AI and machine learning
I can't tell whether anything in this report is binding, or if these are just a bunch of suggestions. And while I'm normally all for making sure web companies are kept in check, I would note that health care isn't exactly the world's most ethical industry. Yeah, we probably don't want Alphabet having too much influence in that field, but if it can help save lives and improve efficiency, we might not want to put up too many barriers either.
On the one hand, this is a great example of how companies like Amazon are making / saving money from machine learning for from using it internally than from selling cloud services. In this case, the article is focused on algorithms that "negotiate" prices with vendors and have already resulted in employee resignations and reorganization. What's unclear is how the tradeoff between efficiency and corporate culture will play out, if the latter has any effect at all. You must read it.
This is a great tool to bring developers into the AWS fold for AI and IoT applications, but can we be done with smart cameras, already? Aside from surveillance and a few edge cases, I'm running low on viable use cases that don't present an unnecessary privacy risk.
This sounds like a pretty big step in NLP, almost like a refined version of word2vec (but I could be way off on the technical part of that). The real goal is to build systems that can determine the meaning of ambiguous words by analyzing their context.
There's a good discussion in the comments section here, too. And I would bet that, at least in the near-to mid-term, relatively short and predefined routes (shuttles, essentially) are going to be the primary and most valuable application of driverless cars. There's still too much danger and uncertainty on the open road.
Yes, and it should. There are certain fields where artistry and fallibility are part of the appeal, but no one needs expensive and potentially dangerous testing, drugs, etc, if a machine could prescribe something better the first time.
So this is really smart, if you buy into the argument that companies need to build AI (and all products, really) to serve the audience they know. Facebook has opened an AI office in Africa, too. Plus, there are universities with smart students and faculty all over the world.
The company, called AutoLab AI has some Autodesk veterans behind it, which is a pretty good crew to have for any sort of systems-design software. It seems the company has about 400 employees and $200 million in funding, but little else is public.
Twitter now uses TensorFlow. Like so many other people and companies (but certainly not all) doing deep learning. I had forgotten that it acquired a computer vision startup back in 2015.
This is early work, but important, especially for robotics, it would seem. Essentially, DeepMind has built a system that can predict the 3-D properties of a 2-D image -- things like what it looks like from other views, shadows, texture, etc. Anything moving in space needs to have a sense of what's next or what's a better way to approach a situation.
This is the Uber block, apparently. And here's another effort that's (1) really useful for Uber and its customers, and (2) potentially really useful for related applications. Uber has to tackle everything from phishing to GPS spoofing.
I can't help but think this data and approach could also be used to solve more pressing problems than food delivery. Uber is collecting a lot of info on parking, traffic, wait times and other things that you'd think could optimize traffic planning overall.
This is one of those pieces of research that sounds impressive, but then you realize we don't need machines that can solve Rubik's Cubes. But, as the article notes, "The real test, of course, will be how this approach copes with more complex problems such as protein folding."
This is kind of creepy -- a system that alters images of people with their eyes closes (or blinking) to show them with their eyes open. I think I could imagine a few valid use cases, but mostly I'm left thinking it's another step toward not trusting images and another way for the vain to perfect their online presences.
CleanNet is a potentially very useful piece of research from Microsoft, designed to clean up noisy image datasets. The goal is to reduce the grunt work / cost of manual labeling by at least pruning the a good portion of irrelevant images from raw datasets.