alexbeloi

Siri/Alexa/GoogleAssist etc. work well in constrained situations (e.g. changing music, turning stuff on/off, pop culture quiz questions), they have trouble parsing more general/complex questions.

Many researchers I've talked to believe the initial self-driving car deployments to the public will be have to be very constrained: fixed routes, possibly isolated self-driving car lanes, only allowed in clear/daytime conditions.

Looks like a good intro, has reference links to keywords/definitions, problems with solutions, text-searchable.

References TI-83 or TI-84 feel dated.

Do you actually think reviews from friends are objective and not grossly positively skewed?

The difference between investing now and staying in now is a single transaction fee.

I don't have any data on this, but to me this seems like a case of confirmation bias + survivor bias.

I don't see skills related to getting good grades vs skills related to innovation as being at odds with each other.

Maybe the play is to get people using it quickly and get bought by Adobe as a way to resolve the legal problem. The owner is putting themselves at a huge disadvantage in negotiation doing this.

>He was fired because he gave up control of his company and for no other reason.

Gave up control or lost control?

>Founders don't get fired because employees find them unbearable or sabotaging. Those employees get fired or quit...unless the founder gives up control and then loses an internal political fight.

see Kalanick, founders don't get fired because of employees, they get fired because of they can't keep their board loyal to them. This was the case with Steve Jobs as well.

Jobs' return to Apple in 1997, he made sure that 4 new board members were appointed[0], which you can be sure were loyal to him first and foremost.

He lost control of his company the first time around, learned from his mistakes and corrected. There is no need to glorify his early mistakes, they don't diminish his accomplishments.

[0] http://money.cnn.com/1997/08/06/technology/apple/

It looks like their metric is: "good" interviewer means the interviewee enjoyed the interview and to want to move forward if asked.

Likely a very useful metric for figuring out if some interviewers are turning people off about the company.

I (and probably everyone else) would really like to see if this at all correlates with future performance metrics.

You are right, I am using money laundering to mean 'anonymous monetary transfer of arbitrary size/value', which most governments don't allow as far as I know.

Their billions are liquid and can be used to buy protection (secluded homes, private transportation), Satoshi's aren't (in so far as he hasn't liquidated them).

>I think one big reason is that regulators don't want the US to miss out on a large wave of value creation.

There is no value created, it is just redistributed through rent-seeking. The US might hope that the redistribution goes in its favor if, as you say, more early adopters are in US rather than abroad.

The global effect of rent-seeking is that money gets pulled out of productive enterprises and into non-producing assets, lowering GDP.

Whether or not the easy access to money laundering channels makes up for the lost GDP or makes it worse, I am curious to see.

You're describing money laundering and selectively choosing to represent the positive examples of it.

The source of the currency is not the issue, it's the longevity/stability of it as a medium of exchange. This startup has very little power (compared to the government) to ensure that the currency will continue to hold value to someone you want to trade with. If faith plummets, they can only guarantee as many assets as they hold themselves.

It's new and interesting, but hasn't outperformed any existing methods yet. High representation efficiency though.

They would peg it to something that everyone did believe was worth something.

>Economists fell for the myth of Black–Scholes. This is a neat way of quantifying risk, but it assumes that risk has a Gaussian distribution. Sometimes it doesn't.

Maybe you're being loose with your terminology, but there is no such assumption. The simplest versions of Black-Scholes does assume that stock price is a lognormal process with constant variance/volatility, which empirically it isn't.

The point of Black-Scholes is to price derivatives not quantify risk, but the price also has an empirical market price. Just as you can use the 'assumed' variance/volatility to try to compute the derivative price (according to B-S), you can use the market price of the derivative to compute the 'implied' variance/volatility.

If you look at the options page for AAPL on yahoo finance[0], you'll see a column for implied volatility, the above method is precisely how it is calculated.

[0]https://finance.yahoo.com/quote/AAPL/options?p=AAPL

My summary:

* deephi built an FPGA based processing unit specialized for sparse low precision tensor operations

* Now they have a dev kit for turning your tensorflow/caffe/mxnet models into compressed (sparse/low precision) binaries that are optimized to run on their processors.

They published a paper about how they balance compression/accuracy. [0]

[0] https://arxiv.org/pdf/1612.00694.pdf

Surely this is a joke or shill account. Not only are you arguing against an entirely different topic than at hand. You're openly suggesting that the best way out of getting screwed is to allow more companies the opportunity to screw you?

History has proved the Luddites wrong time and time again, perhaps this time will be different, but I doubt it.

Two public reasons that are commonly given

* controlling data about Chinese citizens from going outside of Chinese control (e.g. onto international/US servers where it can be subpeona'd by the foreign government host)

* protecting local industries from foreign competition

This site has some opinions on the matter:

https://consumerist.com/2014/05/10/why-starting-a-competitor...

https://consumerist.com/2014/03/07/heres-what-lack-of-broadb...

I use WeChat and Snapchat, they are not at all interchangeable.

>They used the excuse that "people have to much stuff on their feed" but we all know that is BS.

It's really not BS. People's feeds are flooded with content from tons of sources.

Why would posts from company that I've 'liked' get guaranteed display when family/friends/friends-of-friends don't get that treatment. Moreover what if I've 'liked' 100+ bands/restaurants/celebrities/businesses/news-sources, they can easily produce enough posts each day that I would never scroll far enough see all of it, nor would I want to.

The thing you want already exists and is called a newsletter.

>I mean one could imagine a deterministic model that couldn't be "tricked".

The models in reference are deterministic, most models for classification/regression are deterministic.

Adversarial attacks are showing that the models are chaotic (in the dynamical systems sense), they're very sensitive to their inputs.

edit: Not all models have this issue, it's been shown that all the typical image based convolutional neural networks are susceptible to this issue. My guess is that it's a more general problem of high dimensional inputs.

I spoke in hyperbole, there were dozens of papers using generative networks to create real looking simulated data which is the more common approach. You start with noise + some features you want in your image, then you turn that into a realistic image.

The notable exception is if you want to generate segmentation data (where you want labels for each pixel), in this case the set of features you want is essentially an image already.

ICCV 2017

* Photographic Image Synthesis with Cascaded Refinement Networks: https://arxiv.org/pdf/1707.09405.pdf

* Synthetic avatars: https://arxiv.org/pdf/1704.05693.pdf

* Synthesizing cell microscopy data: https://arxiv.org/pdf/1708.04692.pdf

* Material Editing Using a Physically Based Rendering Network: https://arxiv.org/pdf/1708.00106.pdf

* 3D-PRNN: Generating Shape Primitives With Recurrent Neural Networks:

* for more search 'generative' 'generating' 'simulating' 'synthesis' 'synthetic' on: http://iccv2017.thecvf.com/program/main_conference#schedule

Also worth mentioning:

CVPR 2017 best paper, this is the other exception, it's unique because it doesn't require _any_ labeled data

* Learning from Simulated and Unsupervised Images through Adversarial Training: https://arxiv.org/abs/1612.07828

ICCV 2017 this year had dozens of papers using GANs and other methods to make simulated images/videos look more realistic. The hunger for cheap usable data is real.

There are different accounting strategies for when you recognize income/expenses.

Say you land a long term contract to provide a service for 5 years and you need to buy extra hardware to fulfill this service, you can choose to expense the hardware immediately in the first year while you let the earnings come in over the next 5 years. In the first year you would be taking a huge loss (on paper), and in the following four years you would see larger than average profits (on paper).

Alternatively, you could acknowledge the full value of the contract immediately. Or another alternative is to expense the hardware over time.

I'm not an accountant, but this analogy is my basic understanding of the flexibility that companies have.

Yes, optimizes window size for moving average

Optimal detail is measure by:

* minimizing variance of 1-step difference (point-to-point fluctuation)

* maintaining kurtosis (outliers not lost)

+optimizations to reduce search space

Very impressive, the original implementation relied a lot on feature engineering.

I'm surprised they're able to prevent a self-play equilibrium with such a simple loss function.

It's sort of like they are using auxiliary outputs but instead of using them to fit features, they are fitting to multiple ways of arriving at 'best play', through predicting value (SL) and predicting probability for best outcome (RL). In principle, they're doing the same thing but in practice it seems like they are making up for each others shortcomings (e.g. self-play equilibrium with RL).

>Most of the theorems still fell like digging a pool with a spoon. Fermat's last theorem was finally proven, with a lot of effort and diversions.

The Fermat's last theorem equation was an incredibly difficult edge case that is illustrative of very fine boundary of a grand theorem. In the case of FLT, that was the modularity theorem.

I wouldn't call anything put towards proving FLT a diversion, the modularity theorem and other theorems that came as a result of proving FLT are the 'basic' (but complex) discoveries that these problems yield.