reerdna
Joined 4 years ago
Karma
371
Submissions
42
Member since
Oct 7
Recent submissions
Comment · 3 months ago
Cool, but I find rather than just shapes and indexes, tensors with labels are much easier to use and reason about. E.g.:
{
{user:bob, movie:"Heat"}:0.1,
{user:alice, movie:"Frozen"}:0.9,
{user:carol, movie:"Top Gun"}:0.3,
}
Comment · 6 months ago
This brings back good memories from when a friend made C in nynorsk: https://www.pvv.ntnu.no/~andreer/nynorsk/
Comment · 7 months ago
I was able to get part way some years ago by demodulating the bitstream with gnuradio and then making small changes (like replacing one note with the next higher one) and noting the differences. So that is one possible, but probably too inefficient way.
I never got close to finishing or publishing anything. Awesome to see this released and I'll have to play with it!
Comment · 8 months ago
I couldn't help but laugh out loud at the notion of a "held-out test set" for addition of 10-digit numbers.
Comment · 9 months ago
No, not beyond what I already put in the twitter thread. I wanted to wait until I had some cool distributed software running on it too, but then I ran into trouble with the plan 9 wifi drivers for the rpi being unstable and so I'm still working on fixing that. It does serve as a great test bed for that purpose too, as with 8 nodes I can get much more reliable data about how often the driver gets stuck
Comment · 9 months ago
Wanted to reply to you directly also to increase the chance you see this because I think I had exactly the same intrusive thought as you, and actually built such a cluster recently. Would love to hear what you think: https://x.com/andreer/status/2007509694374691230/photo/1
Comment · 9 months ago
I recently built a Plan 9 cluster of 8 raspberry pi zero (2w). From my supplier (digikey) the 2W was the same price and has 4X the cores!
I think it looks quite cool: https://x.com/andreer/status/2007509694374691230
Instead of having to use lots of dongles and usb ethernet, I just wired them all up using brass rods, a small 5V power supply in the base, and boot them over WiFi (just the kernel and wifi config on the sd card).
Raspberry pi's are cheap, easily available, and there is an absolutely massive trove of information about them on the internet. And the scale means that the linux implementation is very stable and "just works" to a degree that is extremely hard for other SBCs to match.
Sure, VMs are the logical choice, but not everything has to be logical. Real hardware does feel more real :-)
Comment · 1 year ago
Yes, but it turns out it doesn't matter much.
What _does_ matter is the huge reduction in the much more dangerous brake dust, as electric vehicles convert the kinetic energy back to the battery charge via generation instead of wasting it via friction: https://news.ycombinator.com/item?id=44666157
Comment · 1 year ago
Same experience in Norway. 95%+ of new cars sold are electric.
I've been riding my bike to work along the same road for 15 years. I used to have to hold my breath along some sections of my ride. Now the air in the city is almost as fresh as in the countryside.
Comment · 2 years ago
For use in retrieval/RAG, an emerging paradigm is to not parse the PDF at all.
By using a multi-modal foundation model, you convert visual representations ("screenshots") of the pdf directly into searchable vector representations.
Paper: Efficient Document Retrieval with Vision Language Models - https://arxiv.org/abs/2407.01449
Vespa.ai blog post https://blog.vespa.ai/retrieval-with-vision-language-models-... (my day job)
Comment · 2 years ago
Sounds a little like this recent paper;
"RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval"
Comment · 2 years ago
Analysis by dhruv___anand in https://twitter.com/dhruv___anand/status/1752641057278550199 suggests that there are three different "resolutions" in the embeddings, for the first 512, 1024 and full 1536 dimensions in text-embedding-3-small.
You can put a subset of the dimensions in your vector database, thus saving a lot of cost by reducing memory/compute when retrieving nearest neighbors.
Then you can optionally even re-rank the most promising top-k candidates by the full embeddings. At least one database supports this natively: https://twitter.com/jobergum/status/1750888083900240182
Comment · 2 years ago
You can turn this into a pocket "lisp machine" by following the instructions on http://www.ulisp.com/show?4JAO - although I don't think this can make use of the LoRa interface (yet).
The hardware is cool, but its just a PCB and the edges of the display are vulnerable. I would recommend 3D printing a case to protect it, especially if you are planning to bring it around with you. There are .stp files for one in the official github repo.
Comment · 2 years ago
I've been lucky to get to work through a few of these already, they do give a decent introduction in how to use Vespa tensor expressions.
I believe it's one of the most flexible ways in any vector (or other) database to express how to rank your query results, and it's compiled down to fast native code for execution, using openblas etc. where appropriate.
You can reach me through [email protected]