sushirain

Perhaps landing into a tank of non-corrosive liquid would prevent such collapses? Or on some kind of "glue"/alloy that quickly freezes.

You can also have spearguns firing cords from the top of the rocket to the ground after landing, to hold it.

Larry and Sergey started Google because they wanted to be rich? Even if they had known for sure that it wasn't the way to the 1%, but say to the 10%, would they not have started Google?

The article mentioned brain drain, but more specifically, Canada let Hinton, Lecun, and Sutskever go... :)

These are very interesting points:

"the average elapsed time between key algorithm proposals and corresponding advances was about eighteen years, whereas the average elapsed time between key dataset availabilities and corresponding advances was less than three years, or about six times faster"

"we might already possess the algorithms and hardware that will enable machines in a few years to author human-level long-form creative compositions, complete standardized human examinations, or even pass the Turing Test, if only we trained them with the right writing, examination, and conversational datasets."

What do you mean by hand engineered? Is a part-of-speech tagger trained on a manually annotated corpus hand engineered? If not, then which "stacks" are hand-engineered and still popular today?

You are right. If deep learning takes more and more of the processing, we will need GPUs or more CPU cores, not necessarily clock speed.

Some of my least favorite quotes:

"A single neural network is never going to be able to write news articles with actual sense or meaning." - why not?

"they're never going to be providing you with an honest opinion of a product" - and a human reporter is going to be more honest?

Highlights

•The Blue Brain Project digitally reconstructs and simulates a part of neocortex

•Interdependencies allow dense in silico reconstruction from sparse experimental data

•Simulations reproduce in vitro and in vivo experiments without parameter tuning

•The neocortex reconfigures to support diverse information processing strategies

From the article: "The list of dimensionless physical constants increases in length whenever experiments measure new relationships between physical phenomena. The list of fundamental dimensionless constants, however, decreases when advances in physics show how some previously known constant can be computed in terms of others. A long-sought goal of theoretical physics is to find first principles from which all of the fundamental dimensionless constants can be calculated and compared to the measured values. A successful "Theory of Everything" would allow such a calculation, but so far, this goal has remained elusive."

They are effective because:

- They use more parameters (and fewer computations per parameter.)

- They are hierarchical (convolutions are apparently useful at different levels of abstraction of data).

- They are distributed (word2vec, thought-vectors). Not restricted to a small set of artificial classes such as parts-of-speech or parts of visual objects.

- They are recurrent (RNN).

etc.

So, human drivers will be faster and die more often than AV passengers. I choose slower and safer.

My explanation to the scarcity of "No"s: people are used to seeing in these puzzles mostly sub-types of increasing series: exponential, linear, etc. By the time they had ruled out these sub-types, and had resorted to guessing "ascending", they wouldn't have encountered even a single No.

I predict that if the rule was narrower, like "exponential", much more guesses would have yielded No's.

He does mention a "golden braid forever weaving", which is reminiscent of: "Gödel, Escher, Bach. An Eternal Golden Braid"

CyanogenMod is an Android mod that has customizable Android permissions (Privacy Guard).

It's amazing how much was already known decades ago. Elman and others did much more, and hopefully, now the field will take the next step (which was long delayed), with the help of today's computer power.

Who said there is no feature engineering?

"So there is ultimately no replacement for the smarts you put into feature-engineering."

Recently, deep learning changed this. Finding the right network architecture allows the net to learn the features by itself.

Maybe the brain can do one shot learning only after it has trained for years?

Why does a potential user have to even know about "pods"? Or worse: to "choose a pod". Just choose one for me. Is diaspora for techies, or for people who care about privacy?

I wouldn't worry about it too much. It was a different time back then, when work was mainly guided by linguistic theory. I think that linguistic background is a plus. The quote (maybe not his) means that linguists used to ignore actual language usage.

"Linguistic theory is concerned primarily with an ideal speaker-listener, in a completely homogeneous speech-community, who knows its (the speech community's) language perfectly and is unaffected by such grammatically irrelevant conditions as memory limitations, distractions, shifts of attention and interest, and errors (random or characteristic) in applying his knowledge of this language in actual performance." ~Chomsky,1965

I think that without teachers, the concept is lacking. Even the brightest need teachers. An expert could have helped the students learn more.

Fred Jelinek, who was the pioneer of speech recognition, wrote this. It summarizes the revolution that automatic speech recognition underwent since the 60s: from theory to data-driven science. My new favorite quote on the subject:

"Relationship of linguists to data is the same as that of physicists to their backyard"

The recent advancements using RNN is much more than sign recognition. It's about language generation, control and reinforcement learning, and attention.

If a computer can convince judges that it is human, then it can also convince judges that it is sentient. If it is convincingly sentient, is it moral to program it?

“A woman holding a teddy bear in front of a mirror”. Hilarious.

The jokes will not last long.

Don't forget that long before we have SMI, we will have very powerful AI on our side to improve ourselves biologically and cybernetically. We may be able to grow into a SMI ourselves.

The fear is not absurd, but SMI is not going to materialize soon. We need to get ready, currently, by thinking about it.

Open questions:

* Compare to RNNs with character level input.

* Compare to dedicated methods of sentiment analysis and topic categorization.

The key words are "long term". I am concerned that these obstacles will not be overcome in "3 to 5 years":

* "Chris Urmson of Google has said that the lidar technology cannot spot potholes or humans, such as a police officer, signaling the car to stop." [1]

* "Another big problem for Google is the current cost of its driverless car, which is reportedly outfitted with a whopping $250,000 in equipment." [2]

[1] http://en.wikipedia.org/wiki/Google_driverless_car

[2] http://www.fool.com/investing/general/2015/01/21/3-reasons-g...

Vertical integration is a powerful business strategy.