It makes more sense if Anthropic is assuming that most flagged conversations are false positives (but it wants to keep Mythos away from the true positives).
I agree with you that restricting access to Fable is stupid, but I'm in favor of e.g. GPU export controls. It's certainly annoying, but—well, I don't know where you live, but you don't want to make it easier for China and Russia to build weapons they can use attack to attack Taiwan and Ukraine, right?
And the nice thing about the GPU restrictions is that even if they don't work completely, just making the hardware more difficult and expensive to access is useful.
I, like, can't even imagine how they're dealing with this internally. Presumably Anthropic is using Mythos to train future models, but now many of their researchers can't use what they themselves created? Do they just need to stop work now?
> We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology. [1]
First, I want to state up top that I don't personally agree with any legislation that seeks to ban or limit access to frontier llms.
Having said that, it's difficult to not have a somewhat glib reaction to this news. Surely, if leadership believes what is written above then they are happy with this outcome. Ignoring all perverse incentives or potential market manipulation that might be going on from this current administration, on it's face it's a genies wish granted, "your models are so dangerous that the world should pause development? done. You may start pausing development."
That is literally not at all what this regulation does. This regulation does not pause development. This is designed to make sure non-US citizens cannot do work with/on this model. This is stupid because Anthropic never once said this is what they wanted. They said there should be a global effort undertaken for everyone to take a coordinated approach to slowing down development. They never said "please please make it impossible for Anthropic to develop models while letting everybody else develop what they want"
I feel like every time I try to use this feature, I get burned because I accidentally use the back touchpad gesture, or I accidentally close the tab, etc. Or, I remember one time I was using the Claude iPhone app, and I quickly switched apps to respond to a text message. My phone must have been low on memory or something, because as soon as I switched back to Claude, the app faded to the startup screen and the whole conversation was gone.
Which of course is the entire point, so I don’t really know what I want here, but what they have right now isn’t that usable IMO. Maybe a chat that lasts 24 hours? And, it would be nice if there was a way to convert an incognito chat into a persistent chat.
> Out of curiosity, I sic'd Claude Fable on the original French version of The Three Musketeers and told it to translate accurately, but also try and keep the same jovial tone as the original and do not censor anything. After it was done, I didn't read the entire output, but I did compare a few individual chapters between the Ellsworth translation and the Fable translation.
This isn’t a great test, because Claude almost certainly has multiple translations of The Three Musketeers in its training data.
The things is, this is almost certainly what's happening.
You can (could, maybe they 'fixed' it by now) get sota LLMs to reproduce entire novels near verbatim.
The idea of giving it parallel texts of those novels in different languages, to train it on translation, is so obvious it'd just be strange if the AI labs didn't do it.
In fact DeepL was doing basically that more than 10 y ago.
Oops, I legitimately missed the second-to-last paragraph.
I still think there are better tests you could do. Ideally, you would choose a book that was published recently—after the model’s cut-off date—which is considered to be a good translation. But even something like The Girl With the Dragon Tattoo, which is not particularly new and by no means obscure, would be better than a famous work of literature like The Three Musketeers that has many translations.
Almost certainly correct, though I've noticed that these LLMs like to complain when you give it stuff that is still in copyright. The Three Musketeers is thoroughly public domain everywhere so in that sense it's a good test, but of course because it's public domain everywhere there are lots of translations to crib from so I acknowledge it's not a great test because the training data almost certainly contains a competent translation.
Even if Fable didn't have Ellsworth's translation, it certainly has the William Barrow translation, which would still get it like 80+% of the way there.
My wife speaks Spanish, I should get her to do some kind of comparison with a Spanish book that doesn't have English translations.
I remember how, if you wanted to swap the top and bottom screens, it would spin and spin and spin… definitely a painful experience.
I actually think it was worth the money though, at least for me, because having a pocketable device that could access the internet was so special at that time.
Also, the DSi’s web browser was legitimately good. In addition to being fast enough, the zoomed out bottom screen and zoomed in top screen was great for browsing designed-for-a-desktop websites.
If only the Wii had supported 1080p or even 720p output, it could have been a great way to browse the web in your living room. Many companies have tried this, but none of them had the correct input device. I think this could have changed the overall direction of technology.
Unfortunately, at 640x480 everything was too small, even back then.
It’s definitely usable, it’s just not a first-class experience. Like the difference between browsing a non-responsive website on an iPhone versus a desktop.
It's still a wise idea to take it seriously.
Many years ago, a then 17 y/o old ( German ) friend of mine did mitm on his own console in order to replace a video stream in ps home, recorded it for youtube/laughs.
Not much later he was SWATed by the german police as if he was some kind of terrorist. His parents got a heart attack. Next he received financial death threats from a ultra expensive UK based law firm send by Sony.
I feel like you're going to run up against the definitions of "efficient" and "compression".
For example, a language with a larger alphabet will be able to express more in fewer characters. Is that more efficient?
Similarly, you could think of each word as a sort of lookup table for information in the mind of the reader. We don't define words as we're writing, we expect the speaker to know them already. If a language has more words, each word is more precise, and fewer words can be used to express an idea—but is that efficiency? You're just relying on the reader having more preexisting knowledge.
> a language with a larger alphabet will be able to express more
> in fewer characters.
True, although it's not really the alphabet that determines this, it's the number of phonemes (distinctive sounds) in the language. For example, writing /s/ (the sound) sometimes with 's' and sometimes with 'c' does nothing to shorten words in English or Spanish.
But in general, languages with fewer phonemes tend to have longer words (and tone languages often have very short words---in a sense, they have more phonemes than non-tone languages). Morphology (particularly compounding) often obscures this.
reply