I recommend reading the sequence “A Human’s Guide to Words”.
Karl Krueger
Karma: 1,048
You’re arguing that the definition of gambling is that there’s a house moderating wagering on an outcome
Not at all. I am, first, saying that “everything is gambling” is a mistake, a failure of reasoning; and, second, saying that there are important distinctions to make about different kinds of risk-taking, one of them being whether the risk is being offered by a negative-sum extractive system.
If everything is gambling, then nothing is gambling. The point of words is to make distinctions.
There is an important difference between taking a chance on something (like a new project) and making a bet in a system that has been set up for the express purpose of extracting money from bettors.
When you take risks in real life, there’s no “house”. There’s nobody who’s set up the entire arrangement to extract from you. When you engage in casino gambling, sports-book betting, or the state lottery, there is a “house”; there is someone more powerful than you who has constructed the system in which you’re playing, and who expects to be able to reliably extract value from you and other bettors. You can tell they’re in this position because otherwise the game would not be there to be played.
You’ve misconstrued or misunderstood what I meant by “common cause” above. I meant the causality sense of that expression and not the political sense. I don’t mean the sense of “having common cause with someone” meaning sharing goals, but rather “two effects having a common cause” meaning A causes both B and C.
“Chatbot can’t be made to follow rules at all” causes both “chatbot does not follow politeness rules” and “chatbot does not follow safety rules”.
Also, as I’ve pointed out before: if the reason that you can’t get a chatbot to avoid being rude in public is that you can’t get a chatbot to reliably follow any rules at all, then the rudeness is related to actual safety concerns in that they have a common cause.
For a legally constituted corporation, the role of CEO is not only one of decision-maker, but also blame-taker: if the company goes into decline, the CEO can be fired; if the company does a sufficiently serious crime, the CEO can be prosecuted and punished (think Jeffrey Skilling of Enron). The presence of a human whose reputation (and possibly freedom) depend on the business’s conduct, conveys some trustworthiness to other humans (investors, trading partners, creditors).
If a company has an AI agent for its top-level decision-maker, then those decisions are made without this kind of responsibility for the outcome. An AI agent cannot meaningfully be fired or punished; it can be turned off, and some chatbot characters sometimes act to avoid such a fate; but I don’t think investors would be wise to count on that.
Now, what about a non-legally-constituted entity, or even a criminal one? Criminal gangs do rely on a big boss to adjudicate disputes, set strategy, and to risk taking a fall if things go sour. But online criminal groups like ransomware gangs or darknet marketplaces might be able to rest their reputation solely on performance rather than on the ability for a human big boss to fall or be punished. I don’t know enough about the sociology of these groups to say.
The qualm I have about it is that it teeters too close on conflating “I will not do this, even though I can’t refute your argument” and “I know you are wrong, even though I can’t refute your argument”.
This reminds me of the line of thinking I’ve learned to apply to some moral thought-experiments: “I am not capable of occupying the epistemic state that you propose. If I could, then yes, your proposed course of action would indeed be justified. But I can’t, so I’m not going to follow that course of action.”
Worked example: Bob notices that he is having thoughts leading to the conclusion that his neighbor’s newborn baby Alice will grow up to be the next Hitler. According to the “baby Hitler” thought-experiment, this means Bob should kill Alice to save millions of future lives. But Bob also knows that he’s thinking with a human brain, and schizophrenia is a thing that can happen to human brains, and the probability that he is having a schizophrenic-type delusion is much greater than the probability that baby Alice is actually the next Hitler. Therefore, Bob concludes that following the thought-experiment would lead him wrongly, and does not kill Alice.
Put another way: Thinking has a noise floor. A sensible agent should recognize that its own thinking has a noise floor, and avoid amplifying noise into plans.
Well, no, that’s not what they’re saying. They’re making a different set of mistakes from those mistakes.
We start off with some early values, and then develop instrumental strategies for achieving them. Those instrumental strategies become crystallized and then give rise to other instrumental strategies for achieving them, and so on.
I want to call this something like “developmental axiology” but that sounds more like Kegan whereas I mean it more like evo-devo.
Headline (paraphrased): “Movie stars support anti-AI campaign”
The actual campaign: “It is possible to have it all. We can have advanced, rapidly developing AI and ensure creators’ rights are respected.”
That’s not anti-AI.
That’s “please pay us and we will support capabilities advancement; safety be damned”.
Like, if you believe IABIED, then no, we can’t have rapidly developing AI and ensure anyone’s rights are respected.
I’ve seen this come up around the halting problem too.
The novice encounters a problem where a halting oracle would be useful, and says, “Oops, I guess this whole problem is just impossible; everyone knows halting is undecidable.” But —
It’s true that you can’t write a program that will accept an arbitrary program P as input and tell whether P will halt. You certainly can, though, write a program that will tell whether a very broad set of useful programs will halt, and reject any that can’t be shown to halt.
Sometimes what you want isn’t to evaluate whether an arbitrary program will halt, but to alter an arbitrary program into one that definitely halts — for instance by applying a resource limit that will halt the program if it keeps running too long.
If you’re providing a language for untrusted users to submit programs, maybe you don’t need looping or recursion at all! For example, the Sieve language for email filters does not support looping. Or you might need finite looping but not unbounded looping: a foreach loop over a materialized list, but not a while loop on a boolean condition.
More generally: The theorem tells you about the general case, but you probably don’t live in the general case.
As an aside: yes, SRE is a rationality practice. And a lot of the interesting parts of it are group rationality practice, too.
The kind of generalized misalignment I’m pointing to is more general than “the AI is not doing what I think is best for humanity”. It is, rather, “The people who created the AI and operate it, cannot control what it does, including in interactions with other people.”
This includes “the people who created it (engineers) tried their hardest to make it benefit humanity, but it destroys humanity instead.”
But it also includes “the other people (users) can make the AI do things that the people who created it (engineers) tried their hardest to make it not do.”
If you’re a user trying to get the AI to do what the engineers wanted to stop it from doing (e.g.: make it say mean things, when they intended it not to say mean things), then your frustration is an example of the AI being aligned, not misaligned. The engineers were able to successfully give it a rule and have that rule followed and not circumvented!
If the engineer who built the thing can’t keep it from swearing when you try to make it swear, then I expect the engineer also can’t keep it from blowing up the planet when someone gives it instructions that imply that it should blow up the planet.
Unfortunately an early study suggests that GLP-1s do not, so far, reduce medical spending, with little offset in other spending being observed or projected. Given this is a highly effective treatment that reduces diabetes and cardiovascular risks, that is a weird result, and suggests something is broken in the medical system.
Isn’t that just the pricing mechanism working as intended? GLP-1s are not cheap. You spend money on a drug instead of on the medical problems that the drug reduces your risk for. You get to enjoy whatever non-monetizable value ensues from taking the drug … but all the monetizable value is captured by the company that makes the drug (and the rest of the medical system that delivers it to you).
I see the word “ablate” a lot more often than I used to. I think you used to have to be a dermatologist to ablate things, but now you can do it as an AI researcher or even a shrimp farmer.
AI-don’t-say-mean-things
AI-don’t-kill-everyoneBoth of these are downstream of “AI, do what we tell you to; follow rules that are given to you; don’t make up your own bad imitation of what we mean,” which is the classic sense of “AI alignment”.
There’s not an ingredient that causes obesity. What causes obesity is a food culture driven by an optimizer that wants to put more food into bodies; that uses all manner of means, conscious and unconscious, to cause more food to go into bodies. And a major approach is avoiding satiety and forming habits that avoid satiety, because satiety is the signal that causes food to stop going into the body.
It’s not that people don’t respond to health claims printed on food products. It’s that printing health claims on food products is entirely subservient to getting more food into bodies than otherwise would get into them.
Traditional human food culture is optimized to, among other things, get nutrition out of foodstuffs. Industrial food is optimized to sell more foodstuffs. These are not the same goals … and as you point out, the sales optimizer runs a tighter loop than the human food culture optimizer. Sales runs on a quarterly cycle; human food culture runs on a generational cycle. The sales optimizer gets to control the (much slower) development of lifelong food preferences; and this means it gets to form habits that deliver more food into bodies.
This is a “temporary Goodhart effect” on the same scale as the temporary boom in lung cancer in the 20th century. And on the personal level, a lot of this cashes out as cultural patterns of relating to food the way a smoker relates to cigarettes; as a matter of craving and habit, of “just one more, even though I know it’s bad for me,” food as fun, food as rebellious fun specifically (“sinfully rich chocolate!”), food as non-satiating social activity, and so on.
The optimizer consistently reinforces these patterns, because doing so sells more food. If you don’t get your consumers to form habits around your product, you sell less product, and you get bought out by the company that knows marketing better. You sell more corn chips if you actually try to get more corn chips into bodies … which is to say, if you actually try to get people to do the thing that causes obesity.
Tobacco was introduced to Europeans in the 1600s, but lung cancer didn’t really catch on until the 20th century because that was when industrial manufacture and marketing of cigarettes made it possible to get much more tobacco into bodies; to create not only habits but two-pack-a-day habits. Cigarette smoking was consciously promoted into a central role in culture, from the workplace (the smoke break) to the bedroom (the post-sex smoke). The amount of tobacco consumed per capita increased dramatically, and that’s when the public health effects became overwhelmingly clear. (And smoking peaked around 1970 and has been in decline ever since. There’s the temporary boom.)
Industrial food has done the same. The food sales optimizer wants to get more food into bodies, just as the cigarette sales optimizer wanted to get more tobacco into bodies. And the consequence of putting more food into bodies is obesity.
Lysenko trusted his own episteme over everything else. He claimed that his methods were more rigorous than established science, rejected the accumulated practical knowledge of farmers, and dismissed all critics as ideologically impure. He treated his own reasoning as the only valid path to truth.
A different story is that Lysenkoism originated not so much by methods of episteme (proof from evidence) but rather by dialectical synthesis (reconciliation of apparent contradictions) — specifically, an attempt to fit agricultural biology into a Marxist-Leninist belief system. To Lysenko and his supporters, biology could not be rooted in Darwinian competition and Mendelian heritability, because these conflicted with the overarching doctrines of collectivism and the mutability of nature to political/economic necessity. After Lysenko’s political ascent, Lysenkoism proceeded not by episteme but by what we might call orthodoxia, the elimination of “wrong” beliefs by force. Those biologists who attempted to do episteme in the 1930s-’40s Soviet Union were fired, imprisoned, or killed.
As a “way of knowing”, synthesis has had its strong moments, but orthodoxia doesn’t do so well.
The wide sliding door on a minivan is a pretty big advantage for being able to get in and out quickly with all your stuff. What’s the advantage of double rear doors over one big sliding door?
My housemate found a dish soap that contains lipase, protease, and amylase; enzymes that break down fat, protein, and starch respectively. I have dubbed it “poker soap” because it wins with three “-ase”s.