All posts tagged thresholds

A couple of years back I purchased an AI-powered energy monitoring system for my home.  It clips on to the power mains and monitors amperage/wattage.  I can view an a graph showing energy usage over time, which is really quite helpful to keep tabs on my electricity consumption at a time when electricity is expensive.

The AI part identifies what devices are drawing power in my house.  Based simply on wattage patterns, so they claim, the app will tell me this device is a light, that device is an air conditioner, and so on.  An electric oven, for example, consumes so much power and switches itself on and off in such a pattern that AI can identify it.  The company has a large database of all of the sorts of products that can be plugged into an outlet, and it uses its database to figure out what you have connected.

So far my AI energy monitor has identified ten different heaters in my house.  That’s really cool, except for the fact that I have exactly one heater.  When the message popped up saying “We’ve identified a new device!  Heater #10!”, I must admit I wasn’t surprised.  It did raise an eyebrow, however, given that it was summer and over 100 degrees (38 C) outside.  At the very least, you’d think the algorithm could correlate location and weather data with its guesses.

Many “futurists” who lurk around Silicon Valley believe in a few years we’ll live for ever when we merge our brains with AI.  I’ve noticed that most of these “futurists” have no technological expertise at all.  Usually they’re journalists or marketing experts.  I, on the other hand, deal with technology every day, and it leaves me more than a little skeptical of the “AI” wave that’s been sweeping over the Valley for a few years.

Of course, once the “analysts” identify a trend, all of us vendors need to move on it.  (“SASE was hot last fall, but this season SSE is in!”)  A part of that involves labeling things with the latest buzzword even when they have nothing to do with it.  (Don’t get me started on “controllers”…)  One vendor has a tool that opens a TAC case after detecting a problem.  They call this something like “AI-driven issue resolution.”  Never mind that a human being gets the TAC case and has to troubleshoot it–this is the exact opposite of AI.  We can broaden the term to mean a computer doing anything on its own, in this case calling a human.  Hey, is there a better indicator of intelligence than asking for help?

Dynamic baselines are neat.  I remember finding the threshold altering capabilities in NMS tools useless back in the 90’s.  Do I set it at 50% of bandwidth?  60%?  80%?  Dynamic baselining determines the normal traffic (or whatever) level at a given time, and sets a variable threshold based on historical data.  It’s AI, I suppose, but it’s basically just pattern analysis.

True issue resolution is a remarkably harder problem.  I once sat in a product meeting where we had been asked to determine all of the different scenarios the tool we were developing would be able to troubleshoot.  Then we were to determine the steps the “AI” would take (i.e., what CLI to execute.)  We built slide after slide, racking our brains for all the ways networks fail and how we’d troubleshoot them.

The problem with this approach is that if you think of 100 ways networks fail, when a customer deploys the product it will fail in the 101st way.  Networks are large distributed systems, running multiple protocols, connecting multiple operating systems, with different media types and they have ways of failing, sometimes spectacularly, that nobody ever thinks about.  A human being can think adaptively and dynamically in a way that a computer cannot.  Troubleshooting an outage involves collecting data from multiple sources, and then thinking through the problem until a resolution is found.  How many times, when I was in TAC, did I grab two or three other engineers to sit around a whiteboard and debate what the problem could be?  Using our collective knowledge and experience, bouncing ideas off of one another, we would often come up with creative approaches to the problem at hand and solve it.  I just don’t see AI doing that.  So, maybe it’s a good thing it phones home for help.

I do see a role for AI and its analysis capabilities in providing troubleshooting information on common problems.  Also, data can be a problem for humans to process.  We’re inundated by numbers and often cannot easily find patterns in what we are presented.  AI-type tools can help to aggregate and analyze data from numerous sources in a single place.  So, I’m by no means saying we should be stuck in 1995 for our NMS tools.  But I don’t see AI tools replacing network operations teams any time soon, despite what may be sold.

And I certainly have no plans to live forever by fusing my brain with a computer.  We can leave that to science fiction writers, and their more respectable colleagues, the futurists.