1. You don't have to sell the stock -- one could fund an ultra-wealthy lifestyle by renting expensive things, using stock as collateral. As your stock goes up in value, so does the value of things you have access to.
I was an engineer in Google Books in 2011 and got an impression that this effort turned out to be not as useful as expected. This took away energy from the legal fight. Having access to every printed book seemed amazing in principle, but it didn't translate to amazing demand for the service. My theory is that it didn't lower the barriers to knowledge much -- there's not that much useful information left in offline-only storage.
I think there's a lot of information sleeping in depositories only on paper. The thing is people are busy with the urgent stuff and stay away from the important if they can't get something immediate from it.
I've used extensively Google Books for research of 17th century Mathematics (not just me, also the whole department were I was at that point). I couldn't have afforded to visit all the libraries to look up the originals but thanks to this repository I could download several copies of obscure works, which gives your very interesting insights: sometimes you hit a personal copy of a previous researcher, or there's an ex-libris from an old institution that you know your author was connected to, or you just discover something no one had noticed before simply with a clever string search. You can even backtrack to the forgotten true primary source of a mistakenly repeated fact for instance. It's been a blessing, honestly.
If you are into whatever not too transited alley of history, you realise soon that your (former?) company has produced a treasure trove for future generations that just shouldn't be shut down ever: there are many books there not digitized anywhere else. So thank you very, very much for your work there. It definitely belongs to the important.
Sure. The $2.48/hour per V100 GPU on GCP does not include the price of the CPU host; that is purely the price to rent a single accelerator. By contrast, a network-attached Cloud TPU v3 device includes both a CPU host and four connected TPU v3 chips that collectively deliver up to 420 teraflops. Furthermore, each individual V100 GPU on GCP has 16 GB of memory, whereas the Cloud TPU v3 device has 128 GB of HBM.
The best apples-to-apples performance-per-dollar comparison we have publicly available was published last fall, and it compared the performance and cost of using various Cloud TPU v2 Pod slice sizes with the performance and cost of using various numbers of V100 GPUs attached to a single GCP host:
We went to great lengths to ensure that we trained exactly the same version of ResNet-50 to the same accuracy in the same way across all hardware configurations. The methodology predated MLPerf and is documented in full here:
If you were going to do a similar performance-per-dollar comparison today, the simplest approach might be to try to get the code from NVIDIA's MLPerf 0.6 submissions running at scale on one or more major public clouds using the fastest-available networking technology that each cloud provides:
It would be very interesting to see how distributed training performance using large-scale GPU clusters in public clouds compares with the published on-premise MLPerf performance numbers using exactly the same MLPerf code and methodology. With these measurements in hand, it would then be straightforward to make performance-per-dollar comparisons with Cloud TPU v3 Pod slices of various sizes.
One of the authors here. Thanks for the comment! Yes, we mention this work and number of others in the blogpost and paper. This isn't the first (or the last) paper on the topic but I think we've clarified the large-batch training situation significantly by connecting gradient noise directly to the speed of training, and by measuring it systematically on a bunch of ML tasks and characterizing its behavior.
I've been using Mathematica since 1995 and Jupyter/colab for 5+ years. Most recently I've been using them both in parallel. While Jupyter is probably the future in terms of mass adoption, there are still some areas where Jupyter is lagging.
2. Mathematica has more active community. Mathematica-specific questions are likely to be answered within an hour by experts on https://mathematica.stackexchange.com/
3. Mathematica has better tools for simple interactivity. I like to throw in "Manipulate" for a simple graph with a draggable constant, or go to http://demonstrations.wolfram.com/index.php for an idea for more complicated demonstration to use in a presentation
4. Mathematica has more options for advanced visualization, and interfaces are more uniform since graph drawing, 3D drawing, and other kinds of visualizations are developed within a single system. Some examples https://www.wolfram.com/language/11/new-visualization-domain...
Thanks for your feedback, Mathematica has indeed millions of $ to provide more features and advertise them, and Jupyter have only a few full time devs that probably do not advertise enough its features:
1) Binder makes that a git push away https://mybinder.org/ Want to check the discovery of gravitational waves ? Go ahead ! https://github.com/minrk/ligo-binder You know the nice thing ? it does not require you to opt-in, as long as a repo is public you an run it. So you don't need have to deploy, or know it exists. we are _already_ doing that for you.
2) Jupyter is "Just" the frontend. StackOverflow have matpltlolib, numpy, sympy, .. tags. We don't the subdomain (yet), and I actually prefer to have tags to have better searching :-)
3) Sure it's called ipywidgets (https://ipywidgets.readthedocs.io/en/latest/), that's the tech. From ipywidgets import interact, and @interact as decorator on your function... that's it.
MyBinder is great. As well as cocalc and sagemath. There's also https://notebooks.azure.com which is a free hosting of Jupyter notebooks. You get a linux/docker container w a Terminal, Anaconda, etc.
With regards to sharing, both Google and Microsoft have free hosting for shareable jupyter notebooks. Probably not quite as easy to get them from your computer to the cloud as a deploy command, but it probably wouldn't be hard to create a module that does exactly that (if one doesn't already exist)
