I run the compute in multiple large scale datacenters. This is such a silly sensationalist article. There is nothing "AI" about this. It is nothing more than a programmed response to inputs.
Also, I don’t believe that’s the proper usage of that term. It would go like this “It cannot be magical, because math is logical.”
Edit for maxdepth reply: The example is looking at any code for a model released in TensorFlow or PyTorch, because it’s 1) quite (exactly) concrete 2) linear algebra with autograd bolted on top. I did a quick search and there are plenty of translators on GitHub that demonstrate that. Here’s an annotated example of neural nets from PyTorch https://pytorch.org/tutorials/beginner/blitz/neural_networks.... What’s incomprehensible about it and would the authors of a ML paper also describe those features as incomprehensible?
In the 1990's, an AI Corp developer (a woman friend of mine) completely automated the management of a large IBM mainframe using old fashioned AI. The computer worked in complete darkness. She called it "lights out computing".
This (presumably) refers to "Good ol' Fashioned AI" aka Symbolic AI (https://en.wikipedia.org/wiki/Symbolic_artificial_intelligen...) . It's a bit of an over-simplification to say it's a bunch of if statements, but it does involve encoding a bunch of rules/symbols by humans (as opposed to learning from data). It's a largely dead paradigm, though aspects of it are still found in various modern approaches (eg task and motion planning, knowledge bases, etc.) often in combination with learning approaches.
Having seen the code for a couple of "intelligent assistant" features: Yeah, basically a bunch of "if" statements. (The IA for the Apple Newton was a ton of special cases. It was clever, but certainly not 'intelligent').
Thus, the technology of AI has improved from a sea of spaghetti code that no human can understand, to a bunch of neural net weights that no human can understand. I love the march of progress. :-)
Having worked in MS's DCs, they are, in fact, hazardous. They absolutely push the limits on heat exposure for humans -- those hot aisles are HOT, and we all carried heat index alarms as a result.
Further, I think its weird you don't think of the sound as a hazard. When its so loud you must wear 30+NRR hearing protection, it sure seems like a hazard to me.
They absolutely push the limits on heat exposure for humans
That sounds very different from my experience (not with MS DCs) --- where it was loud (though not quite as loud as e.g. an automotive shop) and windy, but the temperature was kept at a controlled 68F (20C) and alerts were raised if it got even slightly hotter than that.
Yes, I suppose there is more risk than sitting at a desk all day, but I think the parent is comparing it to many other jobs like those in the heavy equipment industry, mining, farming, petro, etc.
I'd agree its very different even from my own experience at other data centers. MS really were pushing it in the locations I worked. In early May, the hot aisles were already topping 112F! Given that, the policy in place was, IIRC, 15 minutes of work w/ a mandatory 45m break & fluid consumption.
Certainly there are riskier jobs. Certainly, by and large, data centers are safe places to work. But they are not without hazards and I (obviously) find it irksome when those hazards are dismissed so easily.
I've toured power plants, both hydro and coal, which are dangerous places. On a scale of 1-10 with a corporate office being 1 and a power plant being 10, I'd have rated a data center at 1. But now having seen the inside of one, I'd rate it more like 3 or 4. The biggest danger is the high voltage, but there are more mundane dangers like hearing loss, heavy equipment, heat, room-pressure, holes in the floor, etc. It's an industrial facility for sure, and PPE and safety procedures are a BIG DEAL for the workers.
