physicsyogi

Feynman's Lectures are great for understanding conceptually, but they don't teach you what you need to know to solve research problems. Some of what's in TAOCP is directly applicable to research (for example, the combinatorics part).

I agree with a lot of this.

The one thing that I've found with some of these "fundamental" books is that there are gems of information that can be very helpful for specific problems.

For instance, Landau and Lifshitz's statistical mechanics book has first-rate description of entropy and multiplicity. And TAOCP has an incredibly thorough discussion of combinatorics that I found invaluable.

My friend's VPN emailed him saying Comcast asked them for his info because he torrented something and the copyright holder contacted Comcast. The VPN provider didn't give Comcast any info.

I can't speak for others, but I'm far more productive in Clojure. I'm able to write less code in less time, without any loss in correctness. My productivity comparison is in relation to Java and Python, which is mainly what I use at work.

There's a rebuttal to this as well: Attention is not not Explanation. https://arxiv.org/abs/1908.04626

Focus is usually directed at a single thing though, it's a narrowing of attention. Attention mechanisms can pay attention to more than one thing at a time.

Yes. I don’t know if that’s all that they do. They often port new Tensorflow models to PyTorch as well. They provide straightforward APIs, nice documentation, and clear tutorials. I use their stuff pretty regularly.

I'd probably be happy with "update", which is similar to Clojure's API for its persistent vector and map.

https://clojuredocs.org/clojure.core/update

Persistent data structures are thread-safe and are useful for concurrency. That's partly why they are a core part of functional programming languages like Clojure and Haskell. Persistent data structures are also useful for version control.

There are a few numerical computing benchmarks for which Rust is sometimes faster than C: https://benchmarksgame-team.pages.debian.net/benchmarksgame/...

This isn't cool, but it doesn't bother me too much because iOS app store search has been useless from day one. It should almost never be used unless you already know what you're looking for. There are better places to find music or notetaking apps to download.

Language in Thought and Action, by S.I. Hayakawa. I read it in grad school and it opened my mind to how much of human language is subjective and made me think about the problems that arise because of that.

I recently finished reading The Overstory. I thought it was pretty interesting. It got me interested in video game programming (as did reading Ready Player One).

I may be wrong, but my guess is that Jeff Dean’s presentation is what lead to Google’s efforts with Swift for Tensorflow.

There are two parts to this: (1) labeling something as a date or time and (2) normalizing it to a time stamp. The first part is the tagging. The second part is temporal normalization.

There are several libraries for temporal normalization:

- Duckling: https://github.com/facebook/duckling - JChronic: https://github.com/samtingleff/jchronic - There's also Chronic (Ruby version that jchronic was made from).

Stanford NLP and SpaCy also do tagging: - https://github.com/stanfordnlp/stanfordnlp - https://spacy.io/usage/linguistic-features#named-entities

Edit: Stanford NLP does not do temporal normalization. Added SpaCy

The title of the article suggests that Siri is not privacy-minded, but that is definitely not true. Having more choice in virtual assistants would be great though. And, if it's good, hopefully it will not go the way of Viv.

This is a nice intro to RNNs. The article title says RNNs are the workhorse of language modeling, but the transformer architecture used in models like BERT and GPT-2 handily beat RNNs.

One of my friends has a Model S and attaches a wrist weight to the wheel when he's on long interstate trips.

It can be difficult to predict whether a given research topic will be useful or not in the future. It's easy when something is developed with an application in mind. It can be far more difficult for "pure math". There are plenty of examples of math that we don't have an application for today becoming very relevant in the future. For example:

- Fast Fourier transform (Gauss discovered and didn't publish it b/c he thought it wasn't very useful. Others rediscovered it later.) - Number theory was considered useless for a long time, now it's used for cryptography

> If Google owns Search, Gmail, Android, Analytics, Chat, Google Fi, Google Fiber, Chrome, GPay, etc... then they are able leverage one monopoly to establish monopolies elsewhere.

You're right that Google leverages its portfolio of services to get a foothold in other ventures, but not one of those listed above is a monopoly. There are solid and well-performing alternatives/competitors to nearly all of Google's services. Their (and Facebook's) online advertising business is another story.

Murphy is a nice book, but I’ve always felt like it was more of a survey text and not one made for diving deep into a given subject. For instance, if you want to go from theory to writing code, Murphy isn’t necessarily the best book for that.

> Scientists believe that time is continuous, not discrete—roughly speaking, they believe that it does not progress in "chunks," but rather "flows," smoothly and continuously.

The jury may still be out on that. A few of the quantum gravity theories (like loop quantum gravity, causal dynamical triangulation, or causal sets) propose a discrete spacetime as a consequence of background independence.

Edit: formatting

During WWII, Germany was also trying to build the bomb. All things considered, it's probably better that the Manhattan Project beat them to it.

Word2vec doesn't really address multiple meanings (polysemy). There has been some progress on this though. Sebastian Ruder has been tracking the state-of-the-art in this here: [1].

[1] https://nlpprogress.com/english/word_sense_disambiguation.ht...

Edit: formatting

(PhD, 2005 in condensed matter)

I left academia after my postdoc and went to work at one of the think tanks in the DC area for a few years. When I got tired of being a government contractor and left to work in natural language processing and machine learning.

Now, I build machine learning software to detect financial misconduct like insider trading. I also built software that helps law enforcement find minors who are being trafficked.

Some of the mathematical and computational methods used in physics are used in machine learning, so it was a relatively painless transition.

I'm also developing some mobile apps in areas where there isn't already saturation.

> Yeah, one way to reduce costs would be to remove all the extraneous coursework that's of zero value to society.

Some of that "extraneous coursework", I'm thinking of things like theather, drama, and dance, brings billions of dollars to the economy every year. For example, Broadway alone generates billions in spending annually [1], with some of that spending going towards the transportation, hospitality (hotel, food) [2], and merchandising industries [3].

[1] https://variety.com/2014/legit/news/broadway-economic-impact...

[2] https://www.americansforthearts.org/by-program/reports-and-d...

[3] https://www.nytimes.com/2016/06/12/theater/hamilton-inc-the-...

(Edited for formatting)

It looks like the US has reversed its position and the 737 Max is being grounded. [1]

[1] https://www.nytimes.com/2019/03/13/business/canada-737-max.h...

Someone could also replace the carbon with silicon.

The Swift for Tensorflow team has added some python interop to Swift. So you’ll be able to, for instance, do an “import numpy” in your Swift code.

The article says the accuracy of the stethoscope’s predictions is 87%. I wonder what the precision and recall are. Precision saves money, recall saves lives.