mufasachan

I appreciate the author and his post, but the pro's and con's seems really a confirmation biais. Note that I do agree on the part that it will cost users and contributors to use AI. What I think is biais is the technical consequence. I genuinely ask how backed is the "AI does not help productivity". This is stated as a truth while we are only starting to discover how to use it, including guardrail, review system where durability (understand ownership of the source) is the goal. This can be viewed differently, "ethic"/environment set apart. New technology emerges, the community explores methods and overall the evaluator is time that measures durability and engagements. Short-term costs are the reject of exploiter (the ones that does not explore) and cost of exploration. So, for the productivity, the best strategy would be to explore partly, not being all exploiter of current method.

Insightful comment ahead:

Is the gravity set very high or am I getting too old to play Flappy Bird with Transformers?

The blog is technical, technical terms in the TL;DR seems relevant to me.

Me too, I set up a WG tunnel to access this.

Same experience, my workflow is to run the container from a podman run command, check it runs correctly, podlet to create a base container file, edit the container file (notably with volume and networks in other quadet file) and done (theorically).

I believe the podman-compose project is still actively maintened and could be a nice alternative for docker-compose. But the podman's interface with systemd is so enjoyable.

I appreciate the motivation of the author but the examples of this blog do not convince me at all. For every situation, there would be an alternative (with init or not) that addresses the design issues. The first example is solved with namedtuple[1] which is not a recent feature. Then the rest of article is about IO. Although, the special methods for the context managment interface and the std lib for context managment (which do a great job at async btw) are not mentioned.

__init__ in Python already has a lot of alternatives. IMO a situation encountered similar to the example is when you do not take advantages of the alternatives or your code design is bad.

Also, specifically for this case, if I need an object for this IO operation, I __would not__ uses dataclass. The first solution would be to create a context manager[2] and a NamedTuple (or dataclass if needs some mutations). If it's not a fit, the classmethod and the dunders __enter__, __exit __ are good if this fd implementation is important for the design of my object. Alternatively, I would just make a builder function outside my object if the implementation of the fd capability should be considered as extra and not a builtin for the object. The dataclass decorator could be added to save 1 line of code but would confuse the intent of the object which is a content manager with some state management.

A side geniune question, why not a type[3] statement to alias the int as a fd? What is the difference with using NewType[3]? In which situation creating a NewType from a base type - here int - is useful?

[1]: https://docs.python.org/3/library/collections.html#collectio... [2]: https://docs.python.org/3/library/contextlib.html#contextlib... [3]: https://docs.python.org/3/reference/simple_stmts.html#type [4]: https://docs.python.org/3/library/typing.html#typing.NewType

My bad! As you said, this part of Python code bases is chaotic (:

Eh, ML/scientific Python is large and not homogeneous. For code that should work on cluster, I would lean towards a Docker/container solution. For simpler dependancy use cases, pyenv/venv duo is alright. For some specific lib that have a conda package, it might be better to use conda, _might be_.

One illustration is the CUDA toolkit with torch install on conda. If you need a basic setup, it would work (and takes age). But if you need some other specific tools in the suite, or need it to be more lightweight for whatever reason then good luck.

btw, I do not see much interest in uv. pyenv/pip/venv/hatch are simple enough to me. No need for another layer of abstraction between my machine and my env. I will still keep an eye on uv.

I do not know the people's background of a lot of comments here. They might have much more experiences than me with tensors. But, in my deep learning code and works, when I need to design an operation that involves a mix as little as 3 tensors with 4+ dimensions, I always struggle. I need to draft some slices to understand which slice should be contracted etc.. Many times the shape of the output is not even clear in my mind. Plus, add some padding maskS on the tensors and it confuses me quite a lot. I really like this notation, the last example of 1.1 is readable in its sum formulation, but the diagram formulation is much more "alive" in my mind.

I am really lost here if I have missed something about indices notations with tensors or some visualization techniques. Or maybe the confusion of a tensor operation depends of the field? Or maybe I just miss practices and experiences with indices notations...

Thank you for your answer. Some code base have chaotic execution path through a monolithic code base (by design). So, for these code bases, I heavily rely on the debugger. But I like your suggestion to use both of them, I think it's a good way to transition slowly and efficiently.

Thanks for sharing your experience!

I would be curious to know how you managed to do this. I really tried to do this but the tons of dev tools I am using was too much for transitioning to neovim for my daily work. Namely, I need a DAP, multiple dev tools (lsps, linters, formatters) because I work with several projects which do not have the same tools[^1]. Luckily, I do not mix multiple programming languages. Plus, I containerize all my dev env. There might be some elements missing, but the point is the number of tools is overwhelming and it makes me think that I should do the whole configuration/setup on my free time.

Did you face similar issues? If yes, how did you solve them? Or maybe your work does not need that much tools? Or you have been more minimalistic than me for the number of features to be included in the neovim configuration?

[1]: I work in R&D, I need to tweak and contribute in many papers code or different toolboxes/frameworks on top of the team projects.

fair comparison: tiananmen square versus yt-dl. Hilarious.

(I 100% agree with you, just your comment gave me a funny smile. The comparison is relevant.)

Thanks, my bad.

Thanks for the details!

Great benchmark, very interesting. Although, I am not sure about the extrapolation of the H200 from the lambda bench. From my understanding, Lambda and theirs bench used different models - LLama 405B and Mistral 123B - with different bench and inference libs. Since the study is focused on memory-hungry scenario, I am really curious to know why they took H100 instead of H200.

An inspiring project. I am looking forward to see some gloves connected to a VR device. I think that some cheap sensors, a bit of bayesian modelling and a calibration step can offer a proper realtime hand gesture tracking.* I am already picturing being able to type on a AR keyboard. If the gloves are more expansive there might be some haptic feedbacks. VR devices might have more open OSes in the future or could use a "streaming" platform to access remote desktop environments. I am eager to see all the incoming use cases!

*: a lot of it. Plus, the tracking might be task-centered. I would not bet on a general hand gesture tracking with cheap sensors and bayesian modelling only.

Arguably, it's off-topic, though I agree with the point. Lebanon has been struck by poverty, and as a result, they might have far fewer choices when it comes to providers in general. Manufacturing within Lebanon or trading with neighboring countries might not be affordable for them.

It’s important to take a step back before generalizing an economic or political statement that may not be applicable in other contexts. There are little chances that the supply chain in Lebanon is in the same state as Europe countries' ones, for instance. Thus, this is not another example.

> Although the author OCR’ed the SAT questions and believes that they weren’t in the training data

I agree that the author of the tweet fairly underestimates the potential portion of OCR'ed contents in OpenAI's training data. In late August, Nougat[1] is released by Meta, this is an OCR model. Its performance are wild and the model is open source.

I hardly believe that OpenAI does not spend effort on getting more training from OCR content. I also hardly believes that OpenAI waits for a Meta paper to have an internal performant OCR model.

[1]: https://arxiv.org/abs/2308.13418

Disclaimer: This is just my intuition, I do not have knowledge about LoRA on small models.

It's possible that does not work. LoRA (for Low Rank) benefits from the "small changes" introduced during finetuning of a model. The update of the weights has a low rank. If you take a smaller model, it might induce that the rank is not so low, resulting in degradation in metrics by LoRA compression. I would be interested to see if LoRA still has a benefit in this configuration.

One explaination https://x.com/yampeleg/status/1707127722743325106?s=46&t=Cxa...

I would be curious what does he mean by "semi-automated system for detecting benchmark leaks. " though.

A simple research on Scholar with "facial recognition liars" leads to a few niche articles about the topic. So it seems that facial recognition for liars is a thing. After a quick look on the articles you can see those methods are very simple and they are not studied at the state of the arts level. It looks like much more an academic research field rather than an industrial field. So her claims are preventive alarmist.

Although, I would say she is not the most informed person in surveillance task in AI. Another surveillance field in AI is the "person re-identification" task. It aims to determine if two pictures of a person from surveillance cameras depict the same person or not. This is my field of research and you have a tons of papers each year in big AI venues and also big company in the author (alibaba, tencent, mostly big chinese tech companies, I saw one paper with Google once on a related task).

Also, I would like to highlight something. She said we need some people to annotate the datasets. For this surveillance task, actually, we have a bunch of methods that aims to less rely on human annotations : LUPerson and LUPerson-NL datasets and respective methods (CVPR 21 and 22) for example.

Finally, I would like to conclude that I mostly agree with her. A bit of nuance would be to see the interaction with the nature of data and the surveillance capabilities. AI on camera data are obviously surveillance. When it comes more blur it's what "surveillance" is for "code taken from the internet to train Github copilot method"? Or what is the surveillance capability from database of digital arts? Those questions seem to be much more relevant and hard in regard of the use of data for surveillance. AI is just a name of a method that we use to process big data at scale in this context.

TL;DR: One is about distance in space, the other is about spread in space.

KL-Divergence is not a metric, it's not symmetric. It's biased toward your reference distribution. Although, it gives an probabilistic / information view of a difference in distributions. One of the outcome of KL is that it will highlight the tail of your distribution.

Euclidean distance, L2, is a metric. So it is suited when you need a metric. Also, it does not give an insight of any distribution phenomenons expect the means of the distribution.

For example, you are a teacher, you have two classes. You want to compare the grades. L2 can be a summary of how far the grades are apart. The length of tails of both grades distribution won't have impact if they have the same mean. That's good if you want to know the average level of both classes. KL will give the point of view how the class grades spread are alike. Two classes can have small KL Divergence if they have the same shapes of distributions. If your classes are very different - one is very homogeneous and the other one is very heterogeneous - then your KL will be big, even if the average are very close.

To be honest, OP seems to be at the beginning of its reading journey. His method is laborious. This "chain of summaries" seems to be hardly effective. I think the fractal method is close to be atomic - so a good one IMO.

I would suggest to read the book multiple times. The books mentioned here are not literature masterpieces, they are philosophical or they are essays. So the point is to extract information. So it's ok to read the book multiple times. If you understand 60% on the first read, then 80% on the second and 95% on the third, it's alright (Figures are illustrative). By the way, at 95% of understanding of a book, you are already a "specialist" of the book.

Next, take notes when you feel it's important about some concepts. Taking notes is an experience things. The best way to boost this learning would be to learn the semantic of the book before reading, actually it's the role of the first reading. I would not advise to take notes on the whole book structure before the second reading.

Finally, use Zettelkasten-like method to link your notes together. It will reduce redundancy, labour and keep all relevant information.

Everything is in the SO post. Also, this is my first SO question, if you have any remarks regarding it - putting aside the problem in itself - I am all ears.