Personally I've downgraded my phone intentionally, from an iPhone 14 Pro to an iPhone SE 2020 (the previous iPhone I had). It's slow, it overheats, the battery is short. But that's what I want -- the friction makes me really consider every use of the phone, and the occasional slowdowns frustrate me just enough to reconsider taking it out of my pocket for everything. It kind of "works" for me. Might get a battery pack though hahah
Yeah, I guess "creative writing" in this case is a shortcut for essentially saying they aren't that good at conforming to user-specified constraints. They can generate "creative" texts but can't necessarily constrain and iterate on their output in a conversational setting effectively.
Even if UBI is funded by taxation, rather than deficit spending, isn't it possible that demand will increase for some products (e.g. basic necessities that UBI recipients will suddenly be able to spend more on), causing inflationary pressure on the prices of those products?
I mean is it determined by money creation, or just influenced by/one of the factors? If business perceive people generally having more spending money, won't they raise prices accordingly?
I believe the deflection is the high-level explanation. Things like the Bernoulli effect and the air on the top of the airfoil travelling faster (it does -- far faster than the equal transit time theory implies actually), are the "instantiation" or outcomes of the air deflection. This is my understanding. Hence airplanes can fly upside down because even if the airfoil is upside down, it's still deflecting the air, just perhaps less efficiently (I think it's true that planes flying upside down need a more extreme angle of attack to maintain lift, so this makes sense)
This is the point everyone mostly misses -- internally inside the model, there is no difference between generating something that happens to correspond to the truth, and a "lie". It's all the same statistical generation, and all efforts in pushing AI right now are trying to increase the frequency with which the generations happen to correspond to reality.
There needs to be a way to opt out of all AI features in apps, yesterday. Or even better: opting out of all algorithmic/recommendation/infinite feed content. I think it's time we acknowledge that it is a net negative on most people's lives.
So here's a modern Greek perspective to add to the choir. In modern Greek zero is indeed plural, you would say 0 βαθμοί for example to say 0 degrees. However, in contrast to English:
1.5 books
ενάμιση βιβλίο -- singular!
You are functionally saying something closer to "a book and a half" in english I guess! Actually this is an interesting duality in english: 1.5 books but a book and a half. Guess it depends on how "separate" linguistically the numbers are: is it one book and a half book or is it a single quantity? Greek for decimals between 1 and 2 picks the "two quantities" approach.
Maybe I'm too cynical, but too many people in power directly benefit fom enshittification for anything about it to change. Even just the problem of fixing the housing market while the majority of politicians own several properties is an example of this. There's zero incentive for anything to change.
Very exciting, although it was a bit disappointing to see that they're hitting just llama1 7b performance by quantizing llama3. but i'm sure the performance gap will close over time!
Unfortunately not, as it captures such a wide spectrum of use cases and scenarios. There are some benchmarks to measure this quality in specific settings, e.g. summarization, but AFAIK nothing general.
One of the main disappointments with Duolingo's recent path for me has been the removal of the forum. It was so useful whenever I was confused by a certain phrase or new concept to just hop directly into the discussion forum page for that exercise and see others asking the same questions and getting well-written intuitive answers to them.
Just boosted my understanding so much plus added a sense of community, that we're all going through this learning journey together and thinking about and getting confused by the same things :P Now I'm relegated to Googling the text of the exercise to find similar info online, just far less useful :(
It seems like it's matching PyTorch's API very closely, which is great. Part of me wishes they took it a step further and just made it completely API-compatible, such that code written for PyTorch could run out-of-the-box with MLX, that would be killer.
Yeah that would save disk space! In terms of inference, you'd still need to hold multiple models in memory though, and I don't think we're that close to that (yet) on personal devices. You could imagine a system that dynamically unloads and reloads the models as you need them in this process, but that unloading and reloading would be pretty slow probably.
My main gripe is trying to run unsigned programs, and having to go into System Preferences every time to approve the first execution. It definitely feels a little ridiculous.
Reminds me a little bit of PICO-8 conceptually. At first glance I thought it was just a set of premade assets more than an actual game engine, but looking at the API more closely it actually does seem to resemble quite a bit the set of primitives that the PICO-8 console provides you, although without the pretty in-console IDE.
But since there isn't really treatment, I don't even know if I'd want to know personally, in advance, if it's not something I can take any steps to prevent.
OK, thanks for letting me know! It's a bit tricky to upload ArXiv links, I have a sense that not a lot of people will actually click through to read the paper or that some people just won't bother to click an ArXiv link at all, so I think that was my reasoning for sticking the abstract in the comments. but you're right that it's a bit hard to read. thanks!
Paper abstract: We attribute grokking, the phenomenon where generalization is much delayed after memorization, to compression. To do so, we define linear mapping number (LMN) to measure network complexity, which is a generalized version of linear region number for ReLU networks. LMN can nicely characterize neural network compression before generalization. Although the L2 norm has been a popular choice for characterizing model complexity, we argue in favor of LMN for a number of reasons: (1) LMN can be naturally interpreted as information/computation, while L2 cannot. (2) In the compression phase, LMN has linear relations with test losses, while L2 is correlated with test losses in a complicated nonlinear way. (3) LMN also reveals an intriguing phenomenon of the XOR network switching between two generalization solutions, while L2 does not. Besides explaining grokking, we argue that LMN is a promising candidate as the neural network version of the Kolmogorov complexity since it explicitly considers local or conditioned linear computations aligned with the nature of modern artificial neural networks.