onetwotree

Working for a "digital innovation lab" where a big part of my job is sorting out things that developers can actually use to make cool products from Gartner report buzzword hell, I have a few thoughts about this.

Historically, AI has seen boom and bust cycles. AI, unlike other buzzword technologies, has a large cultural mythos about it - Asimov didn't write "I, EC2 Instance", after all. Thus, non-technical people have very strong and often unrealistic set of expectations around AI, so every time there are technological advances, people expect that general AI is right around the corner, and are disappointed when it doesn't materialize. Right now, we're experiencing a boom cycle as a result of the emergence of deep learning techniques. After every boom, funding dries up and investors loose interest, but the techniques that really work stick around and become part of a developers everyday toolkit.

Like all "black sorcery" technologies, AI has a ways to go in terms of building convenient tooling. Tensorflow is a huge improvement over writing low level CUDA code, but it's still too low level for folks without a strong background in mathematics and machine learning. That said, it's been improving in usability, documentation, and tooling, and just a couple of weeks I was part of a hackathon where we turned their pet recognition demo[1] into software to detect objects of interest to my company in satellite images.

At an even higher level are a number of startups (such as Clarifai[2]) that offer AI-as-a-service. In Clarifai's case, you can train your own image recognition models and apply them with a few lines in your favorite programming language (yeah, yeah, I'm shilling them a bit, but I really like their product ;-).

So at the end of the day, I think we'll all be building various kinds of AI into our products in the not-so-distant future, but you won't really need to go deep into tensorflow and similarly low level tools to do so.

Li-Ion battery degradation is generally modeled as two (roughly additive) components, called calendar and cycle aging. Calendar aging is what you're talking about in your comment - it's basically determined by the temperature and state of charge. Both of these impact the rate of chemical processes that lead to loss of lithium and active material. This is what happens when the battery is sitting on the shelf.

Cycle aging, on the other hand, happens when you charge or discharge the battery and is driven mostly by the actual volume change in the anode and cathode when they get lithiated (it's quite significant, up to 15% or so in some materials). This introduces mechanical stress, which can break the protective film that forms between the electrodes and electrolyte and allow chemical degradation to proceed at a faster rate. The degree of mechanical stress is mostly determined by the depth of discharge, although rate of charge and discharge is also believed to be important. Cycle degradation is also impacted by the temperature at which the battery is cycled, in a similar way to calendar aging.

With high temperature variance (low temperatures will wreck your batteries too, and I can't recommend keeping them at 5C), calendar aging is the dominant mode. This is absolutely the case with cell phone batteries. However, electric vehicles, especially higher end ones like the Tesla, have an active cooling system that keeps the battery at a constant temperature, even when the car is "off". This is why many EVB warranties will be voided if you let them run out of charge for more than 14 days - at that point, the cooling system isn't working and your poor battery is at the mercy of the elements.

Under closely controlled temperatures, cycle aging is the most important factor, and it's dominated by depth of discharge. If you drain the battery as far as it will allow you every day, it's going to be hosed. If you only go down to about 80% of the allowed charge every other day, it'll last for quite a while. These numbers are roughly correlated, obviously, with the mileage on the car, but the relationship is not simple, and I'd caution against considering mileage to be a good determinant of battery degradation.

Some good sources if you want to learn more: