This is so overcomplicated. You don't need an actual tree that's just a mental model. All you need to do is generate N random hyperplanes (defined via it's normal i.e. generate N normalized vectors in the dimension of your choosing). Then you generate a hash via these by going normal by normal and checking if your query vector `q` is on the left or right side of the normal by checking the sign of the dot product. I.e. `hash = [sign(dot(q, n1)), sign(dot(q, n2)), ...]`.
So given a query you find that hash and you get all the vectors with the same hash (using your favorite hash to array data structure, i.e. a dict) and you can further refine the search within that bucket. It's like 10 rows of code.
Not really irreversible unless they get a monopoly on violence, so as long as the overton window from the publics perspective shifts dominantly towards breaking them up in the future we'll be able to get a better situation.
I initially blocked that popup with ublock origin, but I've since turned it back off and let it block me instead. It's a fantastic way to prevent you from mindlessly watching videos. Now I just use yt-dlp to download videos I know I care about and subscribe to.
Good article but my god why is there like a 2 second delay before every user interaction on this page. Scrolling, selecting text, everything has some kind of 2 second 'bootup' time after which things work normally but if you stop interacting with it for a bit it goes back to some idle mode. Really weird.
I think it's just them being scared of models running on their servers producing things that cause bad PR. And the reason they're not willing to not run things on their servers is competitive advantage as well as just extracting more money out of users that way. (I definitely don't think actual safety is the concern.)
Well they could just not allow prompts that seem to participate in blogspam. If they wanted to stop it they definitely could.
Their argument is that since it's centralized, things like that are possible (while with llama2 you can't), they do "patch" things all the time. But since blobspam are contributing to paying back the billions microsoft expects they're not going to.
It's funny that one argument openai used to keep their models closed and centralized is so they could prevent things like this. And yet they're doing basically nothing to stop it (and letting the web deteriorate) now that profit has come into play.
This is all making me realize investors are just people with money (not these super intelligent market predictors that they're often portrayed). You'd do the same to market a product as you would to skim off the flood of money going into a hyped field I suppose.
They're amusing but they're definitely dangerous to the potential of AI in the medium term. People like to say that "these things will happen anyway" even if it's made illegal, but there's no way 250 million will be invested into a company to develop LLMs if they're made illegal or heavily regulated. ML is resource intensive as it is (in terms of number of PhDs required and compute) and if the decels and censors have their way they definitely can make a dent in progress (i.e. "it'll happen anyway" isn't true, or it'll be slowed down so much nothing interesting will happen within our lifetimes).
These things won't be 'all knowing': things that are kept secret by the government like how to make nuclear weapons won't be known by it, nor can you ask it what your coworker thinks of you and have it accurately tell the answer. They are however great reasoning and creative engines. I look forward to being able to boost that part of my workflow.
So given a query you find that hash and you get all the vectors with the same hash (using your favorite hash to array data structure, i.e. a dict) and you can further refine the search within that bucket. It's like 10 rows of code.