windy.com with a VPN in a private browser window. They can’t track you if they don’t know where you are!

Can you explain please, where I made a mistake?

Your mistake is thinking Earth is 6km in radius! :D 6km is how far you walk in an hour. Either you think Earth 1000 times too small or kilometer 1000 times too big.

For an object heavier than the Earth, 1g radius will be greater than the radius of Earth. For 56 Earth masses that’s sqrt(56) times bigger = 48000km.

A 56 Earth mass black hole will take 5.5e55 years to evaporate according to this calculator. A 100kg black hole (more close to what Richard used to be) is much smaller than the nucleus of an atom and will evaporate in 0.05 nanoseconds.

Curiously there was a paper recently that calculated that even if there was a small black hole in the center of the Sun, it would take millions of years for it to grow, because the aperture is so small not much can fit through, and the infalling gas heats up so much as to repel the rest, creating an internal hot bubble.

Some notes for my use. As I understand it, there are 3 layers of “AI” involved:

The 1st is a “transformer”, a type of neural network invented in 2017, which led to the greatly successful “generative pre-trained transformers” of recent years like GPT-4 and ChatGPT. The one used here is a toy model, with only a single hidden layer (“MLP” = “multilayer perceptron”) of 512 nodes (also referred to as “neurons” or “dimensionality”). The model is trained on the dataset called “Pile”, a collection of 886GB text from all kinds of sources. The dataset is “tokenized” (pre-processed) into 100 billion tokens by converting words or word fragments into numbers for easier calculation. You can see an example of what the text data looks like here. The transformer learns from this data.

In the paper, the researchers do cajole the transformer into generating text to help understand its workings. I am not quite sure yet whether every transformer is automatically a generator, like ChatGPT, or whether it needs something extra done to it. I would have enjoyed to see more sample text that the toy model can generate! It looks surprisingly capable despite only having 512 nodes in the hidden layer. There is probably a way to download the model and execute it locally. Would it have been possible to add the generative model as a javascript toy to supplement the visualizer?

The main transformer they use is “model A”, and they also trained a twin transformer “model B” using same text but a different random initialization number, to see whether they would develop equivalent semantic features (they did).

The 2nd AI is an “autoencoder”, a different type of neural network which is good at converting data fed to it into a “more efficient representation”, like a lossy compressor/zip archiver, or maybe in this case a “decompressor” would be a more apt term. Encoding is also called “changing the dimensionality” of the data. The researchers trained/tuned the 2nd AI to decompose the AI models of the 1st kind into a number of semantic features in a way which both captures a good chunk of the model’s information content and also keeps the features sensible to humans. The target number of features is tunable anywhere from 512 (1-to-1) to 131072 (1-to-256). The number they found most useful in this case was 4096.

The 3rd AI is a “large language model” nicknamed Claude, similar to GPT-4, that they have developed for their own use at the Anthropic company. They’ve told it to annotate and interpret the features found by the 2nd AI. They had one researcher slowly annotate 412 features manually to compare. Claude did as well or better than the human, so they let it finish all the rest on its own. These are the descriptions the visualization shows in OP link.

Pretty cool how they use one AI to disassemble another AI and then use a 3rd AI to describe it in human terms!

I want people to be able to report bugs without any trouble.

Thank you for being aware! I’ve experienced this on github.com. I’ve tried to submit issues several times to open source projects, complete with proposed code to solve a bug, but github shadowbans my account 6 hours after creating it (because I use a VPN? a third-party email provider? do not provide a phone number? who knows). I can see the issue and pull request when logged in, but they only see a 404 on their project page even if I give them a direct link. I ended up sending them a screenshot of the issue page just to convince them this was even possible. Sad to hear gitlab does it even worse now by making phone mandatory.

Am I the only one for whom "open"subtitles.org hasn’t worked in years? I literally cannot find the download button, like in those okboomer memes. Never used the API. Switched to subscene.com and haven’t had problems since.

The joke is that the orbit was clearly originally reported in kilometers, but the article editor “helpfully” converted it to miles and reported it in miles as default, but it makes no sense now because the same “miles” number now equals two different “kilometers” numbers.

far more maddeningly, some sites tell me that my username and password combo are incorrect when I’m using a VPN

Brb, gonna wiretap the judge’s house. It’s not a crime as long as I don’t act on the information I hear so there is no injury.

*Yawn!* Wake me up when they stop requiring phone numbers to sign up.

The list of tor exit ips is publicly viewable. Some IPS block the entire list contrary to Tor Project’s request not to do exactly that.

Exactly! Trying to think outside the box in a trolley problem is like wishing you could wish for more wishes in a genie problem.

Excellent excellent!

If 6 is rolled, then P(X|R=6) = (N-1 choose 9)/(N choose 10)

Might as well reduce that to 10/N to make the rest of the lines easier to read.

If you don’t flip it, you have a 2/3 chance of dying.

There is also a chance that your switch is not connected and someone else has control of the real one. So there is an implicit assumption that everyone else is equally logical as you and equally selfish/altruistic as you, such that whatever logic you use to arrive at a decision, they must have arrived at the same decision.

No matter what your goal is, given the information you have, flipping the switch is always the better choice.

That is my conclusion too! I was surprised to learn though in the comment thread with @pancake that the decision may be different depending on the percentage of altruism in the population. E.g. if you are the only selfish one in an altruistic society, you’d benefit from deliberately not flipping the switch. Being a selfish one in a selfish society reduces to the prisoner’s dilemma.

there’s no way to know which track the trolley is on

It’s a standard trolley meme problem, the trolley will keep going on the main track unless the lever is switched 😁. I thought !science_memes would be familiar with trolley problems, but I guess I get to introduce some of you! You might want to start off on some easier trolley memes first, this is advanced level stuff.

where the real lever sends it

There is not usually ambiguity with the lever. If you wish, you can have an announcement in the headphones “main track… side track…” every time you flip the lever. Your only uncertainty is which track you yourself are bound to, given how you’re blindfolded.

there’s a 0.017% chance

1/6 * 10% = 1/60 = 0.01666… = 1.666…% ~= 1.7%! Careful there!

It’s not really a trolley problem, because in both scenarios a track is empty,

Everything is a trolley problem.

The trolley is like really slow.

You are on the side track in scenario A. You die if you pull. Ironically, you’d be killing yourself. The dice are to make the two scenarios not equally likely.

Half the fun of trolley problems is adapting them to puzzles for which they are utterly unsuitable:

It’s a standard runaway trolley problem. The trolley is traveling down the main track unless the switch is flipped to send it down the side track. The lever is labeled such that there is no ambiguity which way it is set, the blindfolds notwithstanding. The villain is pernicious and will be equally (though not exceedingly so) delighted to see you die by your own action where inaction would have had saved you. You can somehow trust that the announcement in the headphones is true and not a lie. Such as, for example, you have seen this exact situation happen many times before on TV and survivors/witnesses have described the villain to be truthful every time.

Good question to ask, since specifics of selection process may affect the decision outcome! Other variants include growing humans in a vat from scratch on demand, using Star Trek transporter clones, or abducting the necessary number of people from a pre-selected list where your name happens to be the first one. For now, imagine the potential population as the 5 billion living cognizant adults.

as X approaches infinity, the difference becomes negligible

It may be negligible to the 4.999… billion adults sleeping comfortably and securely in their beds tonight, but the problem presupposed that you have already been abducted. It remains underdefined whether you refers to you the specific person reading this meme, or a more general you-the-unfortunate who has been chosen and is now listening to this on the headphones.

Or spit, or blow air at your potential neighbor, or fart in their general direction!