reichardt

With around 4.6 GiB model size the new Qwen3-8B quantized to 4-bit should fit comfortably in 16 GiB of memory: https://huggingface.co/mlx-community/Qwen3-8B-4bit

You likely mean sampler, not decoder. And no, the stronger the quantization, the more the output token probabilities diverge from the non-quantized model. With a sampler you can't recover any meaningful accuracy. If you force the sampler to select tokens that won't repeat, you're just trading repetitive gibberish for non-repetitive gibberish.

Looks awesome! The YouTube video linked inside the GitHub README seems to be unavailable: https://youtu.be/zm0XbLmXtXY Would love to see an up-to-date demo video.

Amazing! The inserted objects are renders of textured 3D models and not generated by a diffusion model + ControlNet? Is there a fixed set of textured 3D models available or are they generated on the fly based on the prompt?

GPT-4 seems to perform better: https://twitter.com/evanthebouncy/status/1641927192975376384

Super interesting! I’m currently writing my master thesis on analyzing relationships between subreddits based on user and semantic similarity. For user similarity I use the Jaccard similarity between the unique set of authors of each subreddit.

Can I send you a message and quote you in my thesis? You can shoot me a short message as well: violets.parr-0c@icloud.com

Still questions that I would assume are semantically similar to the questions you can find in exam prep material all over the internet. My point is that exams are a crutch we use to determine how well a person studied a subject. A crutch we use because we seem to lack better measuring devices. It's very much possible to ace an exam while at the same time being horrible at actually applying/working on a subject. I'd argue therefore that measuring how well LLMs perform on exams designed for humans is simply a more complicated Turing test, with all its shortcomings.

Do you know that the answer key wasn't part of the train set of ChatGPT?

Thanks! Just stumbled across your comment as well.

I recently looked into this as well. Specifically how the two formats differ. As it stands right now the “Feather” file format seems to be a synonym for the Arrow IPC file format or “Arrow files” [0]. There should be basically no overhead while reading into the arrow memory format [1]. Parquet files on the other hand are stored in a different format and therefore occur some overhead while reading into memory but offer more advanced mechanism for on disk encoding and compression [1].

As far as I can tell the main trade-off seems to be around deserialization overhead vs on disk file size. If anyone has more information or experience with the topic I'd love to hear!

[0] https://arrow.apache.org/faq/#what-about-the-feather-file-fo... [1] https://arrow.apache.org/faq/#what-is-the-difference-between...

EDIT:

More information: https://news.ycombinator.com/item?id=34324649

That's impressive. The audio track of your videos is so clean and well understandable that I was wondering if you used a studio setup or voice synthesis software. Well done!

Ah, thanks! I didn’t know that and shouldn’t have used the first result that came up with a google search.

Yes, it's easy to see you put a lot of thought into that. I hope your course receives much more exposure. When I first found your videos a few weeks ago, I was surprised how few views they have given to the quality of the course.

Do you record the voice track of your videos yourself?

Glad to see you published the final lesson about transformers. Was looking forward to that!

I love your course for being very comprehensive and technical while not getting lost in mundane details. Like the opposite of the following quote:

“I didn't have time to write a short letter, so I wrote a long one instead.” [1]

[1] https://www.goodreads.com/quotes/21422-i-didn-t-have-time-to...

Seems like the Apple M1 Max is almost half as fast as a RTX 3090. Pretty cool!

HTML

As shown in the first example you can pipe data into clickhouse-local: https://clickhouse.com/docs/en/operations/utilities/clickhou...

I read it as 1/2 cent as in 1 or 2 cent per render.

Austria—Schnitzel not kangaroo.

Why not create a table with a single JSON column and dump everything in there?

Does this work similar to rqlite or dqlite from a usage standpoint, or does it solve a different use case?

That's great to hear. I'd love to create some vector tiles with building and building height data on zoom levels down to 11. Is this the correct place to perform this change? https://github.com/onthegomap/flatmap/blob/e402a3333bc5127e5...

Or could I even achieve this via the config file? https://github.com/onthegomap/flatmap/blob/main/config-examp...

Fantastic work! Are the vector tiles compatible with OpenMapTiles styles?

Awesome! Thanks.

Any best guesses if and when we'll reach the "Kubernetes: The Good Parts" stage? A condensed subset of the currently available features of k8s which would prevent most of the foot guns and make k8s feasible for <20 developer projects.

Seems to support QXL graphics driver and SPICE? I ran into problems compiling QEMU on macOS with the above mentioned. Did you solve this?

https://stackoverflow.com/questions/59636198/how-to-compile-...

Something I’ve been wondering: what’s the difference between this and buildroot?

Something like https://workers.cloudflare.com/ ?

TensorFlow Lite with SSDLite-MobileNet gets you around 4 fps on a Raspberry Pi 4 (23 fps with a Coral USB Accelerator): https://github.com/EdjeElectronics/TensorFlow-Lite-Object-De...

One of the killer features is the multi-platform support, including web browsers, in combination with synchronization across devices and the Google Docs plugin. Makes working on a paper across multiple devices super easy.