I don’t agree entirely with this post by Patri Friedman, but I think it’s pointing at something important.  Patri is disturbed by the politicization of previously non-political organizations — people being fired from jobs or booted from conferences for their political views, private businesses discriminating against gays, etc.

If I try to defend the intuition behind this pattern, it’s something like “Workplaces should be about work, conferences should be about the conference topic, bakeries should be about baking — there should be an ‘impersonal’ or ‘public’ space that doesn’t care about your personal views or private life.”  It is nicer to live in a world where people mind their own business.  While in principle nosiness doesn’t violate the non-aggression principle, it’s perfectly reasonable to dislike busybodies who take every opportunity to call a referendum on who is a Good Person and who is a Bad Person, based on every detail you’ve ever posted on social media.

In a world where the distinction between public life and private life is fluid, it’s possible to be infinitely nosy and judgmental.  In the limit of this process, everything you’ve ever said or done is searchable and statistically analyzable, and all the minutiae of your life are potentially relevant to any situation.

And in a sense, this isn’t actually false. Everything you’ve ever said or done does reflect holistically on who you are as a person.   Seemingly disparate facts about a person might actually be connected to deep underlying facets of their personality or core beliefs.  I’ve argued that aesthetic tastes have moral relevance.  Scott Alexander has speculated about a General Factor of Correctness, so that people who are in some overall sense “better thinkers” than average are more likely to have correct opinions on all topics.  Jonah Sinick has ideas about a personal quality of “aesthetic discernment” underlying intellectual achievement across many disciplines.  IQ itself is an idea in this vein — the hypothesis that a single factor about human minds determines a wide variety of life outcomes, such that people with high IQs seem to do better at almost everything.  All of these ideas are sort of dancing around the intuition that you can look at somebody’s life as a whole and evaluate how generally good at life they are — and thus that boundaries like “personal vs. professional” or “personal vs. political” or “expertise in one’s field vs. expertise outside one’s field” are arbitrary.

If you believe in something like arete, and you’re a good Bayesian, then all the minutiae of a person’s life really are evidence about how generally excellent that person is.  And this isn’t necessarily crazy.  Beauty, intelligence, wealth, and health all correlate positively. Being successful in one area does make you more likely to be successful in others.  If arete “is a thing” (i.e. is a large principal component, speaking loosely), then the halo effect is an exaggeration of a real phenomenon.

If you believe in taking ideas seriously, and particularly if you are a “hedgehog“, then all your opinions and emotions potentially are reflections of your core beliefs and subconscious assumptions about the world.  The more integrated you are, the more connected your views are, even about seemingly disparate topics.  Aesthetic, intellectual, romantic, political, professional — it’s all part of a single worldview.

The problem with these kinds of “unifying” thoughts is that they make it very easy to be extremely judgmental, in a negative sense. If “the personal is political”, then every detail of your private life can betray the Cause. If aesthetics have profound moral import, then you can be berated for the media you consume (everything from “your fave is problematic” to “Mozart was a Red.”)  If your private life reflects your overall competence, then any photo of you at a party can jeopardize your career.  If there’s no such thing as “minding your own business”, if it’s possible and therefore imperative to judge everyone on everything, the result is endless conflict over minutiae.  I’m pretty far towards the “humorless extremist” side of the spectrum, myself — I was always a big fan of the Transcendentalists — but even I can see that we’ll eventually have to reach Peak Busybody.

The opposite of nosiness is forbearance.

Forbearance is the radical notion that not everybody has to be your soulmate.

A real soulmate probably is aligned with you on fundamental values. And aesthetics, and intellectual interests, and so on.  When someone is your soulmate, your kindred spirit, your twin flame, you’ll see the world the same way, to a very high degree of precision.  If you meet one, you might want to marry them.  If you meet several, you’re very lucky.  Soulmates are wonderful.

But there are seven billion people in the world.

Most people are not going to see the world as you do. Most people do not share your philosophy and aesthetics.  It’s possible to live peacefully with most of them anyway. It’s possible to be neighbors and coworkers and customers and acquaintances with people who aren’t your soulmates. It’s possible to cooperate with people who aren’t your soulmates.

One of the things I’ve learned to respect about the field of sales is that salespeople really understand this.  I have repeatedly thought, “There’s no way this deal could possibly work! The other guys don’t even live in the same reality as us! They don’t share our values at all!  Hey wait, why are you trying to be nice to them?”   But it does work. In a negotiation, the other guy doesn’t have to be your soulmate for cooperation to be possible. There just has to be a ZOPA.

Compared to the exaltation of connecting with a soulmate, forbearance seems kind of boring. Forbearance involves a lot of politeness and conventional formality.  Robert Heinlein said,

Moving parts in rubbing contact require lubrication to avoid excessive wear. Honorifics and formal politeness provide lubrication where people rub together. Often the very young, the untravelled, the naive, the unsophisticated deplore these formalities as “empty,” “meaningless,” or “dishonest,” and scorn to use them. No matter how “pure” their motives, they thereby throw sand into machinery that does not work too well at best.

Forbearance means not talking about stuff — declining to give opinions outside your field, not talking politics at the dinner table, not oversharing about your personal life at the office. It means deliberately compartmentalizing, even when you could unify all these areas of life, for the sake of reducing conflict.

Forbearance means being nonjudgmental, not in the sense of turning off your mind, but in the sense of choosing not to pick every fight.

Forbearance looks uncomfortably like being dull and middle-aged.  (Though it can also mean being funny and self-skeptical and humane, like Robert Anton Wilson or Douglas Adams.)

The reasons to practice forbearance are scope and freedom.  Scope, because you can’t do things that involve large numbers of people — working in a team of more than eight, say — unless you can cooperate with people who aren’t your soulmates.  Freedom, because you can’t really express yourself if your employer or the internet hivemind will shut you down the instant you step out of line.

I wouldn’t want to practice forbearance everywhere and with everyone.  Life would be a lot less passionate.  Forbearance isn’t as necessary with kindred spirits; the whole point of connecting with someone deeply is that you can engage with your whole self rather than an impersonal facade.  And the great advantage of writing in public is that you can find new kindred spirits through correspondence.  But if you can’t ever practice forbearance, you paint yourself in a corner.

We could regrow livers

There are currently 16,000 Americans on the waiting list for a liver transplant, but there are only enough livers for 6000 transplants a year.  Every year, more than 1500 people die waiting for a liver transplant.

One commonly mentioned idea to close the organ donor gap is to pay people for their organs to incentivize more donation.

New science may open other possibilities as well. Eric Lagasse’s lab at the University of Pittsburgh’s McGowan Center for Regenerative Medicine has been experimenting with lymph nodes as a transplantation site.  Simply put, if you put hepatocytes (liver cells) into a lymph node, the node will grow into a functioning mini-liver.  This rescues mice from lethal liver failure.

Injecting cells into lymph nodes also works with thymus cells, which can give athymic mice a functioning immune system.  And it works with pancreas cells, which, when injected into lymph nodes, can rescue mice from diabetes.

The procedure only partially works with kidneys, which are much more structurally complex — the cells implanted into lymph nodes show some signs of growing into nephrons, but aren’t completely functional.

Interestingly enough, this effect was documented in 1963 by immunologist Ira Green. If you remove the spleen or thymus from a mouse, and replace it with ectopically transplanted spleen or thymus tissue, the tissue grows into a functioning, structurally normal, miniature thymus or spleen.

In 1979, researchers found that hepatocytes injected into the rat spleen (which is part of the lymphatic system and analogous to a large lymph node) functioned normally and grew to take up 40% of the spleen.

There have been clinical studies in humans of hepatocyte  transplantation, generally with less than impressive results, but generally these hepatocytes are infused through the portal vein or renal artery, whence a small fraction of the cells reach the liver and spleen. It’s still possible that injection into lymph nodes would be more effective. As the above article states,

One possible explanation for the discrepancy between the laboratory and clinical outcomes may relate to the route of hepatocyte delivery. Following infusion using direct splenic puncture, dramatic corrections in liver function have accompanied hepatocyte transplantation in laboratory animals. In patients with cirrhosis, however, allogeneic hepatocytes have been delivered to the spleen exclusively through the splenic artery.

A natural question to ask is: why isn’t more being done with this?  “Inject liver cells into lymph nodes” is not a particularly high-tech idea (as far as my layman’s understanding goes.)  Nor is it a completely new idea; researchers have known for decades that hepatocytes grow into functioning liver tissue, particularly when injected into the lymphatic system.  You’d think that a procedure that could replace liver transplants would be profitable and that founding a biotech company to do human trials would be a tremendous opportunity, especially since there is less scientific risk than there often is with other early-stage biomedical research (e.g. preclinical drugs).

Part of the problem is that the business model in such cases is unclear.  This has been a pattern we noticed several times at MetaMed; often a medical breakthrough is not a new drug or device, but a novel medical procedure.  It is not permissible to enforce a patent (e.g. to sue someone for infringement) on a medical or surgical procedure.  Medical ethics (for instance, see this statement from the American Academy of Orthopedic Surgeons) generally holds that it’s unethical to patent a surgical procedure.  This means that it’s difficult to profit off the invention of a medical or surgical procedure.  The total value of being able to offer a liver transplant to anyone who wants one would be billions of dollars a year– but it’s not clear how anybody can capture that value, so there’s less incentive (apart from humanitarian motives) to develop and implement such a procedure.

It also means that it’s difficult to disseminate information about new medical or surgical procedures. Learning to perform procedures is an apprenticeship process; one doctor has to teach another. This, combined with natural risk aversion, means the spread of new procedures is slow. If a new surgery had been shown conclusively, by excellent experimental evidence, to be better than the old one, it still would not necessarily sweep the nation; if the clinician who pioneered the new surgery isn’t a natural evangelist, it may never be performed in more than one hospital.

This seems to be an opportunity in search of a viable strategy. There are spectacular results in regenerative medicine (frequently coming out of the McGowan Institute — see Stephen Badylak’s work in tissue regeneration).  It’s not clear to me how one would make those results “scale” in the sense that we’re used to in tech companies.  But if you could figure out a model, the market size is mind-boggling.

If you had a way to regrow organs, how would you validate it experimentally? And how would you get it to patients?  And how would you do it fast enough not to lose tens of thousands of lives from delay?

Aesthetics are moral judgments

I often hear people say things like “It’s ridiculous to judge that someone’s a bad person because of his musical taste!” People assume it’s obvious that aesthetic judgments have no moral weight.

For me, aesthetic judgments are a kind of moral judgment.

I understand “morality” to basically cash out as “priority structure”, “values”, and related concepts. What matters most to me, and what would matter most to me if I knew more and thought more clearly.  With that definition, when I say that kindness is “good” and I say that Camembert is “good”, I’m not using two unrelated meanings of the word — cheese and kindness are both valuable to me.

Aesthetic preferences aren’t really arbitrary; they say things about what you value and how you see the world.

For example, I like Bach.  There’s a pretty well-established correlation between liking Bach and liking math. Godel, Escher, Bach is a pretty strong marker of membership in my tribe.  And I don’t think that’s arbitrary.  The words I’d use to describe Bach’s music are complex and orderly.  Polyphony gives the impression of a giant, intricate clock, moving according to regular mechanisms, steady as the stars in their courses and endlessly interesting.  It gives me a sense of cosmos, of natural law.  And the fact that I like that says something about what my priorities are more generally.

These sorts of connections are associative and probabilistic rather than determined. Not literally everyone who likes Bach is getting the same associations from the music as I do. But associations and resonances can be real tendencies in the world even if they’re not strict logical entailments.  Metaphors can be apt. There are some synesthetic/metaphorical connections that correlate across human minds, like the bouba/kiki effect.  In a “clusters-in-thingspace” sense, it can be sort of objectively true that Bach is “about” cosmic natural order.  You can’t stretch these intuitions too far, but they aren’t completely fictitious either.

And it’s possible to learn aesthetic intuitions.

I used to only like paintings with very crisp, precise textures, rather than the cloudy, fuzzy textures that show up in John Singer Sargent or Turner paintings.  The art blog Opulent Joy taught me to appreciate the soft textures; when I realized “oh! he’s appreciating a broader power spectrum than I am!” I immediately noticed that his aesthetic was like mine, but stronger — more general, more nuanced, and therefore an upgrade I would like to make.

Another example: when I was a kid,  I found industrial landscapes horribly ugly.  Machines seemed like a blight on nature.  The more I came to understand that good things are produced by machines, and that machines are made with care and skill, the more I started to see trains and bridges and construction sites and shipping containers as beautiful.  Factual understanding changed my aesthetic appreciation. And if learning facts changes your aesthetic views, that means that they aren’t arbitrary; they actually reflect an understanding of the world, and can be more correct or less so.

Judging people for aesthetics isn’t crazy.  If someone loves Requiem for a Dream,  it’s a small piece of evidence that they’re a pessimistic person.  If you think pessimism is bad, then you’re indirectly judging them for their taste. Now, your inferences could be wrong — they could just be huge Philip Glass fans — once again, we’re looking at Bayesian evidence, not logical entailments, so being overconfident about what other people’s tastes “say about them” is a bad idea.  But aesthetics do more-or-less mean things.

For me, personally, my aesthetic sensitivities are precise in a way my moral intuitions aren’t.  My “conscience” will ping perfectly innocent things as “bad”; or it’ll give me logically incoherent results; or it’ll say “everything is bad and everyone is a sinner.” I’ve learned to mistrust my moral intuitions.

My aesthetic sensibilities, on the other hand, are stable and firm and specific. I can usually articulate why I like what I like; I’m conscious of when I’m changing my mind and why; I’m confident in my tastes; my sophistication seems to increase over time; intellectual subjects that seem “beautiful” to me also seem to turn out to be scientifically fruitful and important.  To the extent that I can judge such things about myself, I’m pretty good at aesthetics.

It’s easier for me to conceptualize “morality” as “the aesthetics of human relationships” than to go the other way and consider aesthetics as “the morality of art and sensory experience.”  I’m more likely to have an answer to the question “which of these options is more beautiful?” than “which of these options is the right thing to do?”, so sometimes I get to morality through aesthetics. Justice is good because symmetry is beautiful.  Spiteful behavior is bad because resentment is an ugly state to be in.  Preserving life is good, at root, because complexity is more interesting and beautiful than emptiness.  (Which is, again, probably true because I am a living creature and evolutionarily wired to think so; it’s circular; but the aesthetic perspective is more compelling to me than other perspectives.)

It always puzzles me when people think of aesthetics as a sort of side issue to philosophy, and I know I’ve puzzled people who don’t see why I think they’re central.  Hopefully this gives a somewhat clearer idea of how someone’s internal world can be “built out of aesthetics” to a very large degree.

Epistemology Sequence, Part 5: Extension and Universality

One of the properties that you’d like a learning agent to have is that, if your old concepts work well, learning a new concept should extend your knowledge but not invalidate your old knowledge. Changes in your ontology should behave in a roughly predictable manner rather than a chaotic manner.  If you learn that physics works differently at very large or very small scales, this should leave classical mechanics intact at moderate scales and accuracies.

From a goal-based perspective, this means that if you make a desirable change in ontology — let’s say you switch from one set of nodes to a different set — and you choose the “best” map from one ontology to another, in something like the Kullback-Leibler-minimizing sense described here — then when you take preimages of your “utility functions” on the new ontology onto the old ontology, they come out mostly the same.  The best decision remains the best decision.

In the special case where the old ontology is just a subset of the new ontology, this means that the maps between them are a restriction and an extension.  (For example, if we restrict (a, b, c, d, e) to the first three coordinates, it’s just the identity operation on those coordinates, (a, b, c); and if we extend (a, b, c) to (a, b, c, d, e), again the map is the identity on the first three coordinates.)  What we’d like to say is that, when we add new nodes to our ontology, then the function that computes values on that ontology (the f in Part 3 of this sequence) extends to a new f on the new ontology, while keeping the same values on the old nodes.

For example; let’s say I have a regression model that predicts SAT scores as a result of a bunch of demographic variables. The “best” model minimizes the sum of squared errors. Sum of squared errors is my utility function.  Now, if I add a variable to my model, the utility function stays the same, it’s still “sum of squared errors”; so if adding that new variable changes the model but reduces the residuals, my old model “wants” to make the upgrade.  On the other hand, an ‘upgrade’ to my model that changes the utility function, like deciding to minimize the sum of squared errors plus the coefficient for the new variable, isn’t necessarily an improvement unless the “best” model by that criterion also shrinks the sum of squared errors relative to the original regression model.

From the goal-oriented perspective, the only changes you’d want to make to your ontology are those which, when “projected” onto the old ontology, have you making the same “optimal” choices.

(These statements still need to be made precise. There may be a conjecture in there but I haven’t specified it yet. The whole business smells like Hahn-Banach to me, but I could be entirely mistaken. The universality of neural nets might be relevant to showing that this kind of a “rational learner” is implementable with neural nets in the first place. )

Epistemology Sequence, Part 4: Updating Ontologies Based On Values

What’s a good ontology?

Well, the obvious question is, good relative to what?

Relative to your values, of course. In the last post I talked about how, given an ontology for describing the world, you can evaluate a situation.  You compute the inner product


where V represents a value function on each of the concepts in your ontology, and f(N) is a function of which concept nodes are “active” in a situation, whether by direct perception, logical inference, predictive inference, association, or any other kind of linkage.  For instance, the situation of being a few yards away from a lion will activate nodes for “tan”, “lion”, “danger”, and so on.

If you can evaluate situations, you can choose between actions. Among all actions, pick the one that has the highest expected value.

One particular action that you might take is changing your ontology.  Suppose you add a new node to your network of concepts.  Probably a generalization or a composition of other nodes. Or you subtract a node.  How would you decide whether this is a good idea or not?

Well, you build a model using your current ontology of what would happen if you did that. You’d take different actions.  Those actions would lead to different expected outcomes. You can evaluate how much you like those outcomes using your current ontology and current values.

For  modeling the world, the kinds of things you might optimize for are accuracy (how often does your model come up with correct predictions) and simplicity (how few degrees of freedom are involved.)  This is often implemented in machine learning with a loss function consisting of an error term and a regularization term; you choose the model that minimizes the loss function.

Notice that, in general, changing your ontology is changing your values. You can’t prioritize “civil rights” if you don’t think they exist.  When you learn that there are other planets besides the Earth, you might prioritize space exploration; before you learned that it was possible, you couldn’t have wanted it.

The question of value stability is an important one. When should you self-modify to become a different kind of person, with different values?  Would you take a pill that turns you into a sociopath?  After all, once you’ve taken the pill, you’ll be happy to be free of all those annoying concerns for other people.  Organizations or computer programs can also self-modify, and those modifications can change their values over time.  “Improvements” meant to increase power or efficacy can cause such agents to change their goals to those that present-day planners would find horrifying.

In the system I’m describing, proposed changes are always evaluated with respect to current values.  You don’t take the sociopath pill, because the present version of you doesn’t want to be a sociopath. The only paths of self-modification open to you are those where future states (and values) are backwards-compatible with earlier states and values.

The view of concepts as clusters in thingspace suggests that the “goodness” of a concept or category is a function of some kind of metric of the “naturalness” of the cluster.  Something like the ratio of between-cluster to within-cluster variance, or the size of the margin to the separating hyperplane.  The issue is that choices of metric matter enormously.  A great deal of research in image recognition, for example, involves competing choices of similarity metrics. The best choice of similarity metric is subjective until people agree on a goal — say, a shared dataset with labeled images to correctly identify — and compete on how well their metrics work at achieving that goal.

The “goodness” or “aptness” of concepts is a real feature of the world. Some concepts divide reality at the joints better than others. Some concepts are “natural” and some seem contrived.  “Grue” and “bleen” are awkward, unnatural concepts that no real human would use, while “blue” and “green” are natural ones.  And yet, even blue and green are not human universals (the Japanese ao refers to both blue and green; 17th century English speakers thought lavender was “blue” but we don’t.)  The answer to this supposed puzzle is that the “naturalness” of concepts depends on what you want to do with them.  It might be more important to have varied color words in a world with bright-colored synthetic dyes, for instance; our pre-industrial ancestors got by with fewer colors.  The goodness of concepts is objective — that is, there is a checkable, empirical fact of the matter about how good a concept is — but only relative to a goal, which may depend on the individual agent.  Goals themselves are relative to ontology.  So choosing a good ontology is actually an iterative process; you have to build it up relative to your previous ontology.

(Babies probably have some very simple perceptual concepts hard-coded into their brains, and build up more complexity over time as they learn and explore.)

It’s an interesting research problem to explore when major changes in ontology are desirable, in “toy” computational situations.  The early MIRI paper “Ontological crises in artificial agent’s value systems” is a preliminary attempt to look at this problem, and says essentially that small changes in ontologies should yield “near-isomorphisms” between utility functions.  But there’s a great deal of work to be done (some of which already exists) about robustness under ontological changes — when is the answer spit out by a model going to remain the same under perturbation of the number of variables of that model?  What kinds of perturbations are neutral, and what kinds are beneficial or harmful?  Tenenbaum’s work on learning taxonomic structure from statistical correlations is somewhat in this vein, but keeps the measure of “model goodness” separate from the model itself, and doesn’t incorporate the notion of goals.  I anticipate that additional work on this topic will have serious practical importance, given that model selection and feature engineering is still a labor-intensive, partly subjective activity, and that greater automation of model selection will turn out to be valuable in technological applications.

Most of the ideas here are from ItOE; quantitative interpretations are my own.