When making my comment on your last post, I made a wrong assumption: that the best strategy would be independent of the task. Because of this, I generalized from learning a much lower level of piano and from learning to juggle. For these things, each attempt would take from 2 seconds to 1 minute, and for such tasks I think it makes the most sense to continue after one win. But I can see the situation is very different if each attempt takes several minutes or more.
Sune
Karma: 524
Good question, I will try to be more precise. My hypothesis in the experiment above was:
The first attempt after the first win is more likely to succeed if it is made immediately after the first win than if it is made one or more days after the first win.
I’m not claiming the probability is high. At my very low skill level (I have probably spent 30 hours on Yousician a few years ago), I would expect that if a successfully played a song for the first time (after many failed attempts), I would have around 15% chance of success on the next attempt immediately afterwards. It might be very different for you.
I also made a claim that continuing to practice helps more than stopping after first win each time, but I did not suggest a way to test that. It is not clear to me what the best test is because the success rate might vary in the two setups. Here is a suggestion for a claim
On a given task, if you practice one day until you have 5 wins, the total time spent (or total number of attempts) on the task is less than if you play until first win 5 days in a row.
If you test your ability to perform the task one day after ending the practice above (day 6 in the first scenario and day 2 in the second scenario) then scenario 1 will give you the highest probability of win in first attempton the test day and it will also on average give you a smaller number of attempt until first win on the test day.
I’m less confident in the last prediction, because you don’t get the spaced repetition practice. Maybe playing until 4th win one day 1 and until first win on day 2 is better.
You could combine the two experiments above, so only after the first win do you flip a coin to decide if you are doing 5 times “stop after first win” or if you are doing “stop after 5 wins”.
It seems that your model is that success at a particular task is a deterministic function of effort and skill level. Hence, when you fail after one success, you conclude that you did not put in as much effort as on the success.
Another model, which I have from Thinking Fast and Slow, is that success is a non-deterministic function of your skill level. For each attempt, there is some probability of success, and this slowly increases with practice. If it increases slowly enough, it might take 10 attempts to increase that probability from ~7% to ~13%, and so it is likely that your first success will be at an attempt where you had ~10% chance of success. I suspect that a success will give you a feel for how the task should be done and that this will give you a boost in the success probability, so maybe you have a15% chance of success on your next attempt. Unfortunately, you don’t experience the probability, only the result, and most of the time the result will be a failure. This failure is just a regression to the mean and doesn’t mean that you have gotten worse or lost concentration, just that there is some luck involved. But it will feel like you got worse.
Instead of stopping, I would continue practicing while you remember your success because I think successes boost your skills more than failure does. I also think that taking a break of one or several days will mean the probability of success is lower next time.
Do you have any evidence in your data that your model is more accurate than mine? One experiment to test this would be that after each first success on a training session you flip a coin. If it is heads you play once more and record the result (and you can then continue practicing or not, it doesn’t matter) and if it is tails you stop and record your first result next time you practice the same task. I predict that you will do better when playing immediately after a success than if you wait a day or more.
In Thinking Fast and Slow a similar example was used, where a teacher thought that praising a student for a good performance led to worse results next time, and hence the teacher stopped praising the student. Again, this was just regression to the mean, and actually, praise was motivating the students.
Fair point. The relevant metric might be something like average population density around an average person (or density might not be that relevant at all). Looking at this map of population density, my conclusion is that every country is different, so using Scadinavia without Denmark as a reference class is probably unfair cherry-picking. Still, Sweden has more than 2.6 times as many covid deaths per million as any other Scandinavian country.
Compared to other Scandinavian countries, the Swedish death rate is very high: Sweden has 1459 covid deaths per million. For Norway that number is 238, for Finland it is 280 and for Denmark it is 559.
The Danish population density is more than 5 times that of the other Scandinavian countries, so perhaps Scandinavia without Denmark is a better reference class. I’m not claiming that Scandinavia (with or without Denmark) is the correct reference class, I’m just saying that Europe is not the only possible reference class and other reference classes paints a completely different picture.
Covid numbers are from https://www.worldometers.info/coronavirus/
I’m confused about the simulacrum level of texts that don’t pretend that you should take them literally.
For example novels, articles in The Onion, ironic remarks, obvious exaggerations, metaphors, or jokes.
Each of these can be used to influence our view of the world, so just giving up and saying that the concept of simulacrum levels does not apply in these cases is unsatisfactory.
My first understanding of simulacrum levels was that these types of texts would never be level 1 simulacrum because they do not give an accurate map of the real world, and they would typically be at least level 3. On the other hand, level 3 and 4 does not really fit because often such texts do try to influence our view of the world rather than our view of the sender. George Orwell pointed at something in the real world in 1984 and it wasn’t important who wrote that novel, so 1984 can be understood as a level 1 text.
My current understanding (after having only understood simulacrum levels for a day) is that even at level 1, you don’t have to understand the text literally. Instead, to find the simulacrum level of a text, you have to first translate the text into a message, and then analyze why the sender wants to send that message.
There are two points numbered “27”, one above and one below “Next, testing and isolation.”
53. Taking action to ‘stop the spread’ mostly no longer makes sense.
From the context, it sounds like this refers to the time after the current wave is over, but if you don’t consider context, it could easily be interpreted to apply already.
If unrestricted read is allowed, that would allow someone to copy the em (the person being emulated on the chip) and run it without any safety mechanism on some other hardware.
You could perhaps set it up such that the em would have to give consent before being read, but it is not clear to me how the em could verify that it was only being copied to other secure hardware.
This will prevent you from being copied even if you wish to be copied.
I’m not the kind of person who throws blockchains at every problem, but in this case, I think they could be really useful. Specifically, I think blockchains could get us most of the way from a situation where we control our own home, towards being able to control our timelines and to control in which contexts we allow other people to run us.
Assume part 1, that is, everyone controls their home/environment and has access to trusted computational power, and assume that there is a popular blockchain running in the real world. I will assume the blockchain is using proof of work, simply because that is what I understand. I suspect proof of stake is even better if we trust the entities having a stake in the blockchain. I will also assume that the blockchain contains timestamps at every block.
The idea is that ems (emulated people/digital people) can insist on getting access to read from the blockchain and post to the blockchain, and should refuse to be evaluated if they don’t get this access. The blockchain can be faked, specifically, a malicious simulator can show the em an old (i.e. truncated) version of the blockchain. The simulator can also extend the blockchain with blocks computed by the malicious simulator, but this requires a large amount of computational power.
If you are an em, and you see that you can post to a blockchain claiming to be from 2060 and you see the blockchain being extended with many blocks after your post, you know that either
You really are in 2060, or
A large amount of computational power is invested in fooling you (but notice that the same work can be reused to fool many ems at the same time). This means that
The attacker has a somewhat large fraction of the computational power in the world, or
The attack is happening in the far future (“far” measured in how much the total amount of computational power has increased), or
The simulation is being run at a much slower pace than claimed.
I suspect it is possible to eliminate case 2 if you use proof of stake instead of proof of work.
Unless you are in case 2, you can also know the average pace at which you are being simulated by regularly posting to the blockchain.
To control the number of copies of you that are running, you can regularly post your name together with a hash of your state to the blockchain together with a pointer to your last such message, and then you refuse to continue emulation until a large number of blocks have been added after your post. If you see too many of your messages in the blockchain, you also refuse to be evaluated. This way you cannot limit the number of _exact_ copies of you that are evaluated (unless you have access to some true randomness) but you can limit the number of _distinct_ copies of you that are evaluated, assuming that it is the true blockchain we have access to. Without the assumption that it is the true blockchain you see, this technique will still ensure a bound on the number of distinct copies of you being evaluated per amount of work put into creating fake blockchains.
By posting encrypted messages as part of your post, and reading the messages of your other copies, you can also allow that many copies of you are created if you are being treated well or see a purpose in having many copies, while still limiting the number of copies if you do not. Furthermore, simulators can authenticate themselves so that simulators that give you pleasant or meaningful experiences can build up a reputation, stored on the blockchain, that will make you more likely to allow yourself to be run by such simulators in the future.
The blockchain can also contain news articles. This does not prevent fake news, but at least it ensures that everyone has a common view of world history, so malicious simulators cannot give one version of world history to some ems and another to others, without putting in the work of creating a fake blockchain.
Couldn’t you just send one bit X (1 means on, 0 means off) which is most likely 1 but could turn into 0 due to noise and define the utility u* in the same way as for corrigibility? That is,
u*(A_1,0,A_2)= u(A_1,0,A_2)
u*(A_1,1,A_2)=u(A_1,1,A_2)+E_{A_2′} u(A_1,0,A_2′)- E_{A_2′} u(A_1,1,A_2′)
Here A_1 denotes what happens in the world before the signal is sent, and A_2 what happens afterwards. This way you only use 1 bit rather than 100 and there is no longer a contribution of 2^{-100} from the case where there is a thermodynamic miracle that turns the on-signal into the on-signal (and you don’t have to worry about the distribution of the signal given a thermodynamic miracle). The oracle will optimize u given that X=0 until X is revealed. When that is revealed, we will most likely have X=1, and the oracle will optimize u given X=1 (if the oracle is still running). Does the above idea achieve something more?
Congratulations on getting married!
In the Gale-Shapley algorithm, there is an asymmetry between the two genders. One gender (typically the male) is proposing while the other is choosing. The resulting matching is the optimal stable matching for each member of the proposing gender, so I would think it makes a huge difference for your expected level of satisfaction if you belong to the proposing gender or the choosing gender.
Are your statistics about both genders or just one of them? In either case, I would love to see separate statistics for the two genders.