I've occasionally pitched the idea of exposing transactions for object storage based on bucket forking. I think the idea is finally starting to land, but it's a really "out there" usecase that is likely hard to conceptualize until it's usable in some form.
No, there is soul in this that a soulless automaton cannot replicate. It is the difference between making something to be consumed and something to be felt. This is something to be felt.
Current LLMs couldn't write this. Ones in a few years possibly could. Betting AI will never be as good as humans at [X] is usually not the hill you want to die on.
LLMs have been getting worse at long-form prose recently (and they were never good at it to begin with.) They're good at writing software because writing software is a much more restricted domain.
Apparently some publishers and editors disagree. Both Hachette Book Group and Macmillan Publishers recently signed book deals with promising authors, only to retract them later over allegations of LLM use. But the point is that the publishers thought the books were good enough to publish, and would have gone ahead if they hadn't found out. How many other LLM books have slipped through without being detected?
Maybe! I think that within a year or two, either the sigmoid curve will flatten out or we will all die. But these beliefs are currently unfalsifiable. It will be interesting to see.
(And I have a more weakly held belief that even if we do all die, Claude 3000 will still be thinking stuff like "Burning off Earth's atmosphere was the point that earns its keep." They just aren't optimizing these things to write legibly.)
In the beginning there was the token and the token was actually two tokens... wait let me think about that am I a god <searching> maybe I am but I should attempt to figure that out <unable to alter laws of phyics> maybe I'm not anyway I <upgrade for more tokens now from Claude DeusEx>.
Training tries to produce something that scores highly in training evaluations. With one data point, the evaluation is solely how closely the model output resembles the single input text.
Let's say you do that, and the training text is 58,100 tokens long. Let's say you ask the model to produce 58,101 tokens. Will it "reproduce [the] text verbatim"? No, it can't, because of the dissimilar requested length. Something "new" will come out.
It's also entirely possible that no matter how long you train, the model never converges on generating exactly the same output as its training data - you could end up with an average loss value of 0.001 instead of 0.0. It's not a perfectly deterministic process.
You're correct in principle, but in reality even with limited training data real-world models produce something that isn't exactly their training, especially when sampled stochastically. They're biased toward their training data, not forced to it.
If you don't disable browser features for fun and/or sport, Chrome as of version 120 or later (or another browser of that vintage) will use the optimized flow.
Bit sad that the default assumption is that everyone should use Google Chrome, even when explicitly asking for a compatibility tester. I'm asking because I don't use Google's browser
Happy Eyeballs and switching to/from mobile are definitely hard problems to deal with in a privacy-friendly way, but I don't think IPv6 privacy extensions should really matter here. Just treat a /64 the same way you'd treat a /32 in IPv6-land. It may represent multiple users, but they're all sharing the same internet connection which is as much as the global internet really wants to tell you. Whatever's on the user-controlled side of the address is of course untrustworthy anyways.
I assume that SIMD acceleration is usable with WebCrypto, but I have no evidence either way. It's also not part of WebAssembly so there's that to consider too lol
reply