In my opinion, we are WAY too early discussing the direction development of hardware for building neural nets will go. Probably even too early to discuss what the (usual) "geometry" of general purpose neural nets will be once that is taken into account (how can be broken into "building blocks" trading off size and connectivity between these units).

I rather suspect that most of the major players in the "chip" business have teams working on that right now.

I also suspect that at least SOME of the approaches now used will not be what ultimately wins out

There's a real chance I'm wrong. To start with, I'd say it's either Microsoft or Google. Only Microsoft, Google and IBM likely have the clout and interests to request third party firmware and modify it on a regular basis. He was only an intern at IBM, so it's less likely he got so many details. I thought I'd read about Google requesting firmware and making changes elsewhere, but I can't find it, so I can't put too much faith in that. The other big thing is that for a long time, Google's greatest differentiator was arguably its sophisticated infrastructure--no one else was even close. Lots of Google papers describe their actual production infrastructure and techniques, albeit with enough detail left out that third party reimplementations are quite different (https://changelog.com/podcast/250 mentioned that this was why they created Kubernetes--they wanted the world to standardize on their system, not a reimplementation based on what they'd publicly said). Microsoft runs more of a classic research lab where they pay brilliant people like Simon Peyton Jones to write papers about awesome things and then never include them in any project.

I just went to google and confirmed that they do fire workers over leaks (https://www.theverge.com/2017/5/22/1566 ... stop-leaks -- not commenting on the more sensational accusations, just the reality that they try to prevent leaks). I'd honestly be surprised if any big company (> 5k employees) hadn't fired people for leaks, and there are definitely cases where it's justified.

_________________
Occupy Babel!

Okay I reviewed the closing ceremony video on YouTube and re-read the Deepmind's webpage on "AlphaGo's Next Moves" and am of the conclusion that I'm now more and more convinced that this tool, is probably not intended by Deepmind to be a publicly accessible tool for public consumption and use at all.

After re-watching the video here https://www.youtube.com/watch?v=ru0E7N0-kFE and also reading the entire page here: https://www.youtube.com/watch?v=ru0E7N0-kFE

I don't think the upcoming "AlphaGo Tool" that Google mentioned will actually be accessibly to the public for public use at all.

In reference to the 50 games and the upcoming Ke Jie video analysis and the second research paper etc Google uses the word "release" and "publish" but this was never once mentioned when regarding the "tool".

Deepmind at the end said it had a series of "gifts" for the Go community, and that it was the research paper and the 50 games. But never was the tool directly mentioned.

""We’re also working on a teaching tool - one of the top requests we’ve received throughout this week. The tool will show AlphaGo’s analysis of Go positions, providing an insight into how the program thinks, and hopefully giving all players and fans the opportunity to see the game through the lens of AlphaGo. We’re particularly honoured that our first collaborator in this effort will be the great Ke Jie, who has agreed to work with us on a study of his match with AlphaGo. We’re excited to hear his insights into these amazing games, and to have the chance to share some of AlphaGo’s own analysis too.""

It would seem a lot of folks (myself included at first) have "read between the lines" and imagined that this will be a "tool" accessible for "use" by everyone. However it doesn't really mentioned that at all, only that Deepmind is working on a teaching tool, but for all anyone knows this tool could be entirely internal and not meant for public release or access. In particular, it mentions "honored that our first collaborator in this effort will be the great Ke Jie", well this does make it sound like this is going to be an internal tool and not something for wide release to the public. I mean, if this was going to be either a downloadable analysis tool in the form of a client and/or some web based in-browser version that anyone could readily and easily access, why would their "first collaborator" in this "effort" (I'm assuming the effort referring to the creation and use of this teaching tool) need to be with the great Ke Jie? If it was going to be widely released for everyone's use surely this wouldn't be phrased this way? And Deepmind referred to the 50 self playing games as a "gift" to the Go community but never made such a reference with regards to the "tool" itself, which leads me to believe the "tool" is not going to be published or released to the public. Ever heard of the phrase give a man a fish and feed him for a day, but teach him to fish and feed him for a lifetime? If Google considered the 50 games as a "gift" to the Go community, then surely if it were to release such a tool that would be far greater a gift than 50 games, yet the "tool" was never put in the same category.

It also, when doing an linguistic analysis, indicated that Google wrote : "hopefully giving all players and fans the opportunity to [i]see the game through the lens of AlphaGo."[/i] Note what Google didn't write is " to use the tool" or " to play the game" or "to analysis the game". it merely said it hopes to give all players the opportunity to "see" the game....

Strongly suggesting the tool is not going to be released, or at the very least, there is no strong evidence indicating Google making any form of pledge or statement that there will be such a tool available for the public at large.

Just wanted to add: relax, people. AlphaGo has paved the way, but it won't be the end of the road. Crazystone and Leela have already implemented NN techniques, and those are done by single programmers. And they are already very strong programs. I bet there will be many groups that will be inspired/challenged to make something even better. Maybe a Chinese/Korean/Japanese university collaboration, or some commercial undertaking, who knows. But just like with DeepBlue, there will be successors and they will be better than AGo. It's just a matter of years. Other universities know how to handle intricate NNs as well.

What I'm hoping for are programs that help get the concepts of Go across to students, just like a teacher would do. Deepmind talked about something like that, but if it doesn't come from them, someone will do it. Just be patient.