Thousands of people gathered near Google's headquarters on Wednesday to hear the company's vision for the future. In past years, Google has used its developers' conference to unveil all sorts of shiny new toys and services. Not all of them have been smash hits, however. Google Glass had its big coming-out party at I/O in 2012, after all. Google TV was the star of 2010. And remember the Nexus Q, the orb-shaped music player that never even reached the market?

Welcome to Mossberg, a weekly commentary and reviews column on The Verge and Recode by veteran tech journalist Walt Mossberg, now an Executive Editor at The Verge and Editor at Large of Recode. Google announced something for everyone yesterday at its 10th annual I/O developer conference. There were more details of a new version of Android; new messaging and video-calling apps; a built-in new VR platform for Android; and a good-looking Amazon Echo-like smart speaker called Google Home. There was even a cool new research project called Instant Apps that will let users run portions of apps from the web without installing them first. But the biggest theme stressed by Google CEO Sundar Pichai and his lieutenants, over and over again throughout the two-hour keynote, was that Google is doubling down on artificial intelligence as the next great phase of computing.

We have seen the future, and artificial intelligence takes over. MarI/O, created by gamer Seth Bling, is just the tip of the iceberg. This AI had no idea how to play the game at first, but after some simple parameters were set in the form of "fitness" levels, it was incentivized to continue trying new ideas. You read that right, each level used new ideas, and it remembers what worked, while discarding its mistakes. After 34 steps, using NeuroEvolution of Augmenting Topologies, MarI/O figured out that jumping was the key to complete the stage.

Machine learning, which helps computers do things like understand complex voice commands and improve image search capabilities, can be taxing on traditional hardware. Google should know – over 100 of its products and features use this technology to run and improve themselves constantly. The company has revealed that over the past few years, it quietly developed its own custom processor for such tasks. The Tensor Processing Unit (TPU) is built expressly for running TensorFlow, Google's in-house machine learning system that it open-sourced last year. Some of the biggest names in tech are coming to TNW Conference in Amsterdam this May.

Martine Rothblatt, futurist and founder of Sirius XM, says by the time her biotech company's genetically modified transplant organs are in use, drones will likely deliver them. Rothblatt gave her view of the future at The Washington Post's Transformers conference Wednesday. Her United Therapeutics company, which has offices is Silver Spring, Md. and Research Triangle Park, N.C., is raising pigs with genome modifications its researchers hope will improve the animals' organs for transplant recipients. Pigs organs, because of their size and function, make good transplant material, but often the patient is trading their current disease for "a chronic organ rejection kind-of-disease that ultimately takes the life of many, if not most, people who receive transplants," she said. The company hopes to begin trials on organ transplants from genetically-modified pigs by the end of the decade, with regulatory approval ten years from now, Rothblatt said.

However, Google too seems to acknowledge, "great software shines brightest with great hardware underneath". The aim, the search giant says, was to see what they could accomplish with custom accelerators for machine learning applications. The TensorFlow-tailored TPU, which has been running inside the firm's data centers for more than a year now, delivered "an order of magnitude better-optimized performance per watt for machine learning". If that doesn't sound impressive, to put that in perspective Google said that the improvement is roughly equivalent to a technology fast-forwarding an approximate seven years into the future, or three generations of Moore's Law. Ultimately, it all comes down to crazy optimization.

Do we finally have real competition in the deep learning hardware space?

Pepper the robot participates in a Japanese ribbon-cutting ceremony earlier this year. Its manufacturer, SoftBank Robotics, is opening new offices in San Francisco and releasing a development kit for Android programmers. Japan-based SoftBank Robotics announced Wednesday at Google I/O, the company's annual developer's conference, that it is opening a new Pepper-focused outpost in San Francisco and unveiling an Android SDK, or software development kit, in the hopes of enticing programmers to write code for the robot. "Pepper is ultimately an unfinished product, and we just wanted to incentivize developers to expand the ways in which people can engage with a humanoid robot," says Steve Carlin, vice president of SoftBank Robotics Americas, which has an existing office in Boston. Asked if SoftBank will roll out at SDK for iOS developers, Carlin says he wouldn't rule anything out but "for the moment Android is the pervasive language."

I recently wrote a guide on recurrent networks in TensorFlow. That covered the basics but often we want to learn on sequences of variable lengths, possibly even within the same batch of training examples. In this post, I will explain how to use variable length sequences in TensorFlow and what implications they have on your model. Since TensorFlow unfolds our recurrent network for a given number of steps, we can only feed sequences of that shape to the network. We also want the input to have a fixed size so that we can represent a training batch as a single tensor of shape batch_size x max_length x frame_size.

The predictive powers of computers will work nicely in cases where reality does not change dramatically. However, it will fail in any case where there are dramatic, unpredictable, changes in the future. The authoritative science journal Nature announced recently that a computer designed by Google's DeepMind defeated a human master in the ancient Chinese board game, "Go." This impressive achievement once again raised the expectations for a predicted future in which computers will have artificial intelligence, with major media outlets worldwide touting this anticipated future. One of the major questions raised in response to DeepMind's achievement is what are the outer limits, if any, of intelligent machines?