xml

6 ms·

1.
by xml 6mo ago
I'd like to add a few failure modes: - LLM removes/disables/weakens tests (disallowing manipulation of tests is not really possible in Python since the language is too dynamic, so the entire execution has to be sandboxed, whi
2.
by xml 7mo ago
> But on a tangent, why do you believe in mixture of experts? The fact that all big SoTA models use MoE is certainly a strong reason. They are more difficult to train, but the efficiency gains seem to be worth it. > Every thing I know
3.
by xml 7mo ago
Even with inflated RAM prices, you can buy a Strix Halo Mini PC with 128GB unified memory right now for less than 2k. It will run gpt-oss-120b (59 GB) at an acceptable 45+ tokens per second: https://github.com/lhl/strix
4.
by xml 7mo ago
Looks like the pop-ups that cover 70 % of my netbook's screen are still there: https://i.imgur.com/uPbVW2o.png I wonder if getting rid of those pop-ups would have slowed Stack Overflow's decline. The number of new
5.
by xml 7mo ago
To put it into your metaphor: I am not advocating against the existence of bandsaws. I would just rather have bandsaws that do not cut off your fingers if you do not read a book about them first and make it difficult to sew the fingers back
6.
by xml 7mo ago
If most people are not using a tool properly, it is not their fault; it is the tool's fault. Git is better than what came before, and it might be the best at what it does, but that does not mean that it is good. - The interface is unin
7.
by xml 7mo ago
> this code is not copyright protected, therefore you are not allowed to apply a MIT LICENSE to this project. Why not? You still can (and probably should) disclaim warranty and whether the code is copyright protected may vary b
8.
by xml 7mo ago
> Specifically, we collected new data created after January 2025, including: [...] new fiction on Archive of Our Own (Various, 2025), Not sure how to feel about this. From a researcher's point of view, reproducibility is i
9.
by xml 7mo ago
You can still be excited! Recently, GLM-OCR was released, which is a relatively small OCR model (2.5 GB unquantized) that can run on CPU with good quality. I've been using it to digitize various hand-written notes and all my shopping r
10.
by xml 9mo ago
Were there any particular challenges when implementing your library? I have implemented my own serialization library [1] (with a focus on not allowing arbitrary code execution), but had skipped dataclasses for now, since they seemed difficu
11.
by xml 1y ago
A word of caution: There are SVGs which can freeze a page, so make sure that you do not link to any third party SVGs. This is a known bug, but both the Google Chrome and Mozilla team do not want to fix it. Here is an evil example SVG for de
12.
by xml 3y ago
Here is another inference implementation in Python (only dependency is PyTorch). https://github.com/99991/SimpleTinyLlama The new checkpoints did not seem much better and they changed the chat format for some reason, s
13.
by xml 3y ago
I found CLIP to be _amazing_ for all kinds of image search, like search-by-text or search-by-image. I even ported it to NumPy to understand it better. The whole thing is less than 500 lines of Python (including blank lines and comments): h
14.
by xml 4y ago
If you ever use the middle mouse button to open links in a new tab, you should disable this "feature" because it is a security vulnerability which allows for grabbing the clipboard content. Demo: https://randomforest.ne
15.
by xml 5y ago
Can't you just unplug the Ethernet cable/turn off the Wi-Fi module? (At least if the website was not lazy-loading resources.)
16.
by xml 5y ago
I also implemented a batched client-side web tool to resize photos to a certain file size (in a few hundred lines of JS). It only supports JPEG, but it resizes photos almost instantly https://randomforest.net/resizePhoto.htm
17.
by xml 6y ago
There are SVGs which will crash most browsers, so SVGs could still be abused for denial-of-service attacks. For example, you could post one of those SVGs in every issue thread of a GitHub project if you wanted to mess with someone. Not elig