ARCHITECHT Daily: Google's AI tech lead is for real, but that won't be enough in the cloud

If you're reading this, there's a high probability you heard a lot about all the artificial intellige
ARCHITECHT
ARCHITECHT Daily: Google's AI tech lead is for real, but that won't be enough in the cloud
By ARCHITECHT • Issue #78
If you’re reading this, there’s a high probability you heard a lot about all the artificial intelligence news coming out of the Google I/O conference on Wednesday. But in case you missed something, Recode has a solid rundown of the show’s biggest announcements—most of which involve AI. You can read that here, and I’ll link to Google blog posts with more detail at the end of this.
(One thing Recode didn’t include is Google’s research into using machine learning to design machine learning models, which (1) is not entirely novel and (2) could turn out to be a very important tool in the quest for more, better and easier-to-develop AI.)
However, the biggest news of the day from the cloud computing perspective was definitely Google Cloud TPUs—a hosted, second-generation version of the AI-optimized Tensor Processing Units that Google detailed last month. You could see this one coming from a mile away: Google views AI as its ace in the hole against cloud computing competitors AWS and Microsoft, and there’s no way it would get everyone all excited about TPUs if it wasn’t going to productize them.
I’ve written in more detail about this cloud-AI competition a few times already this year:
Here are some details on Cloud TPUs, per a blog post penned by Google Fellow Jeff Dean:
Each of these new TPU devices delivers up to 180 teraflops of floating-point performance. As powerful as these TPUs are on their own, though, we designed them to work even better together. Each TPU includes a custom high-speed network that allows us to build machine learning supercomputers we call “TPU pods.” A TPU pod contains 64 second-generation TPUs and provides up to 11.5 petaflops to accelerate the training of a single large machine learning model. That’s a lot of computation!
Using these TPU pods, we’ve already seen dramatic improvements in training times. One of our new large-scale translation models used to take a full day to train on 32 of the best commercially-available GPUs—now it trains to the same accuracy in an afternoon using just one eighth of a TPU pod.
Google is also giving some researchers free access to Cloud TPUs via its TensorFlow Research Cloud. (You can get more details on why Google decided to build TPUs in the first place, as well as its take on cloud competition, in the recent ARCHITECHT Show interview with Google VP of infrastructure Eric Brewer.)
Nvidia, for what it’s worth, claims its new V100 GPU processors offer 120 teraflops of performance for deep learning workloads, thanks to the inclusion of 640 tensor cores. 
But don’t be fooled into thinking Google cares one way or another about Nvidia, despite the fact that widespread adoption of Cloud TPUs could have an adverse effect on Nvidia sales. Cloud TPUs are all about sticking it to AWS and Microsoft, Google’s biggest and, in fact, bigger competitors for cloud computing workloads. All three providers will offer Nvidia’s best GPUs for rent in their clouds, just like they all offer managed versions of popular technologies such as Spark or MySQL, but they hope to make their marks with homemade technologies you can’t get anywhere else.
In databases, it’s Cloud Spanner, Cosmos DB and DynamoDB. Now, in AI infrastructure, Google has laid down the gauntlet with Cloud TPUs. Expect AWS and Microsoft to announce some sort of answer by the end of the year, if not sooner.
However, there’s also a software element to this AI battle, where Google also has the early advantage. Cloud TPUs are programmed via Google’s open source TensorFlow deep learning framework, and the company has spent the past several years building up that community. Add in its recent acquisition of machine-learning competition platform Kaggle, and—as Kaggle CEO Anthony Goldbloom explained during a recent ARCHITECHT Show interview—Google has a great opportunity to seed an even larger base of users accustomed to running workloads on its cloud using its frameworks and APIs.
But there’s a big, glowing caveat to this whole discussion: before Google can expect to take over the cloud via AI, it needs to get its act together on trust and customer service. Yesterday, for example, Google also announced that it’s open sourcing the SDKs for the Firebase mobile-app platform it acquired in 2014. The Hacker News thread on that announcement—spurred in part by another Wednesday blog post about Firebase making unannounced changes to its data model and jacking up one startup’s bill by 7,000 percent—provides some insight into how some developers feel about Google. 
They’ve been burned by premature deaths of beloved products, and even a Firebase founder acknowledges that customer support has slipped since the Google acquisition. If developers are leery, imagine how CIOs feel.
On the other hand, there’s Amazon. TechCrunch ran a great guest post on Sunday (save for the strange last paragraph) explaining, in the author’s estimation, how Amazon has come to dominate so many industries over the past decade. It wasn’t always via superior technology, but it also was not by alienating its customers.
Through AWS, Amazon still dominates the cloud computing market it all but created, and Google and Microsoft will both need to bring their A+ games to topple it. I believe that’s possible, and that AI could certainly be a strong weapon in Google’s arsenal, but it needs to build up its reputation in other areas in order for that advantage to pay off. As I wrote earlier this month, Amazon might never top Google as an AI innovator, but it will find a way to make money from it one way or another.
As promised, here are some links to more details on Google’s various AI-infused announcements:

Sponsor: Cloudera
Listen to the latest ARCHITECHT Show podcast
In this episode, I speak with Mike Sherwood, CIO of the City of Las Vegas. Sherwood explains the security threats that keep him up at night, as well as how the city is using AI software to help identify potential threats. He also talks about how Las Vegas plans to become a hub of innovation, ranging from smart stoplights to connected trash cans, and from open data to cloud computing.
Artificial intelligence
Big Blue announced a 17-qubit chip, which is its densest yet, but still nowhere near where it wants to get. It’s competing with Google, D-Wave and others to be first to market with a viable commercial quantum computer—and they’re getting remarkably close.
According to Man Group manager Pierre Lagrange. If the sole purpose of an activity is to maximize returns, and AI can help recognize patterns that help with that, it should be used as widely as possible. Right?
The author of this post posits that vision is the killer app for AI. I disagree, at least in the home. Vision is key for cars, drones, etc, and for some fun photo apps, but people don’t need computers telling them what they can already see in their homes.
And to AMD I say, “Good luck.” The biggest challenge here is actually Nvidia this time around, which has a big headstart on GPU design and works with every deep learning library around.
This is less an AI thing and more an attempt to help people with paralysis, but there’s always an undercurrent of the singularity when embedded chips are involved.
GE Healthcare and Partners HealthCare are embarking on a 10-year partnership to improve hospitals with AI. Between IoT and AI, there are a lot of opportunities to improve health care pretty quickly.
Praise the powers of recommendations! But wouldn’t any recommendation engine be better than not providing recommendations? I’m pretty sure this doesn’t need to invoke AI.
This one is pretty comprehensive, and a quick enough read to get a general sense of what’s up. I suggest it if you find yourself struggling to keep up with the evolving terminology in AI.
medium.com  •  Share
This is a nice intro to AI and the work OpenAI is doing. The discussions around the creation of ImageNet and the state of general AI are particularly good.
I can’t find the deal size, but it’s reported to be in the tens of millions. FPGA maker Xilinx was among the investors, as it looks to find an avenue to blunt Intel’s Altera business and also chip away at GU momentum.
When you train an AI model on high-quality labeled data, and then apply it to data where stuff might be missing (e.g., de-identified patient data), it doesn’t always work. These researchers have figured out a way to improve that.
arxiv.org  •  Share
It made the rounds yesterday. Check it out. 
xkcd.com  •  Share
Cloud and infrastructure
Dropbox’s approach to real-time event-processing is probably a good indicator of what will be commonplace in the next few years. Kafka or similar, and some approach to lambdas/functions, will be everywhere.
With Canadian transportation giant Bombardier. To be clear, though, this isn’t a $700 million Bluemix deal. IBM, like other companies, defines “cloud” broadly, and this deal includes services.
And is now valued at $1 billion. It’s a testament to what the company is able to do in terms of investigating hacks, and also to the supreme importance of security going forward.
fortune.com  •  Share
It’s a cloud-based approach, now with Microsoft as a big investor. Virtual desktops are a fine idea in theory, maybe Frame’s approach can help take them mainstream.
blog.fra.me  •  Share
Media partner: GeekWire
All things data
This story is a little depressing, to be honest. But it also makes a lot of sense—I started to sense data overload might be an issue for fitness tracking, for example, and farmers apparently are no different. Blue River’s automatic pesticide-injecting robot is working, but it has a very specific purpose.
www.wsj.com  •  Share
CrowdFlower really wants to become a major source of data preparation for AI and other big data workloads, so it’s running a contest where it’s giving away a bunch of free work. The company also launched a new toolkit focused on computer vision data.
Speaking of data preparation, here’s a Boston-centric story featuring GE and startup Tamr to automate data-cleaning for its parts-procurement process.
fortune.com  •  Share
Listen the the ARCHITECHT Show podcast. New episodes every Thursday!
Did you enjoy this issue?
ARCHITECHT
The most interesting news, analysis, blog posts and research in cloud computing, artificial intelligence and software engineering. Delivered daily to your inbox. Curated by Derrick Harris. Check out the Architecht site at https://architecht.io
Carefully curated by ARCHITECHT with Revue. If you were forwarded this newsletter and you like it, you can subscribe here. If you don't want these updates anymore, please unsubscribe here.