pama

51 ms·

661.
by pama 3y ago
A simple Emacs imitation for eshell: https://karthinks.com/software/jumping-directories-in-eshell
662.
by pama 3y ago
How easy/tricky would it be to bring the pytorch documentation up to speed?
663.
by pama 3y ago
To people who have trouble with this concept I suggest the analogy to reading without voicing out words yet still internalizing the concepts, and often at a much faster rate than when vocalizing during reading (which in turn can be pleasant
664.
by pama 3y ago
Agreed. I hope there will be commercial scale inexpensive solutions of this type in the next ten years.
665.
by pama 3y ago
Thanks for the insights. What is your typical Emacs workflow for using and comparing the models?
666.
by pama 3y ago
Thanks for the neat estimate. It would be hard to dehydrate DNA to a significant degree, so adding some water and counterions would probably reduce your estimate by a bit. (Error correction would also reduce things by a bit.) Overall the e
667.
by pama 3y ago
Global variables are true for viruses or single cell organisms. There is locality between different cells and partial locality within cell compartments. So not everything is a global variable.
668.
by pama 3y ago
It is a reasonable bet to assume that openAI had some powerful models along the lines of Chris Re’s group well before mamba came out. They hired people with the right background and certainly the cuda optimization would not be a problem for
669.
by pama 3y ago
Interesting and expected core idea. The MoE models in table 1 have drastically larger total parameter count than the baseline transformer, which in turn is super tiny for the purpose of NLP (25M parameters). I suppose inference is of simil
670.
by pama 3y ago
Why not include redshift in the mix?
671.
by pama 3y ago
In fairness, the title of the page is “Dataframes for the new Era”. The “Get Started” link below the title links to a document that points to the GitHub page, which explains what the library is about to people with data analysis background
672.
by pama 3y ago
"Unfortunately popular git GUIs all fail on giant repositories or become unresponsive and unusable." I have only witnessed that in magit (or pyright) when I have super deep directory trees (say a wide dataset collection) not yet
673.
by pama 3y ago
A minor detail here is that the analysis in the blog shows that the linear model built/trained on the the activations of an internal layer has a representation of the board that is probabilistic. Of course the full model is also probab
674.
Isomorphic labs kicks off 2024 with 2 pharmaceutical collaborations(isomorphiclabs.com)
2 points by pama 3y ago|0 comments
675.
by pama 3y ago
Also: https://www.isomorphiclabs.com/articles/isomorphic-labs-kick...
676.
Isomorphic Labs deals with pharma companies(endpts.com)
19 points by pama 3y ago|2 comments
677.
by pama 3y ago
Emacs has an excellent csv mode.
678.
by pama 3y ago
As a distinct difference to personal computers replacing typewriters and the internet enabling self publishing, both of which transformed writing, AI is much faster than humans in reading and analyzing books and thus can also be part of a s
679.
State of the Pandemic(erictopol.substack.com)
5 points by pama 3y ago|0 comments
680.
by pama 3y ago
Sometimes ssh -C helps on slow connections. Also mosh is great for those cases, but optional and has different setup requirements (see below). As others have said start a -daemon on remote and use emacsclient -nw when you connect. Spend so
681.
by pama 3y ago
Do you have a performance comparison for inference on the same hardware using each of these implementations?
682.
by pama 3y ago
Agreed on the accidental reimplementation of something trivial. The distinction of what is trivial will be harder to clarify in the future, however the language models in the future can also help identify the provenance of substantial parts
683.
by pama 3y ago
Nobody has to look at the code to violate the GPL though. An acccidental identical implementation or one spewed out by an LLM counts as a simple translation would largely place the new code under GPL at a technical level. I don’t see the
684.
by pama 3y ago
Modern DL architectures are complicated. If you want to gain insight if a loss function is convex (bowl shaped), you can calculate the second derivatives matrix (Hessian) at arbitrary points and test that all its eigenvalues are non negati
685.
by pama 3y ago
I thought the main point was that this is a very fast way (in terms of wall time) to beat state of the art, not a fair comparison by size; if one made E5 bigger, then E5 would be even slower to train.
686.
by pama 3y ago
Here is the main arxiv page (html is a mess on lockdown mode): https://arxiv.org/abs/2312.08361
687.
by pama 3y ago
We need the full pipeline of tools. What jart did is helping the future users of AI gain familiarity early.
688.
by pama 3y ago
Excellent work—thanks! Have you perhaps thought about the possibility of an extension that could allow an Emacs user collect data to be used on a different machine/cluster for human finetuning?
689.
by pama 3y ago
In addition to the project you mentioned, there is quite an ecosystem for Emacs working with API-based LLM: https://github.com/emacs-openai/chatgpt https://github.com/stuhlmueller/gpt.el https:&#
690.
by pama 3y ago
It is great to see that the simplicity of Emacs wgrep inspires new tools. The Emacs interface is more powerful than this new tool because one can more easily do selective or arbitrary edits in the search results, but the new tool is easier

More ›