If by running models you mean just the inference phase, then even today you can run large family of ML models on commodity hardware (with some elbow grease, of course). The training phase is generally the one not easily replicated by non-corporations.
Thanks for the update. I had a feeling that this might have been the case due to the launch crunch and it's perfectly understandable. Really excited for the future of this project!
Currently the only way to sign up for a personal account is through google. Is there another way that I am missing, or are there any plans to provide email based signups in the future?
I am looking forward to performance improvements. It would be great if the core window services were optimized to use less resources to improve the windows experience on less impressive hardware.
> I don’t care about Kubernetes, but someone might read this post and decide that it is a good idea to “simplify” an existing project by rewriting it in 17000 lines bash script.
I think the 17kloc would discourage that kind of thinking.
Great project! I've had success using camelot-py (https://camelot-py.readthedocs.io) to extract tabular data from PDFs (for images, I use imagemagick to convert those to PDF). If your table has borders the default method (lattice) works quite well. For non-bordered table there is the option to use 'stream' option but usually requires bit more preprocessing to get usable results.
I am considering building a PC for programming and am genuinely curious if WSL2 can be considered a replacement for dual booting Linux and Windows? I am currently rocking a dual boot system (with legacy bios) and haven't had any issues so far.
For those who might not be familiar, graph subsampling is a method of extracting smaller graphs from a large graph while preserving some notion of overall structure. The subsampling can be focused on finding representative nodes, edges or combinations of those (as well as optimizing for some other local or global properties).
It has applications wherever large graphs are encountered - social networks, biological processes etc. An example application of the library could be as follows - you have a very large graph and you want to do spectral clustering. Throwing spectral clustering directly at such large graphs might not be feasible, so one can first perform a node based subsampling and then apply spectral clustering on the smaller graph on the representative nodes found by the subsampling.
The original content was a series of tweets and the link is a website that summarizes long twitter threads, so each newline indicates a follow-up tweet.
The jitter artifacts could be removed by applying some kind of temporal consistency constraints on the network predictions. But such an approach will also definitely introduce a lot of edge cases where a sudden change is actually ignored due to enforcement of the constraint. An adaptive consistency parameter could probably work better, but it's non-trivial to have a meta-algorithm to figure out the paramter.
The inventor has about 134 patents [1] and many of which seem (to me) very generic. I am not very familiar with patent jargon so I can't definitively say if those patents have enough substance to them. Maybe someone with experience could expand on this?
Spacy also has one of the best documentations that I have come across. I really like the use of Tufte-style notes for code examples. https://spacy.io/api
Such competitions do not usually result in a comprehensive "solution" by themselves - pushing the state-of-the-art is more common. Also the value is not going to be derived solely from the algorithm but more from its deployment to real world applications and the surrounding infrastructure to make it possible.
Though neuromorphic architectures are interesting, the Loihi is essentially an ASIC for Spiking Neural Networks [1]. The lower power consumption than GPUs is due to the limited nature of computation that can be performed on these chips (which is generally true for ASICs).
Looks cool. Iirc the transformer architecture doesn't allow any constraints on the learned language model. For code completion settings, a model aware of the (programming) language constructs explicitly and then augmented with code samples would be much more efficient (you could greatly reduce the search space for next token etc).
The article talks specifically about usability testing - though it isn't in the title but it is in the first line of the article. I don't think pricing strategies or matching algorithms and such would fall under this domain.