Science and Technology

I’ve recently noticed some possibly important confusion about machine learning (ML)/deep learning. I’m quite uncertain how much harm the confusion will cause.

On MIRI’s Intelligent Agent Foundations Forum:

If you don’t do cognitive reductions, you will put your confusion in boxes and hide the actual problem. … E.g. if neural networks are used to predict math, then the confusion about how to do logical uncertainty is placed in the black box of “what this neural net learns to do”

On SlateStarCodex:

Imagine a future inmate asking why he was denied parole, and the answer being “nobody knows and it’s impossible to find out even in principle” … (DeepMind employs a Go master to help explain AlphaGo’s decisions back to its own programmers, which is probably a metaphor for something)

A possibly related confusion, from a conversation that I observed recently: philosophers have tried to understand how concepts work for centuries, but have made little progress; therefore deep learning isn’t very close to human-level AGI.

I’m unsure whether any of the claims I’m criticizing reflect actually mistaken beliefs, or whether they’re just communicated carelessly. I’m confident that at least some people at MIRI are wise enough to avoid this confusion [1]. I’ve omitted some ensuing clarifications from my description of the deep learning conversation – maybe if I remembered those sufficiently well, I’d see that I was reacting to a straw man of that discussion. But it seems likely that some people were misled by at least the SlateStarCodex comment.

There’s an important truth that people refer to when they say that neural nets (and machine learning techniques in general) are opaque. But that truth gets seriously obscured when rephrased as “black box” or “impossible to find out even in principle”.
Continue Reading

Book review: The Rationality Quotient: Toward a Test of Rational Thinking, by Keith E. Stanovich, Richard F. West and Maggie E. Toplak.

This book describes an important approach to measuring individual rationality: an RQ test that loosely resembles an IQ test. But it pays inadequate attention to the most important problems with tests of rationality.

Coachability

My biggest concern about rationality testing is what happens when people anticipate the test and are motivated to maximize their scores (as is the case with IQ tests). Do they:

  • learn to score high by “cheating” (i.e. learn what answers the test wants, without learning to apply that knowledge outside of the test)?
  • learn to score high by becoming more rational?
  • not change their score much, because they’re already motivated to do as well as their aptitudes allow (as is mostly the case with IQ tests)?

Alas, the book treats these issues as an afterthought. Their test knowingly uses questions for which cheating would be straightforward, such as asking whether the test subject believes in science, and whether they prefer to get $85 now rather than $100 in three months. (If they could use real money, that would drastically reduce my concerns about cheating. I’m almost tempted to advocate doing that, but doing so would hinder widespread adoption of the test, even if using real money added enough value to pay for itself.)

Continue Reading

Two and a half years ago, Eliezer was (somewhat plausibly) complaining that virtually nobody outside of MIRI was working on AI-related existential risks.

This year (at EAGlobal) one of MIRI’s talks was a bit hard to distinguish from an AI safety talk given by someone with pretty mainstream AI affiliations.

What happened in that time to cause that shift?

A large change was catalyzed by the publication of Superintelligence. I’ve been mildly disappointed about how little it affected discussions among people who were already interested in the topic. But Superintelligence caused a large change in how many people are willing to express concern over AI risks. That’s presumably because Superintelligence looks sufficiently academic and neutral to make many people comfortable about citing it, whereas similar arguments by Eliezer/MIRI didn’t look sufficiently prestigious within academia.

A smaller part of the change was MIRI shifting its focus somewhat to be more in line with how mainstream machine learning (ML) researchers expect AI to reach human levels.

Also, OpenAI has been quietly shifting in a more MIRI-like direction (I’m very unclear on how big a change this is). (Paul Christiano seems to deserve some credit for both the MIRI and OpenAI shifts in strategies.)

Given those changes, it seems like MIRI ought to be able to attract more donations than before. Especially since it has demonstrated evidence of increasing competence, and also because HPMoR seemed to draw significantly more people into the community of people who are interested in MIRI.

MIRI has gotten one big grant from OpenPhilanthropy that it probably couldn’t have gotten when mainstream AI researchers were treating MIRI’s concerns as too far-fetched to be worth commenting on. But donations from MIRI’s usual sources have stagnated.

That pattern suggests that MIRI was previously benefiting from a polarization effect, where the perception of two distinct “tribes” (those who care about AI risks versus those who promote AI) energized people to care about “their tribe”.

Whereas now there’s no clear dividing line between MIRI and mainstream researchers. Also, there’s lots of money going into other organizations that plan to do something about AI safety. (Most of those haven’t yet articulated enough of a strategy to make me optimistic that that money is well spent. I still endorse the ideas I mentioned last year in How much Diversity of AGI-Risk Organizations is Optimal?. I’m unclear on how much diversity of approaches we’re getting from the recent proliferation of AI safety organizations.)

That kind of pattern of donations creates perverse incentives to charities to at least market themselves as fighting a powerful group of people, rather than (as the ideal charity should be) addressing a neglected problem. Even if that marketing doesn’t distort a charity’s operations, the charity will be tempted to use counterproductive alarmism. AI risk organizations have resisted those temptations (at least recently), but it seems risky to tempt them.

That’s part of why I recently made a modest donation to MIRI, in spite of the uncertainty over the value of their efforts (I had last donated to them in 2009).

[Caveat: this post involves abstract theorizing whose relevance to practical advice is unclear. ]

What we call willpower mostly derives from conflicts between parts of our minds, often over what discount rate to use.

An additional source of willpower-like conflicts comes from social desirability biases.

I model the mind as having many mental sub-agents, each focused on a fairly narrow goal. Different goals produce different preferences for caring about the distant future versus caring only about the near future.

The sub-agents typically are as smart and sophisticated as a three year old (probably with lots of variation). E.g. my hunger-minimizing sub-agent is willing to accept calorie restriction days with few complaints now that I have a reliable pattern of respecting the hunger-minimizing sub-agent the next day, but complained impatiently when calorie restriction days seemed abnormal.

We have beliefs about how safe we are from near-term dangers, often reflected in changes to the autonomic nervous system (causing relaxation or the fight or flight reflex). Those changes cause quick, crude shifts in something resembling a global discount rate. In addition, each sub-agent has some ability to demand that it’s goals be treated fairly.

We neglect sub-agents whose goals are most long-term when many sub-agents say their goals have been neglected, and/or when the autonomic nervous system says immediate problems deserve attention.

Our willpower is high when we feel safe and are satisfied with our progress at short-term goals.

Social status

The time-discounting effects are sometimes obscured by social signaling.

Writing a will hints at health problems, whereas doing something about global warming can signal wealth. We have sub-agents that steer us to signal health and wealth, but without doing so in a deliberate enough way that people see that we are signaling. That leads us to exaggerate how much of our failure to write a will is due to the time-discounting type of low willpower.

Video games convince parts of our minds that we’re gaining status (in a virtual society) and/or training to win status-related games in real life. That satisfies some sub-agents who care about status. (Video games deceive us about status effects, but that has limited relevance to this post.) Yet as with most play, we suppress awareness of the zero-sum competitions we’re aiming to win. So we get confused about whether we’re being short-sighted here, because we’re pursuing somewhat long-term benefits, probably deceiving ourselves somewhat about them, and pretending not to care about them.

Time asymmetry?

Why do we feel an asymmetry in effects of neglecting distant goals versus neglecting immediate goals?

The fairness to sub-agents metaphor suggests that neglecting the distant future ought to produce emotional reactions comparable to what happens when we neglect the near future.

Neglecting the distant future does produce some discomfort that somewhat resembles willpower problems. If I spend lots of time watching TV, I end up feeling declining life-satisfaction, which tends to eventually cause me to pay more attention to long-term goals.

But the relevant emotions still don’t seem symmetrical.

One reason for asymmetry is that different goals imply different things for what constitutes neglecting a goal: neglecting sleep or food for a day implies something more unfair to the relevant sub-agents than does neglecting one’s career skills.

Another reason is that for both time-preference and social desirability conflicts, we have instincts that aren’t optimized for our current environment.

Our hunter-gatherer ancestors needed to devote most of their time to tasks that paid off within days, and didn’t know how to devote more than a few percent of their time to usefully preparing for events that were several years in the future. Our farmer ancestors needed to devote more time to 3-12 month planning horizons, but not much more than hunter-gatherers did. Today many of us can productively spend large fractions of our time on tasks (such as getting a college degree) that take more than 5 years to pay off. Social desirability biases show (less clear) versions of that same pattern.

That means we need to override our system 1 level heuristics with system 2 level analysis. That requires overriding the instinctive beliefs of some sub-agents about how much attention their goals deserve. Whereas the long-term goals we override to deal with hunger have less firmly established “rights” to fairness.

Also, there may be some fairness rules about how often system 2 can override system 1 agents – doing that too often may cause coalitions within system 1 to treat system 2 as a politician who has grabbed too much power. [Does this explain decision fatigue? I’m unsure.]

Other Models of Willpower

The depletion model

Willpower depletion captures a nontrivial effect of key sub-agents rebelling when their goals have been overlooked for too long.

But I’m confused – the depletion model doesn’t seem like it’s trying to be a complete model of willpower. In particular, it either isn’t trying explain evolutionary sources of willpower problems, or is trying to explain it via the clearly inadequate claim that willpower is a simple function of current blood glucose levels.

It would be fine if the depletion model were just a heuristic that helped us develop more willpower. But if anything it seems more likely to reduce willpower.

Kurzban’s opportunity costs model

Kurzban et al. have a model involving the opportunity costs of using cognitive resources for a given task.

It seems more realistic than most models I’ve seen. It describes some important mental phenomena more clearly than I can, but doesn’t quite seem to be about willpower. In particular, it seems uninformative about differing time horizons. Also, it focuses on cognitive resource constraints, whereas I’d expect some non-cognitive resource constraints to be equally important.

Ainslie’s Breakdown of Will

George Ainslie wrote a lot about willpower, describing it as intertemporal bargaining, with hyperbolic discounting. I read that book 6 years ago, but don’t remember it very clearly, and I don’t recall how much it influenced my current beliefs. I think my model looks a good deal like what I’d get if I had set out to combine the best parts of Ainslie’s ideas and Kurzban’s ideas, but I wrote 90% of this post before remembering that Ainslie’s book was relevant.

Ainslie apparently wrote his book before it became popular to generate simple models of willpower, so he didn’t put much thought into comparing his views to others.

Hyperbolic discounting seems to be a real phenomenon that would be sufficient to cause willpower-like conflicts. But I’m unclear on why it should be a prominent part of a willpower model.

Distractible

This “model” isn’t designed to say much beyond pointing out that willpower doesn’t reliably get depleted.

Hot/cool

A Hot/cool-system model sounds like an attempt to generalize the effects of the autonomic nervous system to explain all of willpower. I haven’t found it to be very informative.

Muscle

Some say that willpower works like a muscle, in that using it strengthens it.

My model implies that we should expect this result when preparing for the longer-term future causes our future self to be safer and/or to more easily satisfy near-term goals.

I expect this effect to be somewhat observable with using willpower to save money, because having more money makes us feel safer and better able to satisfy our goals.

I expect this effect to be mostly absent after using willpower to loose weight or to write a will, since those produce benefits which are less intuitive and less observable.

Why do drugs affect willpower?

Scott at SlateStarCodex asks why drugs have important effects on willpower.

Many drugs affect the autonomic nervous system, thereby influencing our time preferences. I’d certainly expect that drugs which reduce anxiety will enable us to give higher priority to far future goals.

I expect stimulants make us feel less concern about depleting our available calories, and less concern about our need for sleep, thereby satisfying a few short-term sub-agents. I expect this to cause small increases in willpower.

But this is probably incomplete. I suspect the effect of SSRIs on willpower varies quite widely between people. I suspect that’s due to an anti-anxiety effect which increases willpower, plus an anti-obsession effect which reduces willpower in a way that my model doesn’t explain.

And Scott implies that some drugs have larger effects on willpower than I can explain.

My model implies that placebos can be mildly effective at increasing willpower, by convincing some short-sighted sub-agents that resources are being applied toward their goals. A quick search suggests this prediction has been poorly studied so far, with one low-quality study confirming this.

Conclusion

I’m more puzzled than usual about whether these ideas are valuable. Is this model profound, or too obvious to matter?

I presume part of the answer is that people who care about improving willpower care less about theory, and focus on creating heuristics that are easy to apply.

CFAR does a decent job of helping people develop more willpower, not by explaining a clear theory of what willpower is, but by focusing more on how to resolve conflicts between sub-agents.

And I recommend that most people start with practical advice, such as the advice in The Willpower Instinct, and worry about theory later.

I’ve substantially reduced my anxiety over the past 5-10 years.

Many of the important steps along that path look easy in hindsight, yet the overall goal looked sufficiently hard prospectively that I usually assumed it wasn’t possible. I only ended up making progress by focusing on related goals.

In this post, I’ll mainly focus on problems related to general social anxiety among introverted nerds. It will probably be much less useful to others.

In particular, I expect it doesn’t apply very well to ADHD-related problems, and I have little idea how well it applies to the results of specific PTSD-type trauma.

It should be slightly useful for anxiety over politicians who are making America grate again. But you’re probably fooling yourself if you blame many of your problems on distant strangers.

Trump: Make America Grate Again!

Continue Reading

Book review: Notes on a New Philosophy of Empirical Science (Draft Version), by Daniel Burfoot.

Standard views of science focus on comparing theories by finding examples where they make differing predictions, and rejecting the theory that made worse predictions.

Burfoot describes a better view of science, called the Compression Rate Method (CRM), which replaces the “make prediction” step with “make a compression program”, and compares theories by how much they compress a standard (large) database.

These views of science produce mostly equivalent results(!), but CRM provides a better perspective.

Machine Learning (ML) is potentially science, and this book focuses on how ML will be improved by viewing its problems through the lens of CRM. Burfoot complains about the toolkit mentality of traditional ML research, arguing that the CRM approach will turn ML into an empirical science.

This should generate a Kuhnian paradigm shift in ML, with more objective measures of the research quality than any branch of science has achieved so far.

Burfoot focuses on compression as encoding empirical knowledge of specific databases / domains. He rejects the standard goal of a general-purpose compression tool. Instead, he proposes creating compression algorithms that are specialized for each type of database, to reflect what we know about topics (such as images of cars) that are important to us.
Continue Reading

MIRI has produced a potentially important result (called Garrabrant induction) for dealing with uncertainty about logical facts.

The paper is somewhat hard for non-mathematicians to read. This video provides an easier overview, and more context.

It uses prediction markets! “It’s a financial solution to the computer science problem of metamathematics”.

It shows that we can evade disturbing conclusions such as Godel incompleteness and the paradox of the liar, by expecting to only be very confident about logically deducible facts (as opposed to being mathematically certain). That’s similar to the difference between treating beliefs about empirical facts as probabilities, as opposed to boolean values.

I’m somewhat skeptical that it will have an important effect on AI safety, but my intuition says it will produce enough benefits somewhere that it will become at least as famous as Pearl’s work on causality.

One of the weakest claims in The Age of Em was that AI progress has not been accelerating.

J Storrs Hall (aka Josh) has a hypothesis that AI progress accelerated about a decade ago due to a shift from academia to industry. (I’m puzzled why the title describes it as a coming change, when it appears to have already happened).

I find it quite likely that something important happened then, including an acceleration in the rate at which AI affects people.

I find it less clear whether that indicates a change in how fast AI is approaching human intelligence levels.

Josh points to airplanes as an example of a phase change being important.

I tried to compare AI progress to other industries which might have experienced a similar phase change, driven by hardware progress. But I was deterred by the difficulty of estimating progress in industries when they were driven by academia.

One industry I tried to compare to was photovoltaics, which seemed to be hyped for a long time before becoming commercially important (10-20 years ago?). But I see only weak signs of a phase change around 2007, from looking at Swanson’s Law. It’s unclear whether photovoltaic progress was ever dominated by academia enough for a phase change to be important.

Hypertext is a domain where a clear phase change happened in the earl 1990s. It experienced a nearly foom-like rate of adoption when internet availability altered the problem, from one that required a big company to finance the hardware and marketing, to a problem that could be solved by simply giving away a small amount of code. But this change in adoption was not accompanied by a change in the power of hypertext software (beyond changes due to network effects). So this seems like weak evidence against accelerating progress toward human-level AI.

What other industries should I look at?

I started writing morning pages a few months ago. That means writing three pages, on paper, before doing anything else [1].

I’ve only been doing this on weekends and holidays, because on weekdays I feel a need to do some stock market work close to when the market opens.

It typically takes me one hour to write three pages. At first, it felt like I needed 75 minutes but wanted to finish faster. After a few weeks, it felt like I could finish in about 50 minutes when I was in a hurry, but often preferred to take more than an hour.

That suggests I’m doing much less stream-of-consciousness writing than is typical for morning pages. It’s unclear whether that matters.

It feels like devoting an hour per day to morning pages ought to be costly. Yet I never observed it crowding out anything I valued (except maybe once or twice when I woke up before getting an optimal amount of sleep in order to get to a hike on time – that was due to scheduling problems, not due to morning pages reducing the available of time per day).
Continue Reading

Why do people knowingly follow bad investment strategies?

I won’t ask (in this post) about why people hold foolish beliefs about investment strategies. I’ll focus on people who intend to follow a decent strategy, and fail. I’ll illustrate this with a stereotype from a behavioral economist (Procrastination in Preparing for Retirement):[1]

For instance, one of the authors has kept an average of over $20,000 in his checking account over the last 10 years, despite earning an average of less than 1% interest on this account and having easy access to very liquid alternative investments earning much more.

A more mundane example is a person who holds most of their wealth in stock of a single company, for reasons of historical accident (they acquired it via employee stock options or inheritance), but admits to preferring a more diversified portfolio.

An example from my life is that, until this year, I often borrowed money from Schwab to buy stock, when I could have borrowed at lower rates in my Interactive Brokers account to do the same thing. (Partly due to habits that I developed while carelessly unaware of the difference in rates; partly due to a number of trivial inconveniences).

Behavioral economists are somewhat correct to attribute such mistakes to questionable time discounting. But I see more patterns than such a model can explain (e.g. people procrastinate more over some decisions (whether to make a “boring” trade) than others (whether to read news about investments)).[2]

Instead, I use CFAR-style models that focus on conflicting motives of different agents within our minds.

Continue Reading