rxm

Nice project. I’m curious to see how it writes after instruct.

What used to be feature engineering a decade or more ago now seems to have shifted to developing distributed representations. LLMs use word tokens (for words or the entities in images). But there are many more. The 3D Fields (or whatever they have evolved to) developed by Fei-Fei Li's group represent visual information in a way better suited for geometrical tasks. Wav2Vec, the convolutional features for YOLO and friends, and these sentence representations are other examples. I would love to read a review of this circle of ideas.

I also prefer to number just the important equations. One of the advantages of numbering all equations is that the reader can binary search an equation by number. I try to mitigate that loss by providing page numbers or section numbers.

Providing “contrast” is something I appreciate as a reader. It is much easier to provide in person, on a whiteboard. Maybe we need papers to have references to YouTube derivations :)

One of my pet theories is that flimsy homes reflect an economics signaling problem: it is hard to prove to a buyer the value of quality. Whereas size is easier. As a suburban home owner, I would have preferred the quality-square footage point to lean a bit more towards quality.

I also like Contact, which I think of as Sagan’s struggle with reconciling religion and science for himself. Also, I nlike what the article mentions, I recall the book being based on the script.

I got through graduate school and over a decade of research with my inexpensive but very smooth Uni-ball cylinders (or classic roller) micro, 0.5mm pens. They were light, with a minimal and elegant design, inexpensive, and wrote great. Then, about a decade ago, something happened to them. I now go between a Jetstream (a ballpoint) and the Uni-ball Eye (which feels identical to the Vision to me).

I would have included symbolic algebra systems to the list. Many famous calculations in physics and mathematics may not have been attempted without them. Macsyma and Reduce paved the way for Mathematica and Maple. There are still many symbolic algebra algorithms waiting to be implemented. The love and care many open source symbolic systems are receiving is a reflection of their need and importance to science and technology.

Amazing things can happen when cellular automata have symmetries. Conserve number and momentum and you're a few steps from the HPP gas [0]. Make the momentum flow a bit more isotropic, and you are on your way to the FHP gas [1]. The history I remember is that Hasslacher and Wolfram had overlapped at the IAS early 1980s. On the summer of 1985 Hasslacher visited Frisch in Nice. It was there that they guessed that the hexagonal lattice might brake the strange conservation laws of HPP. After that it was a few days until they had a set of rules for the lattice gas fluid in 2d. From France Hasslacher called Shimomura at Los Alamos and asked him if could implement them. In a few days Shimomura has a crowd staring at his screen watching a simulation of a fluid flow past an obstacle shedding vortices in real simulation time. We can do that today with our smartphones, but at the time it must have been a first.

[0]: Hardy, Pomeau, & Pazzis (1973) JMP: 10.1063/1.1666248

[1]: Frisch, Hasslacher, & Pomeau (1986) PRL: 10.1103/PhysRevLett.56.1505

Some URL shorteners provide a copy of their mappings to the Internet Archive as a promise that if they stop functioning for any reason the archive will continue to provide the mapping (and make the mapping file available).

[0] https://archive.org/details/301works-faq

Until the early 2010s we ran our own tiny data center (six to eight racks in two separate locations). When we re-evaluated whether to have the racks hosted somewhere else, it worked out way cheaper to run the machines ourselves. But when we included the costs of the sporadic stuff, cloud providers became cheaper.

The sporadic stuff adds up. This is the "other stuff" one needs to do to keep the machines running. It includes the time it took someone to swap a failing drive, or the loss of productivity when the system went down (because of faulty air conditioning maintenance procedure), or the costs of procuring hardware, or dealing with the network provider, and so on. This stuff adds up. We ran a tight ship. Every minute someone spent dealing with the racks of computers was a minute they did not spend creating a product for our customers.

The only thing we had some difficulty with was intermediate reliability storage, at the level of non-raided spinning disks. But that was then, when the offerings where thinner.

any Internet user who uploads a copyrighted work could find themselves subject to a largely unappealable $5,000 penalty

Shared a viral amateur video? I understood that this law would allow a troll firm to contact the video creator and offer to sue you and all those that did so.

In most organizations I have interacted with getting data with known sampling is always a problem. “The temperature sensor sends a measurement every 5 minutes.” Except, I learned after a lot of debugging, when the air-conditioner gets maintained and the readings become a fixed value.

It’s not the organizations fault. Discovering and maintaining a known sampling for the data is part of the process. I consider it a win when I can get a team to accept that the data process will need to be debugged just like code.

There are many aspects of a library of technical books that I miss online: the curation process, the higher bit rate once you are holding the book, and the lack of distractions in the library.

What I fear may go away is the long form books offer. When exploring a new topic, articles are often too dense. In books, authors can establish the background, help create a shared vocabulary, and motivate what is often so condensed in a research article.

What awk and send are to files of lines, jq is to json. I find it indispensable on the command line. Jq is a functional language. It took me a bit to get used to it, but now it is one of the first tools I install in a new machine. The linked article is a curated list of jq related resources. After the manual, I found the cookbook helpful.

One of the conversations was about reading. Let the kids read anything: comic books, fun novels, fantasy picture books. He re-assured me that they would eventually expand into other things. Martin was right.

When I was a graduate student, in the late 80s, I was waiting for a flight to Los Alamos in Albuquerque. There was only one other person in the lounge, obviously also going to Los Alamos. I started chatting and he introduced himself as Martin Kruskal. Idiot me had no idea whom I was speaking to. I said “The only Kruskal I know invented this amazing card trick having to do with dynamical systems.” He stared at me, obviously pleased. “Not many people know my brother and I are into magic.” He did not want to take credit for the trick and attributed it to his brother. He would come often to Los Alamos and I miss chatting with him (got lots of advice on bringing up girls).

Thanks for the example. What you just did is one of the hardest things to communicate about data. A casual user would never notice this. I find "data curation" doesn't always get the importance it deserves.

I still use textmate. I started with vi, which I still use for small edits, switched to sam, back to vi when I stopped having X on my Mac, and then to textmate. I have tried sublime and atom. I am sure they can be configured to work as I would like, but out of the box the distance seemed too far.

My desire to configure is strongest when I start writing reports (in LaTeX or markdown). I long for a distraction free mode with easy spell checking.

Does anyone know when this post was written?

You are so right. Good design doesn't just pop out of nowhere. It does emerge early on, but requires work and time while the complexities of the problem are groked.

Hardy's rule can be adapted to programming: to code it better, think it gooder. One big difference in productivity comes from making good design decisions early on and not taking shortcuts that weeks later turn into spaghetti hacks.

Nice read on how to check your rsyncs.

The mechanics of creating a manuscript are not the big problem. Meaningful and honest peer review, assignment of credit among authors and from future developments, better matching between manuscripts and readers are some of the things that I would rank higher.

It seems to me that art and science need to go together. Places rich in technological development are also great art, culinary, literary, ... hubs. To feed the cutting edge you need crazy analogies that only diverse fields provide.

Thank you. These, and the other answers, are great links. The Pearls book seems the right thing to get you thinking functionally.

Functional data structures are another beast altogether. Okasaki's book is the only one I have on the topic. Any other suggestions?

And many of these kind-of-free tools are built on top of open source code. I also feel that Open Source movement is changing.

I wonder if security patches will be free.

I was on a slow connection, so it took some time to download. This appears to be a draft of a Mathematical Methods for Engineers textbook. Its 502 pages start out with multivariable calculus, then go through lots of linear algebra, and end with dynamical systems.

The timing of this post is uncanny given how some large crowds voted.