AI Report

[Phoebe]

I think there's a link to the actual report in this article but if not it's at the Electronic Frontier Foundation. I think there's a link to the actual report in this article but if not it's at the Electronic Frontier Foundation. Unrelated tangent: for a small donation to the eff you can get a really Kick-Ass T-shirt with a cat. I'm just saying. Anyway back to the point, has anyone read this report about concerns over artificial intelligence capacities and what we would have to do to control them appropriately? I'm curious about the report generally and the issue generally.

[bralbovsky]

One statement in the article is that AI has failed to live up to it's hype. While the article says that's changing, many in the industry disagree (as do I) even Elon Musk (who sells a personal flamethrower just for fun, in case you were worried that he was growing risk averse) wants a ban on lethal robots and drones.

Brains are complex. They evolved over eons, so the lizard parts that were useful are still in there. Something so layered is difficult to reverse engineer. We barely know how people learn when they're sitting at our kitchen table. How do we break that down so that it can be entered into code?
What we do, and why we do it are so disconnected sometimes, from the thirteen year old who might truly not know what caused him to do the stupid thing he's in trouble for, to the criminal who is way more afraid of Russian mobsters than he is of federal prison.

I'm messing this up, sorry. Let me begin with three challenges (not the only three), and you'll likely be able to deduce the hazards that result, but if not I can post again
1 - Emulating the functions we do automatically turns out to be really hard. Think of your job. I don't know what it is, but if you're any good at it, there are tons of things a well-trained rookie isn't as good at. How does he understand that while his miscalculation may not have immediately cost the company money, it annoyed an important client, who will add that annoyance into the next buying decision. Those metrics, the expression on a customer's face, the angle of a product on a shelf...we don't even notice we notice them, so they're tough to input into the computer's learning algorithm.
I have a student who is trying to get computers to read insurance papers, doctor's notes, diagnosis reports. They take time to read, because they're unique. Finding the most important information in each one is a challenge, partly because the nurse doesn't know the blood pressure reading he took is going to connect to the blood numbers, which connects to the referred pain the patient reports. We do that synthesis pretty well, with training. But it's not automatic, which is why it's the subject of so many cop shows.
The obstacle of soft metrics is huge and difficult. Exploiting the code in a game or social media algorithm turns out to be one source of abuse. Machines love numbers, fiddle with them and you can make the machine do almost anything, even things that don't make sense.

2 - Life is analog. Paper, for example, is processed, not manufactured. Machines have to deal with varying fiber density, moisture absorption, everything that makes paper unique also makes paper jam. Roads are uneven and subject to mother nature, so navigating them is much more difficult than navigating a map. This doesn't even add the complication that life is 3D while screens (and some satellite perception) is not. So sometimes, apart from not understanding what they're looking at, the machines make incorrect assumptions or can't deal with values outside their working parameters. In a security situation, a lost five year old, or an endangered bird gets shot as an intruder because, those things don't exist for the computer

3 - We succeed by a series of non-fatal mistakes. This is the AI ceiling that many of my recent conversations have wrestled with. The Jimi Hendricks note that convinced us all he was a genius was an almost wrong note. Certainly by conservative definitions, it was absolutely a wrong note, but it was the absolute perfect wrong note, held for the perfect incorrect time. Let's not even enter the creativity conversation yet, just the virtuosity conversation. Machines might love that predictability and repetition, but evolution didn't happen that way. Evolution is the result of perfect and useful errors.
An athlete gets better, faster. Her bones get stronger with stress. Machines just age. T2 is scary, but it also happens to defy laws of physics. While one might argue that the first 500 miles allows the car to smooth out operations, after that, it's all downhill. Meanwhile, the athlete, the musician, gets better. Typewriters were designed to be slow because of material limitations. In order to be faster they needed a redesign. The printer has a peak speed, which is way faster than a monk, but it'll never get better.

If AI, machine learning, is really going to take off, it has to figure out which mistakes are the best ones, what are the really important metrics in a world populated by unique objects, and then figure out how to grow from them.
Sadly, it's much easier to fool the AI by knowing what its parameters are and exploiting them. If the cars follow the white lines on the side of the road, repaint them, and all the cars will crash. If the security program recognizes certain features, maybe we build a mask, or maybe we alter the code. If machines emulate learning, it becomes very critical what they are taught.

This is too long, my apologies, and off the top of my head, but there's a beginning.