560 karma · joined September 20, 2016
Current startup founder / CTO
[1]: https://globalnews.ca/news/8480752/ontario-covid-pcr-testing...
Many up and coming (young) players are also really exciting: Alireza, Duda, and Dubov. Duda’s win over Karjakin in the World Cup was pretty spectacular.
On the other hand, Carlsen is well-known for his mental fortitude. It’s something that Vishy Anand, the previous champion, has talked about extensively. Carlsen’s style is often called “squeezing blood from a stone” because he will play equal-looking endgames and mentally exhaust his opponent into making a tiny mistake that he will convert into an advantage.
It was indeed a tough match-up for Nepo, and most knew it would end tragically for Nepo. Despite being an excellent player, he’s not on the same level as Carlsen.
I think it’s more reasonable to assume that they’ll survive any future calamity (especially something so close as 20-30 years) than to assume that our current age is somehow special.
I would be very happy with that, together with a 10% speed increase in single-threaded performance that they said would come as a result of these changes.
This is not correct. There is a lot of data analysis code in the research and scientific community written in python. A lot of PyCon attendees and speakers come from these communities. Oftentimes, it’s not easy to write code to perform a task entirely in numpy, and then you incur massive slowdowns (often 20-100x). This is a common and contemporary problem in the python ecosystem.
GVR initially didn’t think that python needed to be faster either, but recently changed his mind. You can find a presentation on his motivations here: https://github.com/faster-cpython/ideas/blob/main/FasterCPyt...
Sam Gross comes out of that community so he’s familiar with the motivations for making raw python faster.
8 years later, it still works wonderfully.
[1]: https://nationalpost.com/news/canada/toronto-ice-storm-2013-...
I made a mistake in the article, where I meant to write "there are at least 2 solutions and here is one example". I will fix this. Nice catch.
Unlike a sibling comment, I never knew that the lithography process did not correspond to any physical characteristic in the chip. I naively assumed that the headlines saying that Intel was behind in process technology were correct, and used that to inform purchasing decisions. I found the article enlightening, and I think many on HN with the same background as myself will agree.
As an aside, I tried to read the linked IEEE paper [1] but the page is cut off for me below section 2. If anyone has a link to the full PDF, I would appreciate that.
[1]: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=906...
One question: can you elaborate a little on your point about intellectual diversity?
I look up some boilerplate in stack overflow every week for matplotlib figures to look better.
CSVs generated in excel cause errors when reading in python without a special incantation.
It’s not super easy to set up multi machine training. AWS GPUs are extremely expensive.
sklearn models, once saved, are not easy to rework into a different format to be used with other languages.
I wish python could do compile-time checking instead of running for 30 minutes to then report a type mismatch. mypy is still not very good and in beta
In any case, software notebooks definitely existed before Jupyter like you say (I’m not familiar with Wolf but Mathematica has a similar idea), but I don’t believe they date back to 1996. I might be mistaken. I don’t see the author mention them in the text.
- Jupyter notebooks, for teaching machine learning. I use these for teaching. - OpenCL and other libraries for running scientific simulations on GPUs, gpflow for training on GPUs - Keras and PyTorch (libraries for simple training of deep learning models). More than half of machine learning research exists on top of these libraries
Let’s not even get into the myriad recent discoveries in ML and libraries for them.
Parquet file format and similar formats for mass data storage. Spark and Hadoop for massive parallel computation. Hive and Athena further build upon these innovations. A good portion of distributed computing literature is built on these.
Eventually consistent databases and NOSQL. There’s so much here, hard to list everything.
ElasticSearch and Lucene and other such tools for text search.
Then there is all the low level systems research: new file systems like BTRFS and ZFS. wire guard is something that’s just been built that seems foundational.
I am running out of words but let’s conclude by saying the premise of this article is laughable
The credentials by themselves don’t give execute privileges, or persistence, or in fact any access to non-SolarWinds systems. So both can be true.
[1]: https://en.wikipedia.org/wiki/Paul_Graham_(programmer)#/medi...
But the language problem remains. I think there are a lot more good language options today for backend services than there were in 2010, and if I were writing new backend code for a big website/service today, I might pick one of those.
I remember trying to write a program to simply understand our dependency graph so we could decouple some of our services (we were just starting our SOA efforts at that time). This is something other languages give you for free.
Our build times were exceptionally slow, as a direct result of Python's dynamic nature. For example, test discovery took over 2 minutes on each test run. This also made it painful to have distributed test runs, as splitting up tests across multiple processes would require the test discovery step to be run in each process, incurring a 2 minute penalty per process.
We also had deployment issues, because of Python's dynamic nature. Some code paths were only exercised during a production deployment, and were not sufficiently tested before deployment. This is despite the fact that we had tens of thousands of tests and sunk many hundreds of person-hours into testing as much as was possible. Despite a very professional engineering culture, and a constantly expanding test suite covering these corner cases, we had deployments fail about once a week for this reason.
I adore Python and still use it extensively most days (mostly its data science/ML libraries). But I now think that large projects would really benefit from static typing and strict compiler checks, as well as just being able to actually compile a binary (which would have solved many of our testing woes). In contrast to some other smart people on this thread, I don't think that even disciplined developer teams should use Python for large projects, at least not in the long-run.
However I tried to train it to recognize some images of characters from an anime (so a little different than facial recognition), and I managed to break the model: achieving 64% error with significant number of examples per class. I think one downside is Lobe doesn't expose how potentially overconfident the model is. I would love the ability to take the existing model and test it on a new image that I can import into the app.
EDIT: I would love to see the following in a future version:
1. What are the percentages associated with each image per class. I see that an image was misclassified, but did it at least include my desired class in its top 5 predicted classes?
2. Test the model on unlabeled inputs directly in the app to see how well the model might generalize. I would like to see a "Test" tab on the left once training is complete.
3. View other metrics of model goodness like F-1 score and training details like CV partitions in the app somehow.
Again, this is a really cool idea :)