He said that we have not focused enough on safety, reliability and robustness in AI and that this must change. Much like Eric Horvitz, he drew a distinction between robustness against errors within the scope of a model and robustness against unmodeled phenomena. On the latter issue, he talked about solutions such as expanding the scope of models, employing multiple parallel models, and doing creative searches for flaws - the latter doesn't enable verification that a system is safe, but it nevertheless helps discover many potential problems. He talked about knowledge-level redundancy as a method of avoiding misspecification - for instance, systems could identify objects by an "ownership facet" as well as by a "goal facet" to produce a combined concept with less likelihood of overlooking key features. He said that this would require wider experiences and more data.

There were many other speakers who brought up a similar set of issues: the user of cybersecurity techniques to verify machine learning systems, the failures of cybersecurity as a field, opportunities for probabilistic programming, and the need for better success in AI verification. Inverse reinforcement learning was extensively discussed as a way of assigning values. Jeanette Wing of Microsoft talked about the need for AIs to reason about the continuous and the discrete in parallel, as well as the need for them to reason about uncertainty (with potential meta levels all the way up). One point which was made by Sarah Loos of Google was that proving the safety of an AI system can be computationally very expensive, especially given the combinatorial explosion of AI behaviors.

In one of the panels, the idea of government actions to ensure AI safety was discussed. No one was willing to say that the government should regulate AI designs. Instead they stated that the government should be involved in softer ways, such as guiding and working with AI developers, and setting standards for certification.

Pictures: https://imgur.com/a/49eb7

In between these presentations I had time to speak to individuals and listen in on various conversations. A high ranking person from the Department of Defense stated that the real benefit of autonomous systems would be in terms of logistical systems rather than weaponized applications. A government AI contractor drew the connection between Mallah's presentation and the recent press revolving around superintelligence, and said he was glad that the government wasn't worried about it.

I talked to some insiders about the status of organizations such as MIRI, and found that the current crop of AI safety groups could use additional donations to become more established and expand their programs. There may be some issues with the organizations being sidelined; after all, the Google Deepbrain paper was essentially similar to a lot of work by MIRI, just expressed in somewhat different language, and was more widely received in mainstream AI circles.

In terms of careers, I found that there is significant opportunity for a wide range of people to contribute to improving government policy on this issue. Working at a group such as the Office of Science and Technology Policy does not necessarily require advanced technical education, as you can just as easily enter straight out of a liberal arts undergraduate program and build a successful career as long as you are technically literate. (At the same time, the level of skepticism about long term AI safety at the conference hinted to me that the signalling value of a PhD in computer science would be significant.) In addition, there are large government budgets in the seven or eight figure range available for qualifying research projects. I've come to believe that it would not be difficult to find or create AI research programs that are relevant to long term AI safety while also being practical and likely to be funded by skeptical policymakers and officials.

I also realized that there is a significant need for people who are interested in long term AI safety to have basic social and business skills. Since there is so much need for persuasion and compromise in government policy, there is a lot of value to be had in being communicative, engaging, approachable, appealing, socially savvy, and well-dressed. This is not to say that everyone involved in long term AI safety is missing those skills, of course.

I was surprised by the refusal of almost everyone at the conference to take long term AI safety seriously, as I had previously held the belief that it was more of a mixed debate given the existence of expert computer scientists who were involved in the issue. I sensed that the recent wave of popular press and public interest in dangerous AI has made researchers and policymakers substantially less likely to take the issue seriously. None of them seemed to be familiar with actual arguments or research on the control problem, so their opinions didn't significantly change my outlook on the technical issues. I strongly suspect that the majority of them had their first or possibly only exposure to the idea of the control problem after seeing badly written op-eds and news editorials featuring comments from the likes of Elon Musk and Stephen Hawking, which would naturally make them strongly predisposed to not take the issue seriously. In the run-up to the conference, websites and press releases didn't say anything about whether this conference would be about long or short term AI safety, and they didn't make any reference to the idea of superintelligence.

I sympathize with the concerns and strategy given by people such as Andrew Moore and Andrew Grotto, which make perfect sense if (and only if) you assume that worries about long term AI safety are completely unfounded. For the community that is interested in long term AI safety, I would recommend that we avoid competitive dynamics by (a) demonstrating that we are equally strong opponents of bad press, inaccurate news, and irrational public opinion which promotes generic uninformed fears over AI, (b) explaining that we are not interested in removing funding for AI research (even if you think that slowing down AI development is a good thing, restricting funding yields only limited benefits in terms of changing overall timelines, whereas those who are not concerned about long term AI safety would see a restriction of funding as a direct threat to their interests and projects, so it makes sense to cooperate here in exchange for other concessions), and (c) showing that we are scientifically literate and focused on the technical concerns. I do not believe that there is necessarily a need for the two "sides" on this to be competing against each other, so it was disappointing to see an implication of opposition at the conference.

Anyway, Ed Felten announced a request for information from the general public, seeking popular and scientific input on the government's policies and attitudes towards AI: https://www.whitehouse.gov/webform/rfi-preparing-future-artificial-intelligence

Overall, I learned quite a bit and benefited from the experience, and I hope the insight I've gained can be used to improve the attitudes and approaches of the long term AI safety community.

Similar to the Crazy Ideas Thread and Diaspora Roundup Thread, I thought I'd try making a General-Purpose Questions Thread.

The purpose is to provide a forum for asking questions to the community (appealing to the wisdom of this particular crowd) in things that don't really merit their own thread.

There's a semi-famous paper by Dunn, Gilbert and Wilson: "If money doesn't make you happy, then you probably aren't spending it right". (Proper reference: Dunn, E.W., Gilbert, D.T., and Wilson, T.D., If money doesn't make you happy, then you probably aren't spending it right, Journal of Consumer Psychology, vol 21, issue 2, April 2011, pp. 115–125.) It's been referenced a few times on LW but curiously never written up properly here. The purpose of this post is to remedy that.

There is an earlier LW post called "Be Happier" which among other things references this paper and quotes some things it says, but that post is monstrously long and covers a lot more ground (hence, less details on the material in this paper).

Dunn, Gilbert and Wilson (hereafter "DGW") offer eight principles to follow. Here they are.

1. Buy experiences instead of things.

Many studies have asked people to reflect on past "material" and/or "experiential" purchases and have consistently found that they report greater happiness from (and are made happier by recalling) the latter than the former.

Why? DGW propose 5 reasons. First, deliberately sought-out experiences encourage us to focus on the here and now (something shown to increase happiness substantially); second, when things don't change we adapt to them rapidly, and "material" purchases like cars and tables tend to be pretty stable (whereas ongoing experiences are more varied); third, it turns out that people spend more time anticipating experiences before they happen and recalling them afterwards than they do for material purchases. Fourth, experiences are less directly comparable to alternatives than material things, and therefore less subject to post-purchase regret. Fifth, experiences are often shared, and other people are a great source of happiness.

2. Help others instead of yourself.

Prosocial spending correlates better to happiness than personal spending. If you give random people money and either tell them to spend it on themselves or to spend it on someone else, the latter makes them happier. Reflecting on past spending-on-others makes people happier than reflecting on past spending-on-self. (I am a little skeptical about that one: the right point of comparison would be not the past spending but the past enjoyment of whatever you spent the money on.)

Why? DGW propose two reasons. First, prosocial spending is good for relationships and relationships are good for happiness. Second, when you spend on someone else you get to feel like a good person.

Most people have wrong intuitions about this: they expect spending on themselves to make them happier. Most people are wrong.

3. Buy many small pleasures instead of few big ones.

As we saw above under #1, we quickly adapt to changes. Therefore, a larger number of varied small pleasures may be a better buy than a single big one. There is some evidence for this (though to my mind it seems to bear less directly on DGW's principle than in the other cases we've considered so far). If you correlate people's happiness with their positive experiences, the correlation with how frequent those experiences are is stronger than the correlation with how intense they are. The optimal (for happiness) number of sexual partners to have over a year is one, perhaps because that gets you more sex even if individual instances are less exciting. (I find this less than convincing; individual instances might be better because partners learn what works well for them.)

The other reason DGW suggest why more smaller things should be better is diminishing marginal utility: half a cookie is more than half as good as a whole cookie. (This is, I think, partly because of adaptation, but that isn't the whole story.)

DGW suggest that this is one reason why the relationship between wealth and happiness isn't stronger: "wealth promises access to peak experiences, which in turn undermine the ability to savor small pleasures".

4. Buy less insurance.

We adapt to bad things as well as good, which means that bad things are less bad than we are liable to expect. Our overestimation of the impact of adverse occurrences is one reason why we buy insurance, which notoriously is always negative-expectation in monetary terms.

DGW cite various studies showing that people expect to be made markedly unhappier by losses than they actually are if the losses occur, and that people expect to regret bad outcomes more than they actually do (we overestimate how much we will blame ourselves, because we underestimate how good we are at blaming anything and anyone else for our misfortunes).

5. Pay now and consume later.

The opposite of the bargain proposed by credit cards! Besides the purely financial problems that arise from overspending (which are large and widespread), DGW suggest that "consume now, pay later" is bad for our happiness because it eliminates anticipation. We may derive a lot of pleasure even from anticipating something that we don't enjoy very much when it happens. "People who devote time to anticipating enjoyable experiences report being happier in general."

You might think that moving an experience later would simply mean more anticipation (good) but less reminiscence (bad), but it turns out that anticipation generally brings more happiness. (And, for unpleasant events, more pain.)

DGW suggest two other benefits of delaying consumption. First, we may make better choices (meaning, in this case, ones yielding more happiness overall, even if less in the very short term) when we make them a little way ahead. Second, the delay may increase uncertainty, which may keep attention focused on the thing we're buying, which may reduce adaptation. (This seems a little convoluted to me; DGW cite some research backing it up but I'm not sure it backs up the "by reducing adaptation" part of it.)

6. Think about what you're not thinking about.

That is: when choosing what to spend on, take some time to consider less obvious aspects that you'd otherwise be tempted to neglect. "The bigger home may seem like a better deal, but if the fixer-upper requires trading Saturday afternoons with friends for Saturday afternoons with plumbers, it may not be such a good deal after all." And: "consumers who expect a single purchase to have a lasting impact on their happiness might make more realistic predictions if they simply thought about a typical day in their life." (Rather than considering only the small bits of that day that will be impacted by their purchase.)

7. Beware of comparison shopping.

Comparison shopping, say DGW, focuses attention on the features that most clearly distinguish candidate purchases from one another, whereas other more-common features may actually have much more impact on happiness. It may also focus attention on more-concrete differences; for instance, if you ask people whether they would more enjoy a small heart-shaped chocolate or a large cockroach-shaped one, they generally prefer the former, but if you ask them to choose one of the two they tend to focus on the size and choose the latter.

DGW also point out that the context during comparison-shopping tends to be different from that during actual consumption, which can skew our evaluations.

8. Follow the herd instead of your head.

DGW cite research supporting de la Rochefoucauld's advice: "Before we set our hearts too much upon anything, let us first examine how happy those are who already possess it." Others' actual experiences of a thing are likely to be better predictors of our enjoyment than our theoretical estimates: we may know ourselves better, but they know the thing better.

They also suggest (and I don't think this really fits their heading) looking to others for advice on how we would enjoy something we are considering buying. The example they give is of research in which subjects were shown some foods and asked to estimate how much they would enjoy them, after which they ate them and evaluated their actual enjoyment. The wrinkle is that they were also observed, at the moment of being shown the foods, by other observers, who rated their immediate facial reactions -- which turned out to be better predictors of their enjoyment than the subjects' own assessments. So "other people may provide a useful source of information about the products that will bring us joy because they can see the nonverbal reactions that may escape our own notice".

Outline: I will discuss my background and how I prepared for the workshop, and then how I would have prepared differently if I could go back and have the chance to do it again; I will then discuss my experience at the CFAR workshop, and what I would have done differently if I had the chance to do it again; I will then discuss what my take-aways were from the workshop, and what I am doing to integrate CFAR strategies into my life; finally, I will give my assessment of its benefits and what other folks might expect to get who attend the workshop.

Acknowledgments: Thanks to fellow CFAR alumni and CFAR staff for feedback on earlier versions of this post

Introduction

Many aspiring rationalists have heard about the Center for Applied Rationality, an organization devoted to teaching applied rationality skills to help people improve their thinking, feeling, and behavior patterns. This nonprofit does so primarily through its intense workshops, and is funded by donations and revenue from its workshop. It fulfills its social mission through conducting rationality research and through giving discounted or free workshops to those people its staff judge as likely to help make the world a better place, mainly those associated with various Effective Altruist cause areas, especially existential risk.

To be fully transparent: even before attending the workshop, I already had a strong belief that CFAR is a great organization and have been a monthly donor to CFAR for years. So keep that in mind as you read my description of my experience (you can become a donor here).

Preparation

First, some background about myself, so you know where I’m coming from in attending the workshop. I’m a professor specializing in the intersection of history, psychology, behavioral economics, sociology, and cognitive neuroscience. I discovered the rationality movement several years ago through a combination of my research and attending a LessWrong meetup in Columbus, OH, and so come from a background of both academic and LW-style rationality. Since discovering the movement, I have become an activist in the movement as the President of Intentional Insights, a nonprofit devoted to popularizing rationality and effective altruism (see here for our EA work). So I came to the workshop with some training and knowledge of rationality, including some CFAR techniques.

To help myself prepare for the workshop, I reviewed existing posts about CFAR materials, with an eye toward being careful not to assume that the actual techniques match their actual descriptions in the posts.

I also delayed a number of tasks for after the workshop, tying up loose ends. In retrospect, I wish I did not leave myself some ongoing tasks to do during the workshop. As part of my leadership of InIn, I coordinate about 50ish volunteers, and I wish I had placed those responsibilities on someone else during the workshop.

Before the workshop, I worked intensely on finishing up some projects. In retrospect, it would have been better to get some rest and come to the workshop as fresh as possible.

There were some communication snafus with logistics details before the workshop. It all worked out in the end, but I would have told myself in retrospect to get the logistics hammered out in advance to not experience anxiety before the workshop about how to get there.

Experience

The classes were well put together, had interesting examples, and provided useful techniques. FYI, my experience in the workshop was that reading these techniques in advance was not harmful, but that the techniques in the CFAR classes were quite a bit better than the existing posts about them, so don’t assume you can get the same benefits from reading posts as attending the workshop. So while I was aware of the techniques, the ones in the classes definitely had optimized versions of them - maybe because of the “broken telephone” effect or maybe because CFAR optimized them from previous workshops, not sure. I was glad to learn that CFAR considers the workshop they gave us in May as satisfactory enough to scale up their workshops, while still improving their content over time.

Just as useful as the classes were the conversations held in between and after the official classes ended. Talking about them with fellow aspiring rationalists and seeing how they were thinking about applying these to their lives was helpful for sparking ideas about how to apply them to my life. The latter half of the CFAR workshop was especially great, as it focused on pairing off people and helping others figure out how to apply CFAR techniques to themselves and how to address various problems in their lives. It was especially helpful to have conversations with CFAR staff and trained volunteers, of whom there were plenty - probably about 20 volunteers/staff for the 50ish workshop attendees.

Another super-helpful aspect of the conversations was networking and community building. Now, this may have been more useful to some participants than others, so YMMV. As an activist in the moment, I talked to many folks in the CFAR workshop about promoting EA and rationality to a broad audience. I was happy to introduce some people to EA, with my most positive conversation there being to encourage someone to switch his efforts regarding x-risk from addressing nuclear disarmament to AI safety research as a means of addressing long/medium-term risk, and promoting rationality as a means of addressing short/medium-term risk. Others who were already familiar with EA were interested in ways of promoting it broadly, while some aspiring rationalists expressed enthusiasm over becoming rationality communicators.

Looking back at my experience, I wish I was more aware of the benefits of these conversations. I went to sleep early the first couple of nights, and I would have taken supplements to enable myself to stay awake and have conversations instead.

Take-Aways and Integration

The aspects of the workshop that I think will help me most were what CFAR staff called “5-second” strategies - brief tactics and techniques that could be executed in 5 seconds or less and address various problems. The stuff that we learned at the workshops that I was already familiar with required some time to learn and practice, such as Trigger Action Plans, Goal Factoring, Murphyjitsu, Pre-Hindsight, often with pen and paper as part of the work. However, with sufficient practice, one can develop brief techniques that mimic various aspects of the more thorough techniques, and apply them quickly to in-the-moment decision-making.

Now, this doesn’t mean that the longer techniques are not helpful. They are very important, but they are things I was already generally familiar with, and already practice. The 5-second versions were more of a revelation for me, and I anticipate will be more helpful for me as I did not know about them previously.

Now, CFAR does a very nice job of helping people integrate the techniques into daily life, as this is a common failure mode of CFAR attendees, with them going home and not practicing the techniques. So they have 6 Google Hangouts with CFAR staff and all attendees who want to participate, 4 one-on-one sessions with CFAR trained volunteers or staff, and they also pair you with one attendee for post-workshop conversations. I plan to take advantage of all these, although my pairing did not work out.

For integration of CFAR techniques into my life, I found the CFAR strategy of “Overlearning” especially helpful. Overlearning refers to trying to apply a single technique intensely for a while to all aspect of one’s activities, so that it gets internalized thoroughly. I will first focus on overlearning Trigger Action Plans, following the advice of CFAR.

I also plan to teach CFAR techniques in my local rationality dojo, as teaching is a great way to learn, naturally.

Finally, I plan to integrate some CFAR techniques into Intentional Insights content, at least the more simple techniques that are a good fit for the broad audience with which InIn is communicating.

Benefits

I have a strong probabilistic belief that having attended the workshop will improve my capacity to be a person who achieves my goals for doing good in the world. I anticipate I will be able to figure out better whether the projects I am taking on are the best uses of my time and energy. I will be more capable of avoiding procrastination and other forms of akrasia. I believe I will be more capable of making better plans, and acting on them well. I will also be more in touch with my emotions and intuitions, and be able to trust them more, as I will have more alignment among different components of my mind.

Another benefit is meeting the many other people at CFAR who have similar mindsets. Here in Columbus, we have a flourishing rationality community, but it’s still relatively small. Getting to know 70ish people, attendees and staff/volunteers, passionate about rationality was a blast. It was especially great to see people who were involved in creating new rationality strategies, something that I am engaged in myself in addition to popularizing rationality - it’s really heartening to envision how the rationality movement is growing.

These benefits should resonate strongly with those who are aspiring rationalists, but they are really important for EA participants as well. I think one of the best things that EA movement members can do is studying rationality, and it’s something we promote to the EA movement as part of InIn’s work. What we offer is articles and videos, but coming to a CFAR workshop is a much more intense and cohesive way of getting these benefits. Imagine all the good you can do for the world if you are better at planning, organizing, and enacting EA-related tasks. Rationality is what has helped me and other InIn participants make the major impact that we have been able to make, and there are a number of EA movement members who have rationality training and who reported similar benefits. Remember, as an EA participant, you can likely get a scholarship with a partial or full coverage of the regular $3900 price of the workshop, as I did myself when attending it, and you are highly likely to be able to save more lives as a result of attending the workshop over time, even if you have to pay some costs upfront.

Hope these thoughts prove helpful to you all, and please contact me at gleb@intentionalinsights.org if you want to chat with me about my experience.

Please post other Making Less Wrong Great Again memes in the comments

This is a follow-up to last year's report. Here, I will talk about my successes and failures using Spaced Repetition Software (SRS) in the classroom for a second year. The year's not over yet, but I have reasons for reporting early that should become clear in a subsequent post. A third post will then follow, and together these will constitute a small sequence exploring classroom SRS and the adjacent ideas that bubble up when I think deeply about teaching.

Summary

I experienced net negative progress this year in my efforts to improve classroom instruction via spaced repetition software. While this is mostly attributable to shifts in my personal priorities, I have also identified a number of additional failure modes for classroom SRS, as well as additional shortcomings of Anki for this use case. My experiences also showcase some fundamental challenges to teaching-in-general that SRS depressingly spotlights without being any less susceptible to. Regardless, I am more bullish than ever about the potential for classroom SRS, and will lay out a detailed vision for what it can be in the next post.

One or two research groups have published work on carbon dioxide and cognition. The state of the published literature is confusing.

Here is one paper on the topic. The authors investigate a proprietary cognitive benchmark, and experimentally manipulate carbon dioxide levels (without affecting other measures of air quality). They find implausibly large effects from increased carbon dioxide concentrations.

If the reported effects are real and the suggested interpretation is correct, I think it would be a big deal. To put this in perspective, carbon dioxide concentrations in my room vary between 500 and 1500 ppm depending on whether I open the windows. The experiment reports on cognitive effects for moving from 600 and 1000 ppm, and finds significant effects compared to interindividual differences.

I haven't spent much time looking into this (maybe 30 minutes, and another 30 minutes to write this post). I expect that if we spent some time looking into indoor CO2 we could have a much better sense of what was going on, by some combination of better literature review, discussion with experts, looking into the benchmark they used, and just generally thinking about it.

So, here's a proposal:

• If someone looks into this and writes a post that improves our collective understanding of the issue, I will be willing to buy part of an associated certificate of impact, at a price of around $100*N, where N is my own totally made up estimate of how many hours of my own time it would take to produce a similarly useful writeup. I'd buy up to 50% of the certificate at that price.
• Whether or not they want to sell me some of the certificate, on May 1 I'll give a $500 prize to the author of the best publicly-available analysis of the issue. If the best analysis draws heavily on someone else's work, I'll use my discretion: I may split the prize arbitrarily, and may give it to the earlier post even if it is not quite as excellent.

Some clarifications:

• The metric for quality is "how useful it is to Paul." I hope that's a useful proxy for how useful it is in general, but no guarantees. I am generally a pretty skeptical person. I would care a lot about even a modest but well-established effect on performance. 
• These don't need to be new analyses, either for the prize or the purchase.
• I reserve the right to resolve all ambiguities arbitrarily, and in the end to do whatever I feel like. But I promise I am generally a nice guy.
• I posted this 2 weeks ago on the EA forum and haven't had serious takers yet.