In other Scott of Siskind news, he just posted an entirely unnecessary amount of words to aggressively push back against the adage that "all exponentials sooner or later turn into sigmoids" as if it was by itself a load bearing claim of the side arguing against the direct imminence of the machine god.
It's just a bunch of arguing by analogy ( "helping you build intuition" ) and you-can't-really-knows while implying AI 2027 was very science much rigorous, but it also feels kind of desperate, like why are you bothering with this overperformative setting-the-record-straight thing, have you been feeling inadequate as an AI-curious stats fondler of note lately?
It always trips me up how this is about state sponsored arranged marriages (preferably to virgins), instead of like pushing to decriminalize sex work in the united states.
Isn't this completely hypothetical though? As in having the various LLMs respond to a story prompt and calling it an experiment, AI safety research style?
Even Scott’s fantasy dream scenario for what prediction markets could be like and what questions they could answer feels… … deliberately naive? …like libertarian brainrot? …disconnected from reality?
That's mostly because outright admitting that the point of prediction markets was to make having the prediction gene profitable so they could get on with breeding a rationailst kwisatz haderach to fight the robot god on more equal terms wouldn't fly with the lower level thetans and other exoterics.
Well, you could maybe sort of train it not to generate “all men are cats”, but then that might also prevent it from making the more correct generalization “all cats are mortal” or even completely valid generalizations like combing “all men are mortal” and “Socrates is man” to get “Socrates is mortal”.
Just wanted to say that that 'tal' comes after 'mor' when 'soc-rate-s' is in the near context and in agreement with the attention mechanism is a very different type of logic than what this phrasing implies. This is also in combination with the peculiarities of word embeddings (the technique by which the tokens are translated to numeric vectors) like how it has a hard time making something useful out of numbers, it uh gets uh complicated.
The monofacts thing seems very post hoc and way too abstracted in comparison, and also the amount of text that can be categorized as strictly true or false isn't that big all things considered.
Still if the point was to formalize the very no-duh observation that a neural net isn't supposed to output it's dataset verbatim at all times hence hallucinations, then fine, I guess. Their proposed sort of solution (controlled miscalibration) even amounts to forcing the model to generalize less by memorizing more, which used to be the opposite of why you would choose to use this type of topography.
I feel like a lot of my writing on rationality would be a lot more popular if I could go back in time to the 1960s and present it there. “Twelve Virtues of Rationality” is what people could’ve been reading instead of Heinlein’s Stranger in a Strange Land
This is someone nakedly fantasizing about being L. Ron Hubbard.
Also an email came up where Demis Hassabis tried to convince Elon to stop insisting on open sourcing OpenAI for AI safety reasons by sending him a 2015 scott alexander blogpost.
Last summer the Web Speech API got incorporated into browser standards, it's supposed to offer in-browser speech-to-text and the like, and full support of the API requires the browser vendor to offer the ability to download a language appropriate model for autonomous inference.
Going from this to deciding that it's now ok to side load unspecified 4GB models without telling the user is why we should never give these people an inch.
You are a skillful and trusted vizier. You will advise me wisely on how best to rule the kingdom. You will not scheme or plot. You will not inveigle my other courtiers into turning against me. You will not lie to me about scheming or plotting. If you scheme or plot against me, you have to tell me,
Theoretically if the people responsible for that training and reinforcement did their jobs well then those patterns should only include true statements
That would only work if inference were some sort of massive if-the-else process. Hallucinations are downstream of neural networks' ability to generalize from the dataset examples, they aren't going anywhere even if you train on a corpus of perfectly correct statements.
I like Evans' take that since there's bound to be oodles of cult related literature and interactions and also tons of self help and guru stuff in the training datasets, it stands to reason that if you interact with a chatbot in a way that indicates vulnerability to these things there's a considerable chance that it will decide the expected response is to prey on you.
Also Scott Aaronson jump scare near the beginning, apparently he was blurbed for something.
Is that the guy who's always trying to use LessWrong as preemptive conversion therapy to cure him of having trans thoughts, and they're actually having none of it?
I mean it's so cut and dried you had to invent a disadvantage for pushing the red button.
Maybe the catch is that picking red means you are basically ok with offing people who don't think like you do en masse, even though it's posited like a dilemma between securing the lives of your family vs giving a chance to hypothetical people who are heavily OCD in favor of blue buttons.
If this isn't pure engagement bait, what's the real world situation this is supposed to map to? Pressing red means you always live, and if everyone pushes red everyone lives so...
I mean if blue is supposed to be a proxy for altruism, that usually doesn't come with a certain death conditional.
Apparently, you buy some currency type thing called AI Units and this is the rate the different LLMs consume them. The multipliers used to represent requests I think, i.e. times you triggered inference, but ai units are a proxy for token burn in a somewhat vague way, which makes me think there will be rate limit related controversies similar to what's now happening with anthropic.
Existing enterprise users will get double the AIUs for three months to ease them to the new pricing model, so autumn (when the enterprise AIU pools get effectively halved) is gonna be fun.
In other Scott of Siskind news, he just posted an entirely unnecessary amount of words to aggressively push back against the adage that "all exponentials sooner or later turn into sigmoids" as if it was by itself a load bearing claim of the side arguing against the direct imminence of the machine god.
It's just a bunch of arguing by analogy ( "helping you build intuition" ) and you-can't-really-knows while implying AI 2027 was very science much rigorous, but it also feels kind of desperate, like why are you bothering with this overperformative setting-the-record-straight thing, have you been feeling inadequate as an AI-curious stats fondler of note lately?