GOOGL - Google

[formthirteen]

One big risk for $GOOG that has been going through my head a lot lately is the risk of mass adblocker adoption, particularly if bundled in iOS and Safari.  While Chrome, Android, and IE/Spartan are unlikely to bundle adblockers, I don't see why Apple would not.

 

If you are invested in Google, I highly recommend you download ublock, adblock, and an ios adblocker to see the kind of impact it can have on stripping the web of ads. Just something to think about.

 

Regardless I love the company and think mass adblocker adoption is a risk, but not a huge risk.

 

I use ad blockers on both iOS and OS X:

https://adblockplus.org/blog/adblock-plus-for-ios-9-finally-here-and-pssst-it-s-free

 

I expect it to not be a big problem for Google because "everyone has a price":

http://uk.businessinsider.com/google-microsoft-amazon-taboola-pay-adblock-plus-to-stop-blocking-their-ads-2015-2

 

I think the biggest effect ad blockers will have is that they will force advertisers to create ads people want to see. In the end it's a technology arms race and the biggest budget will win.

[blainehodder]

 

 

I expect it to not be a big problem for Google because "everyone has a price":

http://uk.businessinsider.com/google-microsoft-amazon-taboola-pay-adblock-plus-to-stop-blocking-their-ads-2015-2

 

 

I agree with your statement except for Apple. I'm not sure Apple has a price, and can see them bunding ad blockers stock into ios at some point in the future.  Anyhow, Im not saying it is a likely risk, but it is a risk.

[Jurgis]

Google ratcheted Youtube ads a lot in the last year. My guess from user's POW it probably went 200-1000% up compared to beginning of 2015. On the positive side for shareholders, these are harder to block and there's very little substitute for Youtube. It sucks for users of course (anyone has a good solution for blocking?). On the negative side for shareholders, I wonder how much of that is one time uptick and how saturated this is now. In other words, this might not decline, but the huge uptick is behind.

[rkbabang]

Google ratcheted Youtube ads a lot in the last year....anyone has a good solution for blocking?

 

There is one easy way to get rid of all the ads, but of course this is what the company wants you to do and probably why the ads have increased past the annoyance threshold:

http://www.reuters.com/article/us-alphabet-youtube-subcription-idUSKCN0SF2KO20151021

 

[Aurelius]

I use Adblock (getadblock.com). I get zero ads on youtube and facebook, and everywhere else.

I recently introduced it to a friend, who was amazed - loved it and wanted to spread the word!

So far it has blocked 87K advertisements. That has to make a difference somewhere.

 

[Aurelius]

Given the environment we live in today with very low interest rates, which is certainly very different from the one at the time Buffett said he was looking for a 10% return in year 1, a 5.6% return and growing might not be so bad after all… Surely, it doesn’t seem GOOG is as overvalued as the other FANG stocks.

 

AMZN retail did close to 200 Billion GMV in 2015, growing ”paid units” at 26%. Normalizing EBIT margin to 5% (which is less than WMT) to adjust for investments in growth yields 7B after tax. PE=35. But this is before subscribing any value to AWS.

 

AWS has a revenue runrate of 9,4B with a growing (16,9% to 28,5%) operating margin. It's the fastest growing enterprise tech business ever (http://uk.businessinsider.com/aws-estimated-to-be-worth-160-billion-2015-11?r=US&IR=T) AWS could be worth the entire mkt cap in a few years.

 

Both businesses have great secular winds in their back.

A great founder CEO.

On top of that you get optionality on future endeavours.

[giofranchi]

Both businesses have great secular winds in their back.

A great founder CEO.

On top of that you get optionality on future endeavours.

 

+1

 

Cheers,

 

Gio

[giofranchi]

Google Acquihires Work Chat Startup Pie To Build Engineering Team In Southeast Asia

 

http://techcrunch.com/2016/02/18/google-eats-pie/

 

Cheers,

 

Gio

[giofranchi]

Why is Google's Go win such a big deal?

 

http://www.theverge.com/2016/3/9/11185030/google-deepmind-alphago-go-artificial-intelligence-impact

 

Cheers,

 

Gio

[giofranchi]

Google's DeepMind beats Lee Se-dol again to go 2-0 up in historic Go series

 

http://www.theverge.com/2016/3/10/11191184/lee-sedol-alphago-go-deepmind-google-match-2-result

 

Cheers,

 

Gio

[giofranchi]

DeepMind founder Demis Hassabis on how AI will shape the future

 

http://www.theverge.com/2016/3/10/11192774/demis-hassabis-interview-alphago-google-deepmind-ai

 

Cheers,

 

Gio

[rishig]

DeepMind founder Demis Hassabis on how AI will shape the future

 

http://www.theverge.com/2016/3/10/11192774/demis-hassabis-interview-alphago-google-deepmind-ai

 

Cheers,

 

Gio

 

Go is one of the most complex games humans have ever deviced, exponentially harder than chess. There are 10^700 Go games you can possibly play (that's more than all the atoms in the universe). An average move in chess has ~20 possible moves while an average move in Go has ~200 possible moves. The "branching" factor is an order of magnitude higher. Also, figuring out whether a specific move in Go leads to a victory is not something one can determine by using brute force, it requires human intuition. If you were to read any technical papers, this is referred to as difficulty of defining an evaluate function for Go. This is unlike chess where individual pieces have differing values that can be used to define the evaluate function (i.e. determine whether a move is superior to the other possible moves).

 

It has been a long standing challenge of artificial intelligence to beat a pro AI player. To be able to beat the world Go champion, Lee Sedol, 3 to 0 is an absolutely amazing feat. Lee Sedol is considered as the greatest Go player of the last decade, and has won 18 international titles.

 

The big difference between Deep Blue (the machine that beat Kasparov in chess 20 years ago) is that it had a set of hand crafted rules and heuristics hard coded vs. AlphaGo has general purpose algorithms it used to learn how to play Go by watching human games, and training itself to get better than any individual! AlphaGo is more human like in how it thinks about the game.

 

Here's technical details on AlphaGo published in Nature: "Mastering the Game of Go with Deep Neural Networks and Tree Search":

https://vk.com/doc-44016343_437229031?dl=56ce06e325d42fbc72

[Liberty]

Good explanation, Rishi. This is indeed an extremely impressive achievment and a step change in AI's ability to play the game, which has taken most informed observers by surprised.

[Picasso]

So can Google eventually setup a site where you get to play against AlphaGo?  Basically anyone can try and beat the AI but it will keep learning from all the various players and get stronger?  Or would that require way too much computing power?

[merkhet]

Looks like we won one

 

http://www.wired.com/2016/03/go-grandmaster-lee-sedol-grabs-consolation-win-googles-ai/

[rishig]

So can Google eventually setup a site where you get to play against AlphaGo?  Basically anyone can try and beat the AI but it will keep learning from all the various players and get stronger?  Or would that require way too much computing power?

 

It's important to note that the version of AlphaGo that beat Lee Sedol was trained using amateur Go games for its initial training data.  Later AlphaGo played millions of games with itself to generate additional data, but it had no other real world games fed to it. And even if it were fed real world games, the data sample is too small compared to the millions of games it generated by playing with itself to make any difference.This is the most important difference between Deep Blue and AlphaGo.

 

DeepMind's CEO, Demis Hassabis, explains this important difference (called Artificial General Intelligence) here:

 

[giofranchi]

After AlphaGo, what's next for AI?

 

http://www.theverge.com/2016/3/14/11219258/google-deepmind-alphago-go-challenge-ai-future

 

Cheers,

 

Gio

[rmitz]

It's important to note that the version of AlphaGo that beat Lee Sedol was trained using amateur Go games for its initial training data.  Later AlphaGo played millions of games with itself to generate additional data, but it had no other real world games fed to it. And even if it were fed real world games, the data sample is too small compared to the millions of games it generated by playing with itself to make any difference.This is the most important difference between Deep Blue and AlphaGo.

 

DeepMind's CEO, Demis Hassabis, explains this important difference (called Artificial General Intelligence) here:

 

I don’t see anything that says they restricted initial human learning to amateur games.  I would be very surprised. Where’d you see that?

[rishig]

It's important to note that the version of AlphaGo that beat Lee Sedol was trained using amateur Go games for its initial training data.  Later AlphaGo played millions of games with itself to generate additional data, but it had no other real world games fed to it. And even if it were fed real world games, the data sample is too small compared to the millions of games it generated by playing with itself to make any difference.This is the most important difference between Deep Blue and AlphaGo.

 

DeepMind's CEO, Demis Hassabis, explains this important difference (called Artificial General Intelligence) here:

 

I don’t see anything that says they restricted initial human learning to amateur games.  I would be very surprised. Where’d you see that?

 

See Q/A from match 4 at 6:09:39:

[rmitz]

It's important to note that the version of AlphaGo that beat Lee Sedol was trained using amateur Go games for its initial training data.  Later AlphaGo played millions of games with itself to generate additional data, but it had no other real world games fed to it. And even if it were fed real world games, the data sample is too small compared to the millions of games it generated by playing with itself to make any difference.This is the most important difference between Deep Blue and AlphaGo.

 

DeepMind's CEO, Demis Hassabis, explains this important difference (called Artificial General Intelligence) here:

 

I don’t see anything that says they restricted initial human learning to amateur games.  I would be very surprised. Where’d you see that?

 

See Q/A from match 4 at 6:09:39:

 

Human experts, yes, not amateurs, but pros as well.

[rishig]

It's important to note that the version of AlphaGo that beat Lee Sedol was trained using amateur Go games for its initial training data.  Later AlphaGo played millions of games with itself to generate additional data, but it had no other real world games fed to it. And even if it were fed real world games, the data sample is too small compared to the millions of games it generated by playing with itself to make any difference.This is the most important difference between Deep Blue and AlphaGo.

 

DeepMind's CEO, Demis Hassabis, explains this important difference (called Artificial General Intelligence) here:

 

I don’t see anything that says they restricted initial human learning to amateur games.  I would be very surprised. Where’d you see that?

 

See Q/A from match 4 at 6:09:39:

 

Human experts, yes, not amateurs, but pros as well.

 

I think the main point is that it's not the initial training data that makes the difference. It's AlphaGo playing with tweaked versions of itself millions of times over that creates the the "deep neural networks".  As Demis commented in the question, you could feed it 1000 games played by Lee Sedol and it would not change much. More on this (reinforcement learning) here:

http://googleresearch.blogspot.com/2016/01/alphago-mastering-ancient-game-of-go.html

[rmitz]

It's important to note that the version of AlphaGo that beat Lee Sedol was trained using amateur Go games for its initial training data.  Later AlphaGo played millions of games with itself to generate additional data, but it had no other real world games fed to it. And even if it were fed real world games, the data sample is too small compared to the millions of games it generated by playing with itself to make any difference.This is the most important difference between Deep Blue and AlphaGo.

 

DeepMind's CEO, Demis Hassabis, explains this important difference (called Artificial General Intelligence) here:

 

I don’t see anything that says they restricted initial human learning to amateur games.  I would be very surprised. Where’d you see that?

 

See Q/A from match 4 at 6:09:39:

 

Human experts, yes, not amateurs, but pros as well.

 

I think the main point is that it's not the initial training data that makes the difference. It's AlphaGo playing with tweaked versions of itself millions of times over that creates the the "deep neural networks".  As Demis commented in the question, you could feed it 1000 games played by Lee Sedol and it would not change much. More on this (reinforcement learning) here:

http://googleresearch.blogspot.com/2016/01/alphago-mastering-ancient-game-of-go.html

 

I'm sorry, when the video came up it actually came up to a different point (because I was already watching it) and it coincidentally said something about using expert games for training.  At the point you pointed out it was very clear that they said amateur games.  I actually find this surprising because there can be moves in those games that could "pollute" the model, and wonder why they didn't use pro games, except that they 1) probably had more amateur games or 2) that they wanted to prove a point.

 

The comment about 1000 games about lee sedol is a little misleading in that I think it *would* change depending on when those games were introduced, given the difference between the policy and training networks.  Giving those games now would have a minimal difference to the trained network, but could give better moves as suggestions for alphago to look at and consider, potentially.  It would probably not serve as any sort of guide to playing Lee Sedol though.

 

A key point is that you still have to dramatically limit the amount of possible moves that alphago looks at in order to accomplish anything, and while we have some bits and pieces of that, it's really not clear. The policy network clearly plays a role.

 

I understand neural networks pretty well so that part is not in question.  Alphago clearly has a lack of sophisticated time management (the team members said as much) which it could otherwise have taken significant advantage of when it was "surprised" to look deeper.

 

Aside: I love these games.  What they've accomplished here is fantastic.  I am hopeful to see some more papers and even making a version of Alphago for individuals. They already said they have one for running in a non-distributed environment, which isn't as strong, but would be plenty for lots of people.

[giofranchi]

Google's AlphaGo AI beats Lee Se-dol again to win Go series 4-1

 

http://www.theverge.com/2016/3/15/11213518/alphago-deepmind-go-match-5-result

 

Cheers,

 

Gio

[NewbieD]

Apologies if this has been linked Before: http://uk.businessinsider.com/urs-holze-talks-google-cloud-beat-search-2015-11?r=US&IR=T

 

"Urs Hölzle, Google's eighth employee and overall cloud boss, thinks that within the next five years, the company's Google Cloud Platform revenues could surpass Google's advertising revenue."

 

Any realism to this? I guess margins are much lower for Cloud than Ads. But if his belief becomes true it would still be a major source of value.

[giofranchi]

Why Google Wants to Sell Its Robots: Reality Is Hard

 

http://www.bloomberg.com/news/articles/2016-03-18/why-google-wants-to-sell-its-robots-reality-is-hard

 

Cheers,

 

Gio