The difference with sleep is that meditation increases perceptual clarity instead of decreasing it. The experience of seeing your mind turn off and on while seeing clearly everything that is happening is actually one of the most profound things that can happen in meditation because you can see in real time the different parts of the mind shutting down one by one, giving you insight into what those parts actually are.
Experiences of a lack of free will are relevant to understanding the nature of the mind because the usual presumption is that free will is somehow a property of decision-making itself, not a weird tacked-on module of the brain that can easily be dispensed with. If free will is just a thin coat of paint over the decision making module, then it’s not really free will at all. Advanced meditation lets you experience making eggs in the morning, taking a jog, talking to your spouse, writing your emails all without the experience you might usually have called “free will”. What was intuitively thought to be a crucial part of decision making is seen to be not necessary at all. But it’s more than that, because it doesn’t feel like “the free will module stopped”, it feels more like “what I thought was the free will module is actually 5 other different modules, and my lack of perceptual clarity was kind of blending the 5 together in a confused mess”.
I think the main point that people are missing here is that Sam Harris is an experienced meditator with years of intensive retreat practice. This means that he likely routinely enters states where the perception of any “doer” or “observer” in subjective experience goes away, states where no “decision” is ever taken, but events “just happen” on their own. This is why Harris has such strong opinions on free will, advanced meditation practice will directly show you how the appearance of decisions and free will is being constructed moment-by-moment, no need for any philosophical arguments about determinism when you can simply look at your immediate experience and see it lacking any free will. I suspect he wrote the book because he couldn’t just say “trust me, if you meditate for 5000 hours you’ll see you have no free will”, even though that would likely be a more honest answer.
The correct position on the viability axis will depend on the type of business you want to build. Experimenting on real users has one major cost: reputation. Starting too far to the left on the viability spectrum risks users remembering your first crappy product and not trusting you anymore. This reputational cost is very impactful for businesses that rely on subscriptions or long-term relationships with clients, so those businesses ought to start out already kind-of-viable. But if your business is something like a product that you only sell once, without any long-term client relationship then you can afford to alienate early users with a possibly crappy first product, and you should test stuff as soon as possible.
My impression is that a “model” in this context is fundamentally about predicting the future, I don’t see how having the network answer “how many spaceships are on the screen right now?” would give us any indication about it having built something more complex than a simple pattern-recognizer.
Maybe a better way to detect an internal model is to search for subnetworks that are predicting the future activations of some network nodes. So for instance we might have a part of the network that is detecting the number of spaceships on the screen right now, which ought to be a simple pattern recognition subnetwork. Then, we search for a node in the network which predicts the future activation of the spaceship-recognition-node, computing the autocorrelation between the timeseries of the two nodes. If we find one, then this is strong evidence that the network is simulating the future in some way internally.
In fact we can use this to enforce model-building, we could divide the network into a pattern-recognition half, which would be left alone (not extra constraints put on it), and a modelling half, which would be forced to have its outputs predict the future values of the pattern-recognition half (again with multi-objective optimisation). Of course the pattern-recognition half could also take inputs from the modelling half.
But … we have calculators in our pockets. I’m not sure how you get to do so many fermi estimates that it’s actually time-efficient to do the arythmetic in your head instead of the much more accurate and faster calculator on your phone. Doing it in your head isn’t faster and just adds additional uncertainties on top of the uncertainties in each of your inputs.
I expect that it will help with the tedious parts of programming that require you to know details about specific libraries. It will make those parts of programming that are “google → stackoverflow → find correct answer → copy-paste into code” much fater, but I expect it to struggle on truly novel stuff, on scientific programming and anything research-level. If this thing can auto-complete a new finite-element PDE solver in a way that isn’t useless, I will be both impressed and scared.
This will depend on what you mean by “moral world”. In my view morality is a property of the human brain, the universe itself doesn’t have morality built into it apart from the fact that human brains are inside the universe. Do you mean something like a “Just World” where moral behaviour is rewarded and immoral behaviour is punished?
As an argument for not requiring God, I’d say that in my experience immoral behaviour is its own punishment. Anger, hate, jealousy and all of those emotions are very unpleasant to feel, and they bias the brain away from kindness, joy and compassion, which are immensely more pleasurable (purely as a matter of feeling tone). Immoral acts themselves lead away from these positive brain states, and this causes a certain built-in justice to the world, no need for God, a world where immoral acts prevent you from experiencing the most pleasant and refined mind-states is Just on its own.
The best steelman I’ve heard for the approval is that the FDA just created a multi-billion dollar market for Alzheimer drugs. Once people are paying 50k/year for this drug, other companies will have enormous incentives to try to get a piece of the market, which will create better treatments over the long run. So yes, right now there’s gonna be a lot of money wasted on an ineffective treatment, but here the FDA might actually be incentivizing more innovation.
It really wouldn’t be possible to take 4-6 hours to talk to other people, that would completely take you out of it, it would kill a lot of momentum. After 3 days of really spending every waking hour thinking of a subject, it sort of temporarily becomes the new baseline, and the difficulty of thinking about it drops really dramatically, but it is a fragile effect, any context switching at all comes at the cost of a blunting of momentum, and this is especially true of talking to other people. If you’ve spent a lot of time just thinking in silence, 4 hours of talking with others will literally cause a pounding headache. I took about 3 to 4 hours per day of break-time to eat and go walk without feeling the need to think of the material, but in the later part of the 10-day period I usually think about the textbook even during breaks. I’m not sure how you’d manage this with a job and daily commitments, what I do is do one of these 10-day periods every 2 months, but I run a small business that allows me to do that, and I don’t have a romantic relationship.
I might be atypical in this respect, but I find that the best time budgeting strategy for me is entirely different: I take 10 days and focus on a single textbook, I don’t do much of anything else except take walks to think about the material and go back to read the textbook. I don’t check the internet, don’t check email, don’t read any other books, don’t go out with friends or work on other projects, I just focus all my time on that one textbook. This works best if I’m not doing this for an exam, but just for my own goals, an exam distorts the way that I read a textbook and makes me constantly think about gaming my grade.
I started doing this by analogy with meditation retreats. It’s very well-known in meditation circles that a 10-day intensive retreat makes you progress dramatically faster than a simple daily practice. So I started doing study retreats on a single topic. It takes about 3 days to really get into a groove with the schedule of studying 10 to 12 hours a day, but after that it gets easier. Really small distractions have a dramatic impact, even just talking to someone else for 10 minutes have a perceptible impact on the feeling of study momentum.
I personally find notes and visualizations kind of distracting while reading, the cue of trying to “visualize” something like the Schrodinger equation doesn’t help much, visualizing the symbols themselves won’t do anything for your understanding of it. So what I like to do is take a short break after every section in a chapter and ask myself “what does this knowledge imply?”, “What is the next logical step to take from here?”, basically trying to predict what the next pages in the book will contain. This makes it easier to see which leaps of logic in a book were easy and obvious, and which were much harder. It also provides immediate feedback and lots of “goddamnit, of course, how could I be so stupid?!” moments when I read the book take what is obviously the only possible next step.
Oh, I wasn’t claiming originality, just trying to give some background to people who might have stumbled here.
The way I think of the Lottery Ticket Hypothesis is in terms of the procedure for actually finding a lottery ticket: start with a random initialised network, then train it, then look at the weights of the network that are the largest after training (say keep only 10% or 1% of the weights), go back to the initial random network and drop all the weights that won’t end up being the largest, now train that highly sparse network and you’ll end up with close to the same end performance, even though you’re using a much smaller network.
This seems to mean that all those weights we dropped don’t really have any effect on the training, we might have imagined that weird nonlinear effects might mean that getting to a good final solution would require the presence of weights that in the end will be useless, those weights might have been “catalysts”, helping the network arrive at its final solution, but not themselves useful for prediction. But no, it seems that if weights are going to end up useless, then they’re useless from the beginning.
I largely agree that anger and emotional happiness (but not ultimate Happiness) are quite incompatible, doing 1200 hours of meditation over two years has reduced both the frequency and half-life of any anger I feel and it made me realise just how unpleasant it actually was. However, let me try to steelman anger:
First, very advanced meditators actually seem to report that even moments of anger are “perfect” in some way, just like all other moments. I’ve certainly seen Shinzen Young and other zen masters get somewhat angry, not at anyone, but to emphasise a particular point in an emotional way. So anger is certainly useful for communicating your values to an audience. Hearing someone say “I really care about this” in a neutral tone doesn’t quite have the punch of someone getting mildly angry.
There also seems to be a relationship between anger and a sort of aggressive motivation. I notice that my bench press sets seem a lot easier if I get angry first, and they’re also a lot more pleasant, being angry without doing anything is unpleasant, but expending that “angry energy” towards a goal does seem to be pleasant. Being angry about things like the existence of cancer can certainly help with motivation to solve these problems.
Does FTX let you do the Perpetual Future Arbitrage trade on margin? I was seeing 20x leverage offered without additional fees on their website, that would come out to like 50% monthly gain, which seems batshit insane to me.
I agree with the argument conditional on us not being fundamentally confused about anthropics, but as with most of these sorts of arguments, most of my uncertainty is coming from what is meant by “reference class” and “observer”. I have the nagging feeling that all of anthropics stems from us not being able to properly define what an observer actually is. And I think it’s possible that observers actually don’t exist and we’re completely confused about all of this.
Though any belief so extreme wouldn’t really feel like a “belief” in the colloquial sense, I don’t internally label my belief that there is a chair under my butt as a “belief”. That label instinctually gets used for things I am much less certain about, so most normal people doing an internal search for “beliefs” will only think of things that they are not extremely certain of. Most beliefs worth having are extreme, but most beliefs internally labeled as “belief” worth having are not extreme.
In the specific scenario of “Mark Xu asks to bet me about the name on his driver’s license”, my confidence drops immediately because I start questioning his motives for offering such a weird bet.
Though even there, his lectures are famous for only being truly appreciated after you’ve first learned the material elsewhere. They are incredibly good at giving you the feeling of understanding but quite a bit less good at actually teaching problem-solving. When reading them, it was a common occurrence for me to read a chapter and believe the subject was the most straightforward and natural thing in the world, only to be completely mystified by the problems.
I expect any version of “align narrowly superhuman models” which evaluates the success of the project entirely by human feedback to be completely and totally doomed, at-best useless and at-worst actively harmful to the broader project of alignment
There are plenty of problems where evaluating a solution is way way easier than finding the solution. I’m doubtful that the model could somehow produce a “looks good to a human but doesn’t work” solution to “what is a room-temperature superconductor?”. I agree that for biological problems the issue is much more concerning, and certainly for any kind of societal problem, but as long as we stay close to math, physics and chemistry, “looks good to a human” and “works” are pretty closely related to each other.