xitrium

Can you move to a city? This is what most people I know in this situation do. Though I had a great time getting a car and taking myself out for hikes, sauna / spa days, activities and parties in the east bay near SF. Great place for practicing being alone. I had to think about it like dating myself - where would I have taken a date for fun? Try a bunch of things and see what sticks and remember you can appreciate moments by yourself with this mindset and it's like 80% as good.

Came here to say this. Pure slop, wtf is everyone going on about? Are they all LLMs too?

Very true, but it seems like you may not have seen the expose on this field in Science https://www.science.org/content/article/potential-fabricatio...

I appreciate your commitment to modernist capital-S Science here :) I'm familiar with how the field ought to work but after working in Andrew Gelman's lab for some years, also with how it can fail us. Here I think the researcher in question has had a much larger impact than you are allowing for. Here's a choice quote:

Every single disease-modifying trial of Alzheimer’s has failed.

The huge majority of those have addressed the amyloid hypothesis, of course, from all sorts of angles. Even the truest believers are starting to wonder. Dennis Selkoe’s entire career has been devoted to the subject, and he’s quoted in the Science article as saying that if the trials that are already in progress also fail, then “the A-beta hypothesis is very much under duress”. Yep.

And the original expose is quite interesting if you haven't read it yet https://www.science.org/content/article/potential-fabricatio...

The last thing I read about the link between amyloid-β accumulation and Alzheimer's was that the entire field was full of fake data ( https://www.science.org/content/blog-post/faked-beta-amyloid... ). In particular, even treatments that directly reduce amyloid-β in the brain did not restore cognitive abilities.

At least this paper tests both cognitive abilities as well as "amyloid-β pathologies." I'm not at all an expert in this field but gold nanoparticles sounds like something you'd see on a late night infomercial, lol.

The current law is more general; it’s the current policy’s consumer price heuristic that has become a bad approximation to the law. I like “The Economists’ Hour” on the topic.

Statistical Rethinking is the first book that helped me make sense of statistical modeling (and probability as applied to modeling): https://xcelab.net/rm/statistical-rethinking/

Huge fan, can't recommend enough.

I will say up front that I don't think the social good is worth what we are collectively paying for it, but I do think the market hours are a reasonable device. This is basically because there are humans involved and they need to sleep (Matt Levine has written about this).

If you want the best price, you need to have all of the market participants bidding together. Market hours serve as a coordinated period in which ~all market participants agree to be online and bidding. Prices, thus, get stale overnight. But we assume that that is mostly okay, as business is normally conducted during business hours, and we assume that transactions can wait until the next day. ACH transfers take multiple days! (technically so do stocks, but that's mostly invisible to retail traders).

If you're a retail trader, I would caution you somewhat against trading after-hours; there is very little liquidity and it could cost you 100s of bps more.

I like "The market can stay irrational longer than you can stay solvent." -Keynes

Unfortunately, the rethinking package which is a major component of the book itself depends on the V8 engine for some reason.

This is my fault, in a sense. In order to get the new Stan compiler (written in OCaml) distributed via CRAN (which requires everything to be built from source on its antiquated build servers), we decided to use js_of_ocaml to translate the OCaml compiler into javascript. See this thread for more details: https://discourse.mc-stan.org/t/a-javascript-stanc3/11044

When I posted that, I didn't really think we would end up using it.

This is weirdly deliberately misquoted, maybe as a joke? full: https://statmodeling.stat.columbia.edu/2013/07/21/bayes-rela...

I'm surprised to see this argument. Tooling and infrastructure are only as clean as the services they support. I don't think you get to wash your hands because all you did was build e.g. Palantir a giant database that's great for storing locations if you know customers will be assassinating political dissidents with it.

This did not hold up well, imo. Not sure how to count it but by some lists this is stuff like: Tencent, Alibaba, Amazon, Netflix, Priceline, Baidu, Salesforce.com, JD.com.

Others would include Uber/Lyft, Airbnb, GitHub, ...

There are only a few that would qualify that I can think of - Instagram, Snapchat, and WhatsApp.

If folks here have other examples or thoughts on why this does or doesn't hold true would love to hear them.

How do you feel about Alberto Balsam? :)

Thanks! The new .messages error message mechanism is great, though we'll see if it supports the full brunt of user needs soon enough.

Curious if the author could comment on what Menhir is lacking, my estimation is possibly just the concrete syntax trees, though you could definitely emit an AST that has as much of that as you need to do the tasks the author mentions.

Be cautious using this article to try to learn anything. Differentiable programming is not actually related to deep learning; it's another word for automatic differentiation, a technique that is very important in deep learning implementations but useful for a variety of other tasks where having gradients available for arbitrary functions is useful.

The article is correct that "Differentiable Programming" seems to be a rebranding effort that I believe just helped automatic differentiation work from the machine learning world get published in Programming Languages journals. I wouldn't read too much into it.

Gentle plug for my literate Haskell implementation that aims to teach the concepts as they are defined: https://github.com/seantalts/hasktrip/blob/master/doc/MicroK...

GPU and MPI support in Stan coming hopefully this summer!

Bob Carpenter did a great case study on the this topic, for anyone new to it: http://mc-stan.org/users/documentation/case-studies/curse-di...

Check out the graphs, especially of the euclidean distance from the mode of a random point as dimensions increase.

If you care about quantifying uncertainty, knowing about Bayesian methods is a good idea I don't see represented here yet. I care so much about uncertainty quantification and propagation that I work on the Stan project[0] which has an extremely complete manual (600+ pages) and many case studies illustrating different problems. Full Bayesian inference such as that provided by Stan's Hamiltonian Monte Carlo inference algorithm is fairly computationally expensive so if you have more data than fits into RAM on a large server, you might be better served by some approximate methods (but note the required assumptions) like INLA[1].

[0] http://mc-stan.org/ [1] http://www.r-inla.org/

Stan (mc-stan.org, Columbia University) | New York City | Full-time, ONSITE | $80-$100k

Stan (http://mc-stan.org) is an open source probabilistic programming language and Bayesian inference toolkit that data scientists and applied statisticians across the world and many fields use to specify statistical models and fit them to data. Stan emerged from the research group under Andrew Gelman at Columbia, which is focused on state-of-the-art in Bayesian inference methodology and consulted on a variety of statistical problems in the social, biological, and physical sciences, engineering, sports, and business. Since its initial release, we’ve welcomed collaborators and contributors across the globe and have about 30 active contributors at the time of writing.

We’re looking for a tools and infrastructure developer who is interested in science and open source, is a self-starter and intrinsically motivated to help out, wants a flexible lifestyle, and would like to learn Bayesian stats and data analysis. Ideally you’d come in knowing a decent amount about either systems administration or programming (or both) and we would teach you about statistics. We’re looking for help with the following problems initially: * Streamlining installation for our R and Python interfaces for scientists on Windows and Mac, perhaps creating installers for each platform * Maintaining and improving our continuous integration infrastructure (currently Jenkins, Travis, AWS) * Bringing the PyStan interface up to feature parity with RStan and improving both of them (Python) * Higher order autodiff test framework and infrastructure (C++)

We expect that you will need to spend a significant portion of your time on at least some of these problems at first. Beyond that, there is much work to be done on Stan as an ecosystem. For example, we could also use help with the following key underserved areas of the project: * Profiling and performance benchmarks and improvements * Various refactorings, mostly in C++ * Pedagogical materials * GPU and distributed computation support

There are also many other ways to improve Stan that could use helping hands: * Stan 3.0 language and/or compiler rewrite * CloudStan * Graphical modeling language transpiler * New algorithm research implementations * Anything from the list here: https://github.com/stan-dev/stan/wiki/Longer-Term-To-Do-List

Links to some Stan team outputs and other Stan-related resources: https://github.com/stan-dev andrewgelman.com http://mc-stan.org/users/documentation/ https://github.com/stan-dev/stancon_talks great intro Bayesian analysis book: http://xcelab.net/rm/statistical-rethinking/

To apply, email Sean Talts, contact info here: mc-stan.org/about/team/

What idiot called this "Feasibility of cooling the Earth with a cloud of small spacecraft near the inner Lagrange point" instead of "Solar Eclipses As A Service?"

Going along with this sort of fundamentally changed worldview theme - I loved Debt: The First 5,000 Years but for me it was equally important to read a book called The Worldly Philosophers: The Lives, Times and Ideas of the Great Economic Thinkers by Robert L. Heilbroner. This book summarizes the entirety of major paradigms in the history of economic thought, complete with great little biosketches of the authors and their times, from Adam Smith to Joseph Schumpeter. Highly recommended, especially as a counterpoint if you are already a little much in the David Graeber camp as I was.

You probably weren't a culture fit :P

I actually prefer getting feedback about why someone chooses not to go forward with me as it helps me improve. For that reason I think interviewing at Bridgewater can be valuable as you can explicitly ask them what you did right and wrong and they'll tell you without sugar coating it.

And don't forget, the reason most companies don't give this kind of feedback is to prevent any allegations of discrimination, so they're not exactly heroes either.

In actual compiled code, conditional branches (without branch prediction) are translated to jumps to different targets, which are specified inline with the instructions. Specifying a modifiable target would mean fetching it from a register (or worse, memory) and delaying execution until the fetch is complete (several cycles minimum on a pipelined machine). With branch prediction, instructions are predicated on a condition inline and we avoid the costly jump instructions.

Read more: http://en.wikipedia.org/wiki/Branch_predication EDIT: Also: http://en.wikipedia.org/wiki/Branch_predictor

I think we more commonly use the latter, which tries to guess which way the code will branch and load the appropriate jump target. It's actually typically very successful in modern processors.

I like that you link to the open source software involved.