Hacker News new | past | comments | ask | show | jobs | submit | amilios's comments login

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.

I'm cynical enough to basically agree. :-) I do think it may change, but that's what I mean about hitting bottom. It may have to get so bad that there is some kind of violent uprising or societal collapse/fracturing.

Just by the by, ironically, I've heard, from my tenuous connections into ideological spheres outside my own, that a decent number of people voted for Trump out of a similar desire to shake the foundations of the system. Of course they've likely been hoodwinked, but I think the opportunity is there for a Bernie Sanders-esque person on the other side to make some change by whipping up a frenzy of rage at the status quo and promising to get out the pitchforks. The question is whether such a frenzy can be accompanied by enough cool-headed calculation to be effective.


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.


Thanks, any ideas why it's not possible to build a generic eval for this? Since it's about asking a set of questions that's not public knowledge (or making stuff up) and check if the model says "I don't know"?


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 :(


Those forums were mainly a band-aid for DL's complete refusal to actually explain any of the grammar.


There's an overview for each unit section that has some pretty important explanations, and nothing really directs you towards it.

Still, in terms of forums, a lot of the time Googling will lead you to various subreddits that have some good conversations and explanations.


I'm in Unit 6 of Italian, 11 in Russian, and there have been zero of those. I know which section you're talking about, I check it every time I reach a new unit.

Shouldn't have to google. They made the space for it and created the expectation. They're charging a steep price for flashcards and little else.


Crowd-sourcing is a valid form of support.


If that were even implied to be how the site works, that would be acceptable. They also removed all that. That's garbage "support."


As I think I’ve posted previously, they’re one of those companies that have gone to great lengths to made themselves impossible to contact. Try finding a way to get in touch with them. I’ll wait. Except for billing queries, you won’t. Granted, some of the queries will be language questions you might not have the resources to answer, but that’s a consequence of being a language learning site.

Anyway, to me, it’s a big red flag. (For the record I did use Duolingo for over a year as a paid member with a 365 day streak. At which point I stopped.)


I've found ChatGPT great for this. It's way better than a translator at giving you context for grammar and phrases.

My stack is Duolingo + Podcasts + Instagram + ChatGPT:

Duolingo - daily consistency, grammar

Podcasts - "comprehensible input" from intermediate french podcasts (InnerFrench, French with Panache)

Instagram - Just follow a ton of french accounts. They often have subtitles. Listen to native speakers talk at a normal pace about stuff you're interested in.

ChatGPT - answer basic questions about grammar or phrases. I've even started using this for conversation practice too.


How did you find the Instagram accounts?

I made a Spanish twitter account but had no idea where to even start for following Spanish accounts that interest me. I'd probably be more willing to engage if they talked about things I liked


I think I just searched for some translated terms. Once you find some accounts, you can see who follows them, and then see what other accounts they follow, etc.


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.


"import mlx as torch"


Note that there is a Metal backend for PyTorch [0]. Sadly it doesn't work well with codebases that didn't account for it from the start...

[0] https://developer.apple.com/metal/pytorch/


This is (partly) outdated. MPS (metal performance shaders) are now (since torch 2.x) fully integrated in standard Pytorch releases, no external backends or special torch versions are needed.

There are few limitations left when compared with other backends. Instead of using 'cuda' device, one simply uses 'MPS' as device.

What remains is: the optimizations Pytorch provides (especially compile() with 2.1) focus on cuda and it's historic restrictions that result from CUDA being _not_ unified memory, and lots of energy goes into developing architectural work-arounds in order to limit the copying between graphics HW and CPU memory, resulting in proprietary compilers (like triton) that move parts of the python code into proprietary hardware.

Apple's unified memory would make all of those super complicated architectural workarounds mostly unnecessary (which they demonstrate with their project).

Getting current D/L platforms to support both paradigms (unified/non unified) will be a lot of work. One possible avenue is the MLIR project currently leveraged by Mojo.


> This is (partly) outdated. MPS (metal performance shaders) are now (since torch 2.x) fully integrated in standard Pytorch releases, no external backends or special torch versions are needed.

Not sure what you're referring to, the link I provided shows how to use the "mps" backend / device from the official PyTorch release.

> lots of energy goes into developing architectural work-arounds in order to limit the copying between graphics HW and CPU memory

Does this remark apply to PyTorch running on NVidia's platforms with unified memory like the Jetsons?


Your link suggests downloading nightly previews and v1.12 torch which are both slightly out of date info.


Practically does using unified memory mean that the slow transfer of training/testing data to the GPU would be eliminated?


Yes, I believe so.


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.


https://github.com/predibase/lorax does this, it's not that slow, since LoRAs aren't usually very big.


With a fast NVME loading a model is only 2-3s.


Loss was still decreasing for the models, there's a sense that we can push the training data much much further than we currently are.


Yup. I found this article quite enlightening:

https://espadrine.github.io/blog/posts/chinchilla-s-death.ht...


Phenomenal blog post about scaling laws.


Prediction as an objective basically forces the models to model the casual processes that create the text itself. It's not going to stop getting better unless the data is insufficient/unvaried or the architecture creates a bottleneck.

I think by the time the former is an "issue", we'll have a Super Intelligence on our hands anyway.

The latter is looking less and less likely to be a real hurdle. Very little inductive bias to steer away from crucial solutions, very scalable.


The TinyLlama project is trying to do that pushing by training a small 1.1 billion-parameter model on 3 trillion tokens: https://github.com/jzhang38/TinyLlama


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.


Try right clicking on the .app file/folder and click on "Open" in the context menu, this normally skips the trip to the preferences and offers you the "Open anyway" option


Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: