Tag: google

Datacenter performance

How do you know how well your large kubernetes cluster is performing? Is a particular change worth deploying? Can you quantify the ROI? To do that, you’re going to need some WSC-wide metric of performance. Not so easy! The WSC may be running 1000s of distinct jobs all sharing the same underlying resources. Developing a load-testing benchmark workload to accurately model this is ‘practically impossible.’ Therefore, we need a method that lets us evaluate performance in a live production environment. Google’s answer is the Warehouse Scale performance Meter (WSMeter), “a methodology to efficiently and accurately evaluate a WSC’s performance using a live production environment.” At WSC scale, even small improvements can translate into considerable cost reductions. WSMeter’s low-risk, low-cost approach encourages more aggressive evaluation of potential new features.

TFX

TFX also includes a suite of data transformations supporting feature wrangling. As an example, TFX can generate feature-to-integer mappings, known as vocabularies. It’s easy to mess things up when transformations differ in subtle ways between training and serving. TFX automatically exports any data transformations as part of the trained model to help avoid these issues.

Waymo World Simulation

At any time, there are now 25000 virtual self-driving cars making their way through fully modeled versions of Austin, Mountain View, and Phoenix, as well as test-track scenarios. Waymo might simulate driving down a particularly tricky road 100Ks of times in 1 day. Collectively, they now drive 13M km per day in the virtual world. In 2016, they logged 4B virtual km versus a little over 5M km by Google’s IRL self-driving cars that run on public roads. And crucially, the virtual km focus on what Waymo people invariably call “interesting” km in which they might learn something new. These are not boring highway commuter km.

Outrageously Large Neural Networks

We finally realize the promise of conditional computation, achieving greater than 1000x improvements in model capacity with only minor losses in computational efficiency on modern GPU clusters. We introduce a Sparsely-Gated Mixture-of-Experts layer (MoE), consisting of up to 1000s of feed-forward sub-networks. We present model architectures in which a MoE with up to 137B parameters is applied convolutionally between stacked LSTM layers. On large language modeling and machine translation benchmarks, these models achieve significantly better results than state-of-the-art at lower computational cost.

Undercover AlphaGo

The account is simply called “Master”, and since the start of the new year it has made a habit out of trashing some of the world’s best Go professionals. It’s already beaten Ke Jie twice, who is currently the highest ranked Go player in the world. AlphaGo, incidentally, is #2. Ke Jie was “a bit shocked … just repeating ‘it’s too strong'”. By January 3, the number of probably-but-we-can’t-officially-say AI sanctioned beatings had risen to 41-zip

2017-01-04: It’s alphago

The Monkey and the Apple

It has been waaaay more work than I expected. Starting with a more-or-less working game, and tweaking it to work on Cloud and mobile — I mean, come on, how hard can it be, really? Turns out, yeah, yep, very hard. Stupidly hard. Especially since out of brand loyalty I chose Google’s cloud platform, which 3 or 4 years ago was pretty raw. And let’s face it, iOS APIs have evolved a ton in that timeframe as well. So even as “recently” as 2013 I was working with some pretty immature technology stacks, all of which have come leaps and bounds since then.

AMP for standardized measurement

if amp v2 succeeds, we’ll drain the swamp that is today’s web and abp will be unnecessary. this is a far preferable outcome than a bunch of walled gardens.

AMP, through its established `amp-analytics` mechanism, already ships with all the code to perform these measurements. It is vendor neutral and supports a wide range of metrics. This means ads can take advantage of the same “instrument once, report many times” feature that benefits AMP pages today, completely eliminating the bandwidth and runtime cost outlined above.

Transportation-as-a-Service

any discussion of the threat self-driving cars poses to Uber tends to imagine a world where there are magically 10Ks if not millions of self-driving cars everywhere immediately. That simply isn’t practical from a pure logistics standpoint; the time it will take to build all of those cars — and, crucially, get government approval — is time Uber has to catch up.

Moreover, it’s not at all clear that Google will be willing to make the sort of investment necessary to build a self-driving fleet that could take on Uber. The company’s recent scale-back of Google Fiber is instructive in this regard: it is very difficult for a company built on search advertising margins to stomach the capital costs entailed in building out a fleet capable of challenging Uber in more than 1 or 2 markets.

Finally, as the incumbent in the transportation-as-a-service space Uber has the advantage of only needing to be good-enough. To the degree the company can build out UberPool and UberCommute, they can ensure that their own self-driving cars get first consideration from consumers trained to open their app whenever they are out and about.