chaxor

35 ms·

391.
by chaxor 3y ago
H5AD :D
392.
by chaxor 3y ago
Lol yeah, I always reach for paperback books now since they're a great modern replacement for my Kindle. On a more serious note, DuckDB is actually a pretty fantastic replacement for SQLite for some data analytics tasks. Hdf5 is still
393.
by chaxor 3y ago
They are indeed radically different: one is very well understood in terms of how it works, and the other - neuroscientists have absolutely no idea how it works.
394.
by chaxor 3y ago
Great point - a good threat model usually targets all companies first, such as FAANG, since they are the worst aggressors, and are well known to be leaky (data dumps of personal info are leaked almost daily now), whereas the government tend
395.
by chaxor 3y ago
Maybe they rewrite it in Rust :)
396.
by chaxor 3y ago
I thought was impossible somehow? Something something Barry's Paradox...?
397.
by chaxor 3y ago
At least they're using ssh and not some home rolled garbage. If ssh uses keys and not small passwords, it's one of the most secure methods around to get access to a machine. Unfortunately far too many companies are just rolling t
398.
by chaxor 3y ago
Excellent advice. I will now look only for things that I have no chance to understand. Proprietary systems are probably the best bet since we are, by definition, not allowed to understand. Only windows for OS and secure pulse for internet
399.
by chaxor 3y ago
NER, Entity linking, and relationship extraction definitely seem to be 'low hanging fruit' due to LLM improvements, but one of the big problems is that they really need a completely different architecture to limit the decoder voca
400.
by chaxor 3y ago
BERT-like encoder models actually do help here, although unfortunately the dense representation is not as nice and clean as a class, it will at least map to something similar. If the hospital provides even a completely bizarre entity id, t
401.
by chaxor 3y ago
Torrents are the backbone of a lot of academic research downloads. They're the absolute best way to distribute neural network weights, or large models used in physical science studies. That or IPFS. If you use huggingface or dockerh
402.
by chaxor 3y ago
The biggest problems in aware of is sync conflicts, which just make things a little difficult. If they're text files it's not so bad, since vimdiff can easily merge. But if they're encrypted or more complex formats ... :&#x2
403.
by chaxor 3y ago
If you have a task that requires something suggested by "__exact__", then a full LLM is probably not the answer anyway. Try distilling step by step, especially if the goal to to generate a DSL or some restricted language. It can
404.
by chaxor 3y ago
GPT-4 has made some substantial improvements recently in references. A few weeks ago using it would get things sort-of correct (the authors would be nearly the same, with one or two just added in or removed, maybe journal number a bit of
405.
by chaxor 3y ago
I think this is a reasonable response to the terrible atrocity that is medical care (in the US*). The way forward is likely going to be the medical form of 'self-hosted': home medical devices for self diagnosis. There will be man
406.
by chaxor 3y ago
Yeah that's a good suggestion for that use case. I was thinking a bit more along the lines of 2 other use cases: 1) If you have a file locally and want to send a link to friend/family member (yourself or even some random person
407.
by chaxor 3y ago
I always assumed having a raspberry pi with a couple HDs in raid1 with IPFS or torrent would be the best way to do this. Giving another one of these raid1 rpis to a friend could make it reasonably available. I am very interested to know if
408.
by chaxor 3y ago
I'm not sure why these are presented as mutually exclusive
409.
by chaxor 3y ago
Texas wasn't really set up for being so cold, so it caused deaths and considerable damage because the infrastructure was simply not good enough to handle it (despite being very new/modern in most cases). It got politicized because
410.
by chaxor 3y ago
Puns and wordplay are easy? I think most linguists would heavily disagree. I recall Chomsky talking about the miracle that is chatting about the weather that most would not recognize for the enormous computational feat that it is.
411.
by chaxor 3y ago
Pretty sure this is just pasting the link into GPT-4 with browsing summary
412.
by chaxor 3y ago
This is certainly something that needs to be addressed far more often. Too many times have I seen some data scientist trying to parse and write 6 Petabytes of data with multiple cores, while the disk is thrashing about. Spinning disks are s
413.
by chaxor 3y ago
I don't understand this comment. Spinning disks are still used very often in computing. SSD/NVMe/etc drives are great for tiny amounts of storage, but if you need to process something substantial, like petabytes of data, you
414.
by chaxor 3y ago
Spacy is a decent suggestion here. They have pretty good ways of writing tagging rules. All of this does seem to be extremely excessive to choose a book genre though. I would imagine the number of books after a simplistic clustering techn
415.
by chaxor 3y ago
I don't see why a simple TFIDF with ~10 LoC and a few minutes doesn't make this at least reasonably/crudely done.
416.
by chaxor 3y ago
This is an extremely wrong viewpoint. "Many extremely important* papers have been published on arxiv alone - no other publisher. Just off the top of my head, without looking, ELMo from allenAI (the paper that started the contextuali
417.
by chaxor 3y ago
Many people may disable js when browsing. So there's that.
418.
by chaxor 3y ago
Almost none of the "private LLM" ideas are private. There are hoards and hoards of idiots out there that continue to use OpenAI unfortunately. The best thing to do is make your own Dockerfile to run Gaucamo locally. Putting it o
419.
by chaxor 3y ago
I didn't see the actual data or a link or anything to how to get it, but why on earth would it be 100GB? That seems fairly large if it's just tallying accidents and such.
420.
by chaxor 3y ago
They are GNNs with attention as the message passing function and additional concatenated positional embeddings. As for reasoning, these are not quite 'problems well-trained for', in the sense that they're not in the training

More ›