plafl

I'm not thinking just about AD. I would love lots and lots of little simulation modules instead of monolithic solutions.

Now some uninteresting story: the reason to have a look at MuJoCo is personal: I studied Mechanics I, II and Analytical Mechanics at college and thought non-continuos mechanics a solved problem. When doing my masters thesis about helicopter simulation I needed at some point to model ground contact and I thought "easy, just add a damped spring". I feel that thesis was nice for an undergrad overall but that part as you can imagine was a complete and utter piece of crap. And now at my forties I want to have vengeance.

> it would tell you there was a type error, and then delete the source file

Which makes all the sense because Haskell is about avoiding bugs at compile time

I have a proof that NP(Non Deterministic Polynomial)-Hard = Non Polynomial, but it's too long for a HN comment

I find amusing the fractal nature of the universe: we say that some problems are complex (i.e NP-Hard) because no matter the algorithm you design I can devise a problem instance that will take exponential time for your algorithm (informal explanation). Clearly designing metrics for human is even worse: no matter the metric, people will find a way not only to make it hard to measure, but to make the metric fail. Which takes me to the following strengthening of the Turing test:

A true AI will not only pass the Turing test, it will do so by cheating.

Me neither, and yet I maintain my opinion.

I know, I know, but the source code is not yet available. The github repo is a placeholder for now

Julia is the Haskell of numerical computing.

It seems like AD is a solved problem, which is not. I can personally think two instances in which I have hand rolled my own gradients: in the context of recommender systems (dense matrices are evil) and right now in the context of real time collision detection (memory allocations are evil).

One of the domains where I hope Julia will excel is precisely as you comment on physics + ml (I check frequently if MuJoCo source code is available at last).

Anyway, I encourage everyone with any background in scientific computing or DS to have a look at Julia. The ecosystem is nowhere near Python yet, but the language is very good and the tooling is getting better. The performance Julia provides without interfacing C/C++ or Fortran is not just a convenience, it has architectural consequences. It's not about of coding faster, it's about coding "further".

In any complex field, say data science or software development, any newcomer perceives the solutions to problems they still don't have as unnecessary complexity. So for example the whole idea of dependency management, environments or version control will get in their way. Because it's not yet their problem. They are solving simpler problems: how I run this simple script (or with Emacs, say, undo this change, etc...). In that regard Matlab excels at making it easy for newcomers and harder later.

We need gradual complexity, but I think it's very hard to design. It needs knowledge + didactic capability.

> Numpy is just wrappers for BLAS and LAPACK.

Certainly not. I encourage you to have a look at the API of BLAS/LAPACK to perform basic operationa. And that just scratches the surface.

I think the main problem with LP is philosophical: it is used so you can deeply understand the whole program but most people want to understand a part and ignore the rest. So for example extracting how Tex renders math equations is almost impossible without understanding the whole (some people kindly did this and published a paper using SML). This actually is related to your comment because it's impossible to "navigate" or to "debug" the code, or at least I'm unaware of how could it be done in a simple way (that is, with existing tooling) and so if you want to learn how Tex is written you must proceed page by page of "TeX: The Program" instead of randomly jumping through code, hacking here and there. It's like learning mathematics axiom by axiom. I personally find it unproductive as learning for me is highly non sequential. If knowledge is a graph of concepts I want to explore as I wish, not following another's predefined serialization.

It requires a lot of skill to write TeX, but the key is declaring it feature frozen for almost 40 years. I like Knuth's approach to open source: "Here is the code, I wrote it because I needed it, I hope you find it useful. Now, if you don't mind I have things to do. I no longer use email but I read snail mail every six months"

Was "Señor Developer" intentional or an auto correction of some spanish keyboard? Anyway I love it and I'm thinking of keeping it for myself.

If it compiles your predictions are bug free

Is there so much demand in the US that it will impact global wages? I'm not sure.

As a foreigner is not in my interest but I do wonder if, just for the sake of coherence, shouldn't the import of labor be taxated in the same way goods are taxated.

>Unfortunately you are never going to see a project with comprehensive documentation, I'm not sure what that would even look like.

It probably would look like Tex, FWIW.

Edit: I move to the bookcase to look at TeX: The Program. There it is sitting next to my volumes of TAOCP. A reminder of my failures. I can almost feel the disaproving gaze of D. E. Knuth. I'm not worthy.

> Transaction fees kill this from a practicality point

That depends on the amount. If it's small it could be a feature.

They play a role in a different sense: if you see a rifle it may be there because it's going to be fired or maybe because the character likes rifles. Could you imagine The Big Lebowski without bowling? And yet bowling has nothing to do with the story.

Quite dull. That's why I love The Big Lewobski. It's full of irrelevant details. There is a story too but the movie it's about the little things.

Very nice. I have somewhere the book Airship Technology, which I bought when I was curious about the possibility of blimp drones (I was an aerospace engineer some time ago).

I'm glad you are considering smaller blimps, I think previous projects have been too ambitious. Since I have worked on drone navigation I'm curious about your approach in that regard and how you are going to attain autonomy and get it certified. If you manage to get it a completely autonomous and certified that alone will be an asset even if you pivot away from blimps! Are you going to use alternatives to GPS? GPS is very precise when it works but I doubt you can get certification based on that alone.

Python definitely started with very good ideas and whatever success it has achieved is by merit and not marketing power or reputation of a company behind. Clean syntax, a decent object model, good documentation. Good taste in general.

But it's older than Java, think about that. Has it aged worse than other languages? Are Python libraries worse than Java's? Package management? Has Python more cruft? Is it harder to install?

At this point it's part inertia and network effects, although you should check the alternatives in ML: R, SAS, Matlab... I think Julia is better, but it has made some decisions not everyone likes. Static typing for ML, you may think, that would be awesome. Well, I think the language that got closest to it was Scala but it never escaped Spark.

If you want a gentle introduction you may like this book http://the-little-prover.github.io/

I could have written your comment. I sometimes think that it's not that children learn faster, it's just that they are used to be totally lost and keep trying

I'm sure there is so much room for improvement here. The closest thing for me is using TeXmacs: https://m.youtube.com/watch?v=H46ON2FB30U

This is why public health insurance is so important and why I feel so lucky to be covered by one: my brother had brain hypoxia during birth and even in a country (Spain) with a solid public system it places considerable personal and financial strain on our family.

I don't know about your legal system but in Spain at least it must be proved by the insurer that if you lie about a previous condition that lie is directly related to the payment they are trying to deny you.

I'm extremely happy with Julia performance. I'm using it right now to prototype some new algorithms for my PhD.

How fast is fast? Well in my case I'm trying to come up with things related to collision detection and I'm getting <30ns using SAT between OBBs (Oriented Bounding Boxes) and <300ns for minimum distance computation using the GJK algorithm again between OBBs.

In the context of recommender systems I did an experiment when I started learning Julia comparing against Cython and wrote about it, although I was a newbie and I should update it: https://plopezadeva.com/julia-first-impressions.html

I once used clang to extract all input/output structures of a big C project. In my case was about generating tests. It's very satisfactory when the computer saves you so much time.

Not for GPUs AFAIK but still related: https://www.blosc.org/pages/blosc-in-depth/

Google gives me this: https://m.youtube.com/watch?v=rgwNsNCpbhg

I will check it at home, I have started my PhD recently and I find the subject very interesting. It got my attention after reading some papers (i.e. DeepSDF). One of the best pages on the subject is Iñigo Quilez's blog: https://www.iquilezles.org/index.html

Do you have any recommendation for a book? I studied classical thermodynamics (and physics) long ago and I consider I have good probability and statistics foundations but never studied statistical mechanics. From my ignorance I cannot understand how is it possible that entropy considerations (macroscopic) are applicable to a microscopic demon.

I'm Spanish, not Italian, but we also talk quite loud. Maybe all discussions look heated to foreigners.