Homunculiheaded

Have you looked into remote work? I love Reno and have been remote working here at several different companies for many years. Right now, it's fairly easy to get a good front-end web dev gig remotely that pays much closer to SV salaries than Reno ones.

Reno has really blossomed even in the last few years, but the jobs market, especially for skilled people, is abysmal. There's a bit of a chicken and egg problem as well since many people with enough talent realize that Reno's salaries are laughable and end up leaving. Right now there's not enough decent paying work to attract people to the city, but if you where to bring a company here that paid sane wages you'd have a hard time finding talent. I've known enough amazing UNR grads that migrate to the Bay to know that this city does have the potential. It's just a matter of that right window of a reasonable paying company being here and snatching up enough bright people before they move to the Bay.

But for remote work it's hard to think of a better place. Cost of living is very low, there's no state income tax, and SF is an easy 3 1/2 hour drive when you miss parts of the big city experience. Every other major West Coast city is a cheap and quick flight. And there are some really amazing people in this city. If you don't go already, head to Hack Night at the Reno collective some time, it's a great group.

Anyone interested in Logic and Probability should take the time to read through (at least) chapters 1 & 2 of Jaynes' Probability: the Logic of Science [0]. Jaynes' is the arch-Bayesian and in these chapters mathematically develops what is essentially an alternate Universe model of probability which, in his view, arrives as the natural extension of Aristotlean logic. There's no "coin flipping" in these chapters, and when he finally derives the method calculating probabilities the fact that his model matches with coin-flipping models is written off almost as a happy accident. If you're familiar with Bayesian analysis but have not read Jaynes it is very likely that you aren't familiar with quite how (delightfully) extreme his views are.

Jaynes' fundamental metaphor through the book is building a "reasoning robot" so anyone interested in the intersection of logic, probability and AI will get many interesting insights from this book.

[0] PDF of the preprint: http://bayes.wustl.edu/etj/prob/book.pdf

Many of the examples in this post are great examples of 'Modernist Art' and are decidedly not postmodern. This is roughly the equivalent of writing a post on "Functional programming is Anti-Mind" and then demonstrating that with examples from the Gang of Four Design Patterns book.

The real issue with this is that most postmodern art is incredibly accessible. You don't need an art degree to think that Roy Lichtenstein's paintings "look cool", or that Campbell's Soup cans are "neat". One of the quintessential, textbook postmodern film directors is Quentin Tarantino; there are few directors more adored by the general public. Postmodernism is a descriptive term for artists who mostly reject the Western tradition of 'High Art'. Almost all the difficulty and "unintelligibility" lies in postmodern theory, but not in the art theorists consider postmodern. And I would argue that this is because theorists themselves are artifacts of Western high culture and are therefore unable to articulate a response to something that is outside this framework.

Almost all examples of "unintelligible" art fall into some subcategory of High Modernism. High Modernist schools of thought are almost always exploring questions within the context of Western high culture(ie the "What is art?" questions), and for many of these works you need to have a background in the art in question to really engage with and understand the piece.

If you want to critique postmodernism a good place to start is Fredric Jameson's "Postmodernism or, The Cultural Logic of Late Capitalism".

I've worked remote for quite awhile now at a pretty broad range of companies. For the jobs that have had remote teams and local offices I do agree that I'm able to get a lot of communication done quickly when I visit the office.

However I find that the amount of "heads down" work I get done is greatly diminished when I'm in an office. And, much worse, there's a lot of noise in that added communication of being in the office. Remote teams, in my experience, have dramatically less "office politics".

Office space is great for communicating "big ideas" but these aren't anywhere near the bulk of communications being had. For most of the communications needs of software remote works fine (in my experience better).

I work on a quite a few "big idea" projects and I've found the best solution is to visit the office quarterly, get all the big idea brainstorming done, then scurry off to my remote office where I'm not distracted by office politics and can just get things done. A little face time goes a long way, and annual, or semi-annual all hands meetups can do wonders at filling in the gaps created on remote teams.

My experience has been that none of the major Deep Learning libraries (Theano, Torch7, Caffe) offer support for OpenCL, whereas they all make it trivially easy to get models running on a CUDA GPU. On top of that NVIDIA has a library of deep neural network primatives[0], and I don't believe AMD offers anything similiar.

The general consensus I've seen is to just get an NVIDIA card if you're serious about working with deep neural nets on the GPU.

One thing that did surprise me was that there was no mention of using EC2 GPU spot instances for getting your feet wet. If you don't have access to a GPU with CUDA support you can get a spot instance for about $0.07 an hour to at least test out that you have your GPU code configured correctly (and you will see some performance gains). There are even a couple of AMIs out there with Torch7 and Theano already installed.

0.https://developer.nvidia.com/cuDNN

I have a running joke with my machine learning friends that I will write a Data Science/ML book titled "A Thousand Ways to Say 'Singular Value Decomposition'". The number of papers and techniques out there that are SVD with a few minor tweaks and a unique philosophical interpretation of SVD is hilarious.

Here are some examples:

Principal Component Analysis - SVD does dimensionality reduction where some n% of variance should be accounted for.

One layer Autoencoder - SVD done by a neural network

Latent Semantic Analysis - SVD on td-idf matrix we interrupt lower dimensions as having semantic importance

Matrix Factorization - SVD only now we interrupt lower dimensions as representing latent variables

Collaborative Filtering - SVD where we interrupt lower dimensions as representing latent variables AND we use a a distance measure to determine similarity.

> computers are good at some things, humans are good at others

"You insist that there is something a machine cannot do. If you will tell me precisely what it is that a machine cannot do, then I can always make a machine that will do just that!"

-- J. von Neumann

Computers will continue to get better at human things as we continue to get better at understanding how human things work. Look at the recent advances in deep learning. This is using only the most crude approximation of human neurons we can identify and caption images with astounding results. Google currently claims that anything that can be done in 0.1 of a second by a human, they can do as well.

Fraud detection relies heavily on unsupervised learning, and for all of history up until the last few years state of the art unsupervised learning was usually SVD + clustering or some variation on that. The current state of the art, things like deep belief networks, are able to achieve markedly superior results.

Additionally this article seems to imply that they are collected labeled data from customers which should help tremendously in modeling fraud. If even if the labels are a small sample recent advances in semi-supervised learning using deep neural nets is even greater than the advances in unsupervised learning.

While I don't disagree that historically it has been wise to include a human element in fraud detection, I don't believe there is any reason to assume that trend will continue indefinitely into the future.

One thing that this article fails to point out is how often librarians were wrong before Google. In 1986 there was a study[0] that showed that across the board reference librarians were only correct about 55% of the time.

I think most people today would consider a query answering system that had an accuracy of 55% to be an interesting curiosity, but certainly not ready for real-world application.

It's funny how frequently we measure machine learning performance of an "easy for humans" task and fail to compare it to human accuracy on the same data. We just assume humans would do perfectly on it. I'm sure there are a few MNIST digits that I would get wrong.

[0] P.Hernon, C McClure "Unobtrusive Reference Testing: The 55 Percent Rule," Library Journal, 111 April 15,1986

It's always worth it to put your feelers out. Even for jobs I love I'll occasionally respond to recruiters just to keep up with the market and see what's out there. Often this will confirm my happiness with my current role.

But if you're on the fence at all, there's a really good chance you can find another job that has better pay and even better product, coworkers etc.

Interviewing when you're not desperate to leave is much easier. If you have something good to fall back on, all sorts of red flags become more obvious during an interview because you have very little to lose if it doesn't go well. It's also much easier to negotiate salary, if you think you're worth 20k more than you are now, open negotiations at Salary + 30k. Since your primary concern is pay potential employers have very little leverage in terms of pressuring you to take a salary less then you think you are worth, and you already have a clear picture of what you're worth.

I really dislike these numberphile videos as they deliberately construct their arguments such that the answer makes math seem like a magic trick and that math is truly confusing, which is the opposite of helping people gain an intuition around math.

The trick here is that they start with the intuition of counting, in which each number is itself a single thing, but are actually doing a calculation based on each number being a string of single digits.

It's much less shocking if you say: Picking a random 10 digit number is the same as randomly picking 10 single digits. The more digits you pick the more likely it is that you'll get a '3' somewhere in there. So as your string of numbers increases in length the less likely it is that you're string won't contain any given number. If I said to someone "The more dice you throw the more likely it is you'll get at least one 3", I don't think I would get anyone who was surprised by that.

The unfortunate thing is that math and number theory in particular are full of genuinely fascinating observations that don't rely at all on a tricks of phrasing to be revealed, and rather than making less math literate people feel 'dumb' (as these videos tend to do), spark in interest in exploring math further.

Tails answers this question on their site [0]

Their general opinion is that this makes Tails less secure as now you have to trust both the host and the visualization software.

Provided that you trust Tails itself. The expected way to run it is off of a live cd. This way the trusted OS is only ever in ram, and if you use a non-rewritable disk you can also be assured that the Tails disk itself cannot be modified after its creation.

Tails handles the 'only talking to Tor' via iptables. Unless I am mistaken Tails' firewall will not allow clearnet connections.

[0] https://tails.boum.org/doc/advanced_topics/virtualization/in...

To be fair to Kahneman by 2012 he did come around and recognize that there were serious issues with "priming" research[0].

The chapter on priming in "Thinking Fast Thinking Slow" completely ruined the book for me. The very premise of the book is (paraphrased): "I will teach you to overcome bias and think rationally" and then here is an entire chapter that should make anyone with a critical eye strongly question the research. Even if one is ignorant of the issues these experiments have with reproducability the experiment design itself is terrible, and there seems to be no way to draw the wild conclusion that are drawn. Not to mention that if priming actually worked as claimed you would see it become a hotter topic among marketers than SEO. I fail to see how a world expert on cognitive bias can fail to question such an obvious fault in a topic that he covers in his own book.

[0] http://www.nature.com/news/nobel-laureate-challenges-psychol...

I'm surprised that they didn't mention Leo Breiman's famous paper "Statistical Modelling: The Two Cultures"[0]. A worthwhile read for anyone interested in the topic. I could sum it up, but the abstract does a better job:

----------

Abstract. There are two cultures in the use of statistical modeling to reach conclusions from data. One assumes that the data are generated by a given stochastic data model. The other uses algorithmic models and treats the data mechanism as unknown. The statistical community has been committed to the almost exclusive use of data models. This commitment has led to irrelevant theory, questionable conclusions, and has kept statisticians from working on a large range of interesting current prob- lems. Algorithmic modeling, both in theory and practice, has developed rapidlyin fields outside statistics. It can be used both on large complex data sets and as a more accurate and informative alternative to data modeling on smaller data sets. If our goal as a field is to use data to solve problems, then we need to move awayfrom exclusive dependence on data models and adopt a more diverse set of tools.

[0].http://tuvalu.santafe.edu/~aaronc/courses/5352/readings/Brei...

"post-scarcity" is an illusion presented to people living at the very top of the economic pyramid. I'm sure most HNers are in the top 90th percentile of US household income on their own, living in a world mostly surrounded by other affluent individuals working in a similar field, all in jobs with increasingly flexible work hours, better benefits and nearly unlimited job prospects.

The world only looks like it's moving towards post-scarcity because the resources consumed and waste produced have been nearly completely outsourced away, and the pangs of the labor issues created by automation are happening somewhere else.

Post-scarcity is not an issue of 'productivity' but of resource use. As Jevons paradox [0] points out the more efficiently we use resources the more of them we tend to use. Looking merely at productivity and efficiency without looking at resource use and waste production gives a wildly incomplete picture of "post-scarcity". What I see is that the world is increasingly separating into a small sub-population that is presented the illusion of post-scarcity and a another, growing, subset that is feeling the effects of scarcity required to maintain that illusion.

When I see data to support the argument that resource use and waste production are plummeting, then I'll believe in narratives about post-scarcity.

0. http://en.wikipedia.org/wiki/Jevons_paradox

My suspicion is essentially the opposite: Tor is secure, but the two high profile arrests (Freedom Hosting and Silk Road) where given priority to make the general public a.) feel that the entire function of Tor is illegal and often repulsive activity b.) that Tor is not safe.

The latter part of that theory, that law enforcement agencies intentionally stepped up the resources for both the FH and SR cases in order to intentionally create disgust and distrust of Tor, is of course merely conjecture. Basically I find it an amazing coincidence that the two most notorious parts of the Tor hidden service world where busted very quickly after a huge amount of positive public attention was brought to Tor right after the Snowden leaks. Additionally if you actually look at the details of the FH exploit the FBI unleashed it is fairly useless, but very terrifying when you read just the headline. Legally there seems no useful reason to use such an easy to discover exploit that would have delivered no particularly interesting information. However from the stand point of creating public fear it worked marvelously. If you talk to even technical people that don't understand security and Tor well they often assume that the feds "hacked Tor". Which, in my opinion, is exactly what state actors want people to think.

As for the former part of the claim, that Tor is secure, look at the Snowden leaks about the methods that the NSA was thinking about for attacking Tor. Egotistical Giraffe, the attack used on FH, as mentioned was not a particularly useful exploit, and attacks user behavior not the network. Other similar leaks also suggest that neither the NSA nor any other state agency, has the ability to completely compromise Tor.

Finally,if you are a state agency and you have completely compromised Tor, you would actually want the general public to think it is safe. It is an amazing advantage to have your adversary think they are on a secure line when they absolutely are not. On the other hand if you haven't (and probably can't) compromised Tor you want the majority of people to think you have so that they disregard one of their best tools for defense.

Now of course there is plenty of evidence that federal agencies can perform targeted timing attacks against specific individuals. Tor does not and really cannot guard against this, and this has always been the case and fairly well known. If a state agency is targeting you specifically, I don't think there is anything you can do. However, given the information that is available to us, I do think it's reasonable to assume that Tor is secure from general, large scale, untargeted surveillance.

I would highly recommend reading Cryptography Engineering [0] cover to cover. It's amazingly readable, covers the basics, the theory necessary to understand how things works and includes ample practical advice and observations on the industry.

The first thing I did after the Snowden leaks was read through the entire thing and after doing so I really wished I had done this years earlier. There's very few books that I think should be required reading across the board for software engineers, but this is one that I do think everyone writing code should read every page of.

[0] http://www.amazon.com/Cryptography-Engineering-Principles-Pr...

I was interested in Buddhism for awhile, and even spent a fair bit of time reading through translated segments of the Pali Cannon.

I've seen a critique of Job's "Buddhism" discussed quite a bit, but it's important to consider that the soteriology of Buddhism and Christianity is radically different.

A major difference is the time scale and way with which salvation occurs. Christianity focuses on the realization that Christ is the savior, and appropriate adjustment in behavior in a single life time. Buddhism on the other hand is a progress in realizing the dharma over many lifetimes of effort.

Paul's conversion happens instantly on the road to Damascus, his change in action is immediate. The salvation of a bodhisattva occurs over the progress of many lifetimes effort, each new lifetime being shaped by the actions and desires of the old.

The point being that the binary saved/damned dichotomy used for judgement in Christian theology (ie X is a bad Christian) doesn't make any sense in a Buddhist context. A "bad buddhist" makes no sense as each person is simply working out the consequences of past actions during each successive lifetime until the dharma is finally realized. (edit: just a note that this 'progress' itself is a massive over simplification itself)

My personal opinion on this issue is that we already have a pretty good sense of public/private spaces in the physical world. The problem is that most people assume that if your computer is in a private space (ie your bedroom) then it is also private. This is obviously not the case.

What we need is to start building a culture of online privacy. Everyone should have access to an anonymizing VPN, should know how to use Tor and understand PGP. But just like it would be insane to never leave your bedroom and unlock your door, or always speak in a whisper, the same applies to online privacy.

And the important thing is just because you are in your private space doesn't mean you're doing anything wrong or shameful. People use private spaces to snort lines of coke and plan bank robberies, but also just to have some time when they can think and not be bothered... or watched. People need to start using Tor just to browse the web and know they aren't being watched, that neither the government nor advertisers are building a profile on you.

And just like physical privacy means less social, so does online. Online privacy should be about keeping anonymous. Don't talk about where you live, what your hobbies are etc. Be conscious of not leaving a trail of personally identifiable information (just like you close the blinds in your bedroom).

It's not a question of always watched or always hidden, but being conscious of when we are being watched, and when we are free to say and do as we please.

To be fair though a huge difference is that the deaf community is bound together by a common language. Language is hugely important in human social groups. A "free fix" for members of this community comes at the cost of the eventual annihilation of their shared language. This is why you typically don't see the same reaction from medical advancements in vision from the blind community, whose impairment has no effect whatsoever on their primary language (and just to be clear Braille is simply a character mapping to the readers native language while sign languages such as ASL are actually distinct languages).

I've personally found a great alternative to living in major cities is not the suburbs but smaller cities. I grew up in the Northeast and now live in a smaller city in the West (pop ~250,000). I can get to SF in a few hours, but locally I still have most of the things I'd want from a city: great restaurants, art, fun tech community, music, easy to access airport etc. I also have many of the perks of living in a smaller community: very little traffic, incredibly affordable cost of living, get to know the owners of most restaurants/businesses I frequent etc.

The crazy thing is that living in a house 10 minutes from the center of downtown I pay a fraction of what my friends and relatives back East do to live in a suburb so far away from the metro area they live near that they visit it only once or twice a year.

Especially with more and more remote work becoming available, if you're getting sick of living in a major metro area I highly recommend checking the diverse range of smaller cities across the US, imho it is a vastly superior experience to living in an expensive suburb.

The thing with R that I think is important to note is that you don't have interactivity to support code (eg in ruby and python the huge advantage of the repl is to interact with the code you are writing while you're writing it) but rather your code is a way to the make interactive experience better. R is an amazingly advanced calculator.

The reason I use R over python most of the time is because, despite some amazing improvements in this area by the python community, there's no better tool for fluidly interacting with data that offers the same power.

That said, R is not for writing software systems. People used to refer to many interpreted languages as "scripting languages", and while this is clearly not the case for Python and Ruby, this is exactly what R is. There's a good reason in RStudio it says "New File > R Script". The limit of using R is when you have a bunch of scripts that interact with each other to create a bunch of visualizations/reports. If your system gets more complicated then that, write it in something else.

Of the many language/environment combos I've used, I don't think I've found one better for rapid prototyping than R, and following from that R has no place near anything that would be called "production". I also happen to think, if used properly, this is a good thing since it means your "prototype" never accidentally creeps into suddenly being your production system.

Also no graphviz![0] I know gephi is gaining a lot of ground, but for a wide range of visualizations requiring actual graphs it is still very hard to beat the simplicity and power of graphviz.

[0] http://www.graphviz.org/

"The techniques of the police, which are developing at an extremely rapid tempo, have as their necessary end the transformation of the entire nation into a concentration camp. This is no perverse decision on the part of some party or government. To be sure of apprehending criminals, it is necessary that everyone be supervised. It is necessary to know exactly what every citizen is up to, to know his relations, his amusements, etc. And the state is increasingly in a position to know these things."

- Jacques Ellul "The Technological Society" (1964)

I would suggest that you put together a quick monte-carlo simulation for any of the models you're experimenting with to see how well they perform when you actually know the true conversion rates. There's plenty of theoretical issues you can find with any method and the more complex what you're doing is the harder it can be to work it all out with pencil and paper. Likewise, because you're dealing with probabilistic solutions, real-world results can be deceptive (for example conversion rates may naturally fluctuation between weeks or months). I've found that testing with simulations is the best way to get a real sense of how whatever method you wish to employ will work.

Except in that example the author is choosing E based on the observed difference between two means (which is the exactly the unknown you're trying to determine, so it makes no sense to use it as a constant in a formula), rather than the threshold for the minimum distance you care about.

If you're going the classical statistics route the entire point is that you need to determine your sample size before you peek at the data. In that post you would need to replace E with a threshold of difference that you care about, then calculate n before you start the test and not look at the results until you had reached n observations.

There are several things that help. Firstly you're not just looking for a red light/green light significance. Since you're actually modeling the beta distribution for each conversion rate you not only can ask "what's the probability that this test is an improvement?" you can actually sample from both distributions and see what that improvement looks like.

For example I just simulated some bad data. A has 480 observations and a mean conversion of 33%, B has 410 observations and has a mean conversion of 37%. The p-value here is 0.0323 In the traditional A/B testing model we'd be done and claiming better than a 10% improvement!

However when I sample from these 2 beta distributions I see that my credible region is -2% to 34% meaning this new test could be anywhere from 2% worse to 34% better. No magic value is needed to tell you that you really don't know anything yet.

Another huge help is the use of a prior. Until your data overrides your prior belief you aren't going to see anything. Going with the last example, if I had a good prior that the true conversion rate on that page was actually 33% I wouldn't have even gotten a p-value of less then 0.05. On the other hand if I had a strong prior that the conversion rate was 50% that would imply that both A and B were getting strangely unlucky results, which would actually boost the probability that B was in fact an improvement.

On the philosophical side, Bayesian statistics are simply trying to quantify what you know, not give you 'yes'/'no' answers. Maybe the gamble of -2 to 34 is good for you, or maybe you really want to know tighter bounds on your improvement and aren't comfortable with any possibility of decline. Bayesian statistics gives you a direct way to trade off certainty with time.

confidence in your own decisions can also be referred to as a Bayesian prior ;)

I've treated the A/B tests I've run pretty much as a case of Bayesian parameter estimation (where the true conversion of A and of B are your parameter). You then get nice beta distributions you can sample from, as well as use the prior to constrain expectations of improvement and also reduce the effects of early flukes in your sampling.

I hear more and more chanting of "correlation does not equal causation!" which is great if your goal is to form a causal model of the world, but there are plenty of insights you can arrive at from correlation alone.

For starters in the world of machine learning and predictive analytics, it doesn't really matter if X causes Y so long as X is a consistently good predictor of Y. Maybe powerlines being over someone's home are not the cause of cancer, but if their presence can be used to predict cancer rates that's a good thing.

More important imho is the idea of latent or hidden variables. Two things that are clearly correlated but also seem to not have a causal relationship (just as transistors and longevity) may share a latent variable, that may be either non-quantifiable or completely unobservable. For either case measuring outputs that share a common latent variable and thus correlate with each other might be the only way to attempt to measure hidden, non-quantifiable causes.

For example employee happiness might be the cause of employee retention. However you can't currently measure or observe 'happiness', but there may be many, seemingly, unrelated employee activities that correlate with retention because they are also driven by this same latent variable. Studying them is the only way to get a quantifiable understanding of this latent cause.

tl;dr somethimes correlation is just as important as causation.

Feature extraction can be aided by unsupervised data but will certainly work with labeled data. One of the advancements bundled under 'deep learning' is how we can leverage unlabeled data (which is much easier to come by) to improve performance. And of course you can always do unsupervised learning with labeled data, just toss out the labels ;)

It's actually the multiple layers hidden units that perform non-linear feature extraction and the unsupervised pre-training is simply a means to do this better (theoretically, although we don't really know what's happening as much as it would seem).

Most of the current research shows deep neural nets to be state of the art in image classification and nlp. I don't know that it is the case that deep learning techniques do not work out side this area, it's just there hasn't been much published on it either way. Although I do believe the Kaggle Merck contest was neither of these, and deep learning out performed all other techniques http://blog.kaggle.com/2012/11/01/deep-learning-how-i-did-it...

Not articles but there are two very excellent books on the subject that I can't recommend enough:

If you read calculus with about the same fluency as the comic books then "Data Analysis: A Bayesian Tutorial" is awesome http://www.amazon.com/Data-Analysis-A-Bayesian-Tutorial/dp/0...

And if you would like a little more exposition (but still a mathematically sophisticated treatment) "Doing Bayesian Data Analysis: A Tutorial with R and BUGS" is fantastic http://www.amazon.com/Doing-Bayesian-Data-Analysis-Tutorial/...

The latter will also give you more details of how to approach classical, frequentest tests and summary statistics with their Bayesian equivalent.

Honestly I would say get both books as they're cheap and provide different insights. You only need to read a few chapters of each to see how you approach basic experiments from a Bayesian perspective.