Good on you for seeing a problem and making something to solve it!
That said, I'm a bit confused about the use case. Wouldn't it be simpler and more secure to have all the encryption occur on the client side, and have the server be a dumb encrypted blob store?
Put another way, I think OpenADP tries to solve the problem "I don't trust hosting providers in any single sovereign nation" by splitting the trust between multiple nations; whereas it seems like it would be even better not to trust any of them.
Forth supports direct memory access; you can request arbitrarily sized chunks of heap memory and retain pointers either on the stack or aliased by dictionary words.
Here is one - there are finitely many mathematical symbols (or at least, all mathematical symbols can be defined in terms of a finite core of symbols).
That means the set of all mathematical definitions is countable (i.e. you could assign a whole number to each one, putting them into an infinitely long ordered list).
However, the set of real numbers is uncountable (by Cantor's argument).
Therefore the vast majority of numbers ("almost all" numbers, in a mathematical sense) cannot be defined, even in principle.
I think you misunderstood OP- I think he's saying that there was propaganda across the board, including both treatment of people like Gallileo, and also including techniques such as the iron maiden.
I take his point to be precisely that the iron maiden wasn't really used.
Would it be fair to say they signal boosted her personal details on social media?
Irrespective of whether it is literally doxxing (I think you're right that it isn't), it feels irresponsible in terms of the effects that might be anticipated from doing so.
They are not strictly speaking illegal in the UK, and were pretty common in the 2000s and early 2010s.
The law requires employers to pay minimum wage to anyone who is a worker (i.e. is doing work for the company) - but "sitting in" doesn't have to be paid. Exceptions are also made for volunteering, and placements as part of a university course (these can be unpaid).
Enforcement isn't consistently applied so some smaller outfits get away with it. The big FAANGs and other major companies pay.
Wow! I think I dimly intuited your first paragraph already; I directionally get why your second might be true (although I'd have thought L1 was even more so, since it encourages zeros which is kind of like choosing a subspace).
Your third paragraph took me ages to get an intuition for - is the idea that regularisation penalises having "sharp elbows" at the join points of your hyper-spline thing? That's mind blowing and such an interesting way to think about what a ReLU layer is doing.
Thanks so much for a thought provoking comment, that's incredibly cool.
You're correct, and the term you're looking for is "regularisation".
There are two common ways of doing this:
* L1 or L2 regularisation: penalises models whose weight matrices are complex (in the sense of having lots of large elements)
* Dropout: train on random subsets of the neurons to force the model to rely on simple representations that are distributed robustly across its weights
I think there's a sense in which moderns feel that the reals (or at least the rationals) are a natural category: that 5 and 0.3 are the same "kind of thing". Mathematicians talk about the distinction between different classes of numbers, but to most people they're all numbers.
Do you think that sense was shared by the ancients, or do you think that the linguistic distinction mirrored a stronger mental separation between the two? It sounds like it might have done if they had to do the work of duplicating proofs from one to the other. Did they have a single word to describe the shared concept?
I'm no fan of this act but your characterisation is highly misleading.
To pick two examples from the document you linked:
Discussion of being a sex worker would not be covered. The only illegal content relating to sex work would be if you were actively soliciting or pimping. From the document:
* Causing or inciting prostitution for gain offence
* Controlling a prostitute for gain offence
Similarly, discussion of drug use wouldn't be illegal either per se, only using the forum to buy or sell drugs or to actively encourage others to use drugs:
* The unlawful supply, offer to supply, of controlled drugs
* The unlawful supply, or offer to supply, of articles for administering or preparing controlled drugs
* The supply, or offer to supply, of psychoactive substances
* Inciting any offence under the Misuse of Drugs Act 1971
That's very different to criminalising content where you talk about being (or visiting) a prostitute, or mention past or current drug use. Those things would all still be legal content.
The capital W is conventional - when talking about a place, a language, or a grouping of people, it's grammatically correct to capitalise the inital letter.
"I visited a few of America's major cities, and was especially struck by how many New Yorkers speak English as a second language. "
So it's correct to capitalise "Meanwhile, Western countries might [...]" but not "There's a brisk westerly wind coming in over the sea".
The linked article precisely claims that the have overhyped their results by (inadvertently? ) having their test distribution overlap with their training data, and implies that the result won't generalise to novel proteins.
I don't have a specific link, but the general method is called "image streaming". It's pretty googleable so you should be able to find something similar at least.
I remember hearing that Beam Search doesn't work well for LLMs, because it leads to repetitive, generic output.
The majority vote sampling technique in this paper sounds like it'd give similar output to Beam Search, because it's sampling sequences of tokens from a joint distribution. So why doesn't it give repetitive output like Beam Search does? What am I missing?
I think it does come down to semantics. When you say "weights", people will take you to mean the pre-trained parameters of the network.
I agree that in some sense the attention weights are more like meta-weights that are applied to the context of the conversation to decide how to actually weight the various words. So it's totally correct to say that previous words in the conversation affect how future words will be weighted, and I think it's reasonable to call that 'learning': for example, you can tell ChatGPT new words and it will be able to use them in context. Again though, people usually take 'learning' to mean making updates to the trained parameters of the model itself, which obviously isn't happening here.
If I'm going to ask the player to do something multi stage I will usually flag it so they know that there's a few places where things might go wrong. "OK well sneaking up to it is going to be the first hurdle; after that you're going to need to figure out how to get the giant lizard to cooperate as you smuggle it out the back".
That lets the party start thinking ahead a bit, so maybe the illusionist says "how about I conjure the smell of something delicious wafting in from out a back window – that way when the rogue manages to sneak up and untie it it'll hopefully pop right out the window by itself".
Whereas if after the stealth check succeeds you give a bunch of subsequent unexpected checks the player might feel a bit aggrieved that you hadn't signalled how much work it'd be (especially if it uses skills they aren't good at and would have avoided relying on).
Christianity has a high degree of continuity with Judaism, but rejects important aspects of Jewish theology, such as Kosher and Sabbath-keeping. So while Christianity accepts large amounts of Jewish theology, and extends new ideas, it doesn't universally accept everything that came before it.
Similarly, Islam accepts large amounts of Christian theology (and indeed reintroduces some Jewish practice) and brings new beliefs with it, but doesn't bring every tenet along.
Admittedly Jesus' divinity is a very big part of Christianity to reject; but then, the idea of the trinity would be an equally big pill to swallow for a Jew ("hear, O Israel: the LORD your God is one").
That said, I'm a bit confused about the use case. Wouldn't it be simpler and more secure to have all the encryption occur on the client side, and have the server be a dumb encrypted blob store?
Put another way, I think OpenADP tries to solve the problem "I don't trust hosting providers in any single sovereign nation" by splitting the trust between multiple nations; whereas it seems like it would be even better not to trust any of them.