zhanwei

statements like that typically come from best-selling author and keynote speaker. This kind of articles seem valuable to some audience, but it's probably something I can never understand.

"What if a large class of algorithms used today -- from the algorithms that help us avoid traffic to the algorithms that identify new drug molecules -- worked exponentially faster?" -- these problems used as example in the opening line are only tangentially related to the problem of submodular maximisation that the paper it is talking about is tackling. I also don't think submodular maximisation is widely used today for identifying new drug molecules or avoiding traffic.

He already did his job of intellectually engaging the recruiter. If the other party is not interested, it will not be rude if he update his resume and move on.

The guy writes well. The recruiter is the one with bad attitude. But I would say that a practical communicator knows his audience, and in this case might want to give up explaining POSIX, change his resume, and move on.

I agree with this article more. Yes, $86M is sure expensive with lots of efficiencies. But the first one is oversimplifying the technical aspect, which is missing the point on the sources of inefficiencies. The project can't be using openALPR at its current level and it is hopeless to use it as a starting point to get decent accuracy.

So IBM Watson can do all the smart and complex stuff but we still need human to do the dumb stuff like importing excel files where the cost outweighs the benefit of getting Watson to do the smart stuff.

"I still think you're thinking of this as you'll take an existing product and add some AI to it. That’s not what we’re seeing. What we’re seeing is an entirely new kind of product that wasn't possible before."

The major difference from previous platform shifts is that the limits of frontier AI technology are extremely hard to gauge even for experts. Unlike AI tech, previous platform shifts are easy to understand for a person on the street. For e.g., consumers can see that smartphone allows messages to sent over cellular data, which is better than SMS. But it is not obvious whether Echo is better than Siri? Tesla's driver-assist more reliable than a competitor one?

I feel the recent breakthroughs in AI are concentrated in a small number (but very important) areas such as computer vision and to some extent machine translation. There need to be more advances especially in areas of decision making before AI can have a broader and more meaningfully impact as stated in the interview.

"Now that DeepMind knows the approach works, it also knows where its AI system lacks information, so it may ask Google to put additional sensors into its data centers to let its software eke out even more efficiency."

Sounds like active learning to me. It's a type of machine learning where a learner pro-actively ask for interesting data points to be labeled so that he can learn more about the system. :)

This tutorial seems quite good. Covers the basic and various useful UCT extension: https://webdocs.cs.ualberta.ca/~mmueller/courses/2014-AAAI-g...

However, the tutorial doesn't work towards a working implementation. I think you can verify your results against benchmark problems. There are a number of good implementations around:

UCT implementation with many MDP benchmark problems: https://github.com/bonetblai/mdp-engine

My favorite implementation. The code is quite easy to read: http://www0.cs.ucl.ac.uk/staff/D.Silver/web/Applications_fil...

UCT is a very simple idea that works surprisingly well across very diverse domains. Can't emphasize its generality enough, you can throw different problems at it and it can give you decent (may not be the best) result without using any additional domain knowledge.

It is also very easy to work with, you can easily tweak the algorithm and add heuristics for your specific domains.

Also relevant:

UCT applied to partially observable game (Poc-man) http://papers.nips.cc/paper/4031-monte-carlo-planning-in-lar...

Another approach for Monte-Carlo planning http://papers.nips.cc/paper/5189-despot-online-pomdp-plannin...

I think it could be predictability of a self-driving car in an accident that is making people uncomfortable and accident rate does not capture that. For e.g., an accident might means hitting a barrier, a human realizes the terrible situation and will swerve away from the cliff behind the barrier. But what would a self-driving car do? Its sensor might be too damaged to realize the dire consequence.

I see a number of meaningful advances in computing technologies last year. But the term "AI breakthroughs" is getting meaningless these days.

Yes, data is important to AI but open-sourcing TensorFlow doesn't mean that code is not. Rather, it means that data and code have different strategic value. Data is their secret sauce and code is their network. The more people use and contribute to the library, the better the code gets.

Also important to note TensorFlow is probably not the complete package of what they use at Google.

I see it as it always "human-in-the-loop" since the beginning of computing. Human using spreadsheet to do accounting is a human-in-the-loop operation with human deciding what to add, and computer doing the adding. The key difference with AI is that it's now more difficult for the human to understand what the "smart" computer is doing and how to cooperate with it. For e.g., how do the self-driving take over control from and hand over control to human/ How to visualise and give directions to complex algorithms / how gain business insight from data.

next step, check your calendar for scheduling conflict and suggest the right response, i.e., "i'll be there"/"i can't make it".

further extension, pull out quick answers from other emails, i.e., "what's the sales number this month?" "100M" from another email

It seems to me to be all about having the right prior and planning for exploration. Policy search methods (http://arxiv.org/abs/1504.00702) assume that there aren't many trajectories that make sense (based on prior knowledge/testing in simulator) and search for the best ones among those that make sense using real-world data. Even within policy search you need some kind of exploration such injecting gaussian noise in trajectories. The hard part is to come up with a model for exploration.

Vim works well enough for me. I got merlin-mode, ocp-indent and syntax checker to work in vim but not REPL.

1. people like to point out when others are wrong on the internet. It much harder to do so for in-depth writings. Relevant: https://xkcd.com/386/

2. my newsfeed is never ending. Am I going to spend 5-10 min reading an in-depth article? and I know nobody else on Facebook will like/comment/discuss the long article with me.

There are still some elements of uncertainty in car driving. For e.g. driverless cars are confused by cyclist doing track stand[1]. Hopefully with enough data driverless car could overcome these difficulties.

[1] http://road.cc/content/news/162468-cyclist-doing-trackstand-...

He is comparing situations where human decision makers are highly trained professionals such as astronauts and passenger plane pilot to daily driving situations where the average human decision maker is more likely to be an idiot(people switch off while driving) distracted by his smartphone.

Could be practical if there is a protocol for dash cams to communicate with each other. i.e. car behind can auto grab the video feed from the car in front of it if they are near and display it on its own screen.

The "improvements" and "big ideas" the author listed are real and exciting developments. I think about robotics a lot (in academia) but given these developments, a Cambrian explosion (i interpret as robot succeeding in general simple tasks that human can do) is still way beyond my imagination. There are many technical hurdles such as perception, planning, learning and reasoning. The key challenge now seems to be finding scenarios where robots can be successful with current technology.

On a separate note, the robobrain project http://robobrain.me/ seems to fit the running themes in the paper.

You will somehow need to give your personal information for a social network to work and to receive these services in return. For e.g., u need to identify yourself and your friends to make sure only your friends can see your photos and vice versa. Question is how much info is too much.

The founders of Batteriser would know the truth that device manufacturers are not "stupid". The message is targeting the consumers. It's a nice marketing trick. It definitely makes the consumers feel smart because they beat the "stupid" incumbents by buying a Batteriser.

I like how "official account" are lightweight apps. They are like the web with push notifications, online identity, and payment integrated into the web browser. The usual web can fulfil the features of official accounts in theory, but it's awkward. 1. push notification -> email 2. online identity -> login to fb, twitter, github, or website specific login 3. payment -> authorize paypal, bitcoin, amazon, bank

Whenever u want to do something on the web (e.g. buy a ticket), u have to figure out how to login (worse if you have to create a new account), how to make payment, and check your emails for confirmations and updates. That's enough friction to make a user hunt for alternatives. On top of that, the website make not work well on your mobile phone.

The current safety technologies like lights and brakes are absolutely necessary but they only work up to a certain point. The rest are human factor. But to say that cyclist should stay safe by spraying paint on their body is getting a bit too far.

Everyone should be responsible for their personal safety, regardless what everyone else says. But beyond that, i guess this piece is pushing for more responsibility on the motorist end.

Businesses are now free to refuse services to Apple because of their CEO Tim Cook. good luck with that.

We need a model to use Monte Carlo Tree Search... which is missing in this approach since it uses model-free reinforcement learning. Unless the deep convolution net can extract some latent features as state, it would be impossible to do planning on top of it.