Risks of Artificial Intelligence

> Drones with chainsaws however are just plain badass

Original battle bots ++

Indiscriminate visual targeting is a potentially hard problem if you're doing it against adversaries. One of the things that's underappreciated about machine learning is, since it's statistical, it can only be about anomaly detection, so the solution is to not be anomalous. You could conceivably hide from thermal imaging by using ice packs, and anyway, you only have to get to within the range of indifferentiability from the ordinary noisiness of the environment. Optical techniques suffer from similar problems, except that it's even easier to hide. Not to mention the occasinal voodoo that comes out about metamaterials.

Did someone say drones with chainsaws?

https://instagram.com/p/BU2TsdlFtf6/

I'm no fan of the A.I. hype of today, but I can comment that deep learning actually makes the "contentious decision" easier to detect. With the "old way", i.e. jam some features into an SVM, there was no accountability / confidence. At least I've never seen a reliable SVM-based confidence score.

But, for example, confidence scores are actually a byproduct of running a modern CNN for image classification. This is very useful for "finding mistakes" (assuming you have 100% accurate training data, and other prerequisites). It even helps fixing issues in the training data, we use it all the time.

That being said, the big dirty secret that nobody's pointing out is: there has been no real "theory-level" advancement in AI in the last 20 years that have borne fruit (at least in computer vision and neural nets). The models are just bigger (and much better) with some slight tweaks (dropout, etc). It's hardware that's made that possible. Academics have done a thorough and wonderful search of the model space and have some great finds. But I think we still need a few solid theoretical leaps before the singularity.

Technically it's still the programmer who made all the decisions, at least with the current state of NNs. True decision making is philosophically up there with the 'what is consciousness' problem.

When the blame game is played though, the end result would likely be the corporation in question getting sued for a lot of money, who then might choose to take internal action and fire certain employees.

Whoever is using it to make said decisions.

> Some humans are likely to become more powerful than other humans some day. Most such humans will by default develop instrumental subgoals that conflict with other human interests.

Isn't this already true?

That is assuming that humans have control of their AI.

That's why I don't find the paper maximizer analogy very convincing.

Still it's possible AI systems put in charge of running society could result in some undesirable future because we weren't careful enough in designing their goals.

But then again, we have that problem with whoever we put into office.

> If the AI is so dumb that it doesn't know that happiness means different things to different people...

Arguably, we humans have general intelligence, and still some fail to see that point (some might even argue intelligently against it!). Collectively, we have taken over the world regardless.

How do you optimize happiness for all people?

It is not obvious to me that strong AI will recognize and fix all bugs in itself to become what its creators intended it to be.

Fixing bugs which harm its performance is a valid instrumental goal, sure. Fixing all other bugs is not.

Taking over the world is basically a hardware problem. Understanding that we don't want plain happiness is a software problem.

In a sense, nuclear weapons could "take over the world" today. There isn't that much intelligence required.

So we should regulate AI to slow advancements and let other countries take the lead? Would you rather be taken over by China's general AI or the USA's?

Imagine this AI had access to the internet (after all, it'd be using cloud resources). Suppose it was smarter than the smartest human ever. I'm not that smart, and here's some things I can think it could do that would be scary:

- Find security holes in all computer systems, then replicate itself onto vulnerable systems

- Intercept, fake, modify phone calls (including between government members)

- Procure large amounts of money and then pay people to execute its bidding

- Make itself more intelligent

But again, it'd be smarter than me, so these ideas I just came up with in the 5 minutes since I saw your question wouldn't hold a candle to theoretical worst case.

Elon Musk had a thought provoking take on this question. I'll roughly summarize, and add some nuances he didn't mention. He was speaking to a room full of state governors, most of whom seemed fairly unconvinced by his concerns.

He said that he didn't believe his example was real, but that it was illustrative.

Consider an AI whose goal is to maximize stock market returns. So this AI ingests vast sums of historical data, of all kinds, and it considers https://en.wikipedia.org/wiki/Korean_Air_Lines_Flight_007 where the Soviet Union shot down a Korean Air Lines 747 in 1983 when it deviated slightly from its path from Anchorage to Seoul and entered the edge of Soviet protected airspace. About the same time, there was a nearby US aerial reconnaissance mission.

This event greatly increased tensions during some of the worst years of the cold war, and as a result, certain kinds of stock market issues moved in certain directions.

Considering this incident, and the state of the world in 2014, the AI takes three actions:

1. Buys/shorts the same and/or related stocks, as applicable.

2. Creates some fake intelligence chatter, consumed by the Russians, that where was going to be an areal target of interest over east Ukraine on a certain day and time. The location and direction would not clearly match known commercial airline flights.

3. Break into the flight computer of Malaysia Airlines Flight 17 on 17-July-2014, and cause it to fly slightly off course. https://en.wikipedia.org/wiki/Malaysia_Airlines_Flight_17

The result? Another civilian airliner was shot down by the Russians, which increased global political tensions, and probably caused some stocks to move in relatively predictable ways. In this example, the AI met the goal it was given in an exceedingly unexpected and unpleasant way.

This is clearly NOT an end of the world scenario, and is a hypothetical.

But it is, in my mind, illustrative of some of the dangers of AI. Systems that can do powerful, unpredictable things in order to meet mundane and reasonable goals.

The design seems to be heavily influenced by Edward Tufte and is likely based on https://edwardtufte.github.io/tufte-css/

I think the better question is: Should we really be taking real-life advice from cheesy action movies?

Now my question is: who programmed the rogue AI? We should be even more afraid of him. /s

Whether or not you're joking, I think the real problem with using Terminator as an example is that it's overly optimistic. The story is roughly:

1) US military builds up arsenal of autonomous killing machines and nuclear missiles

2) US military connects all of these to the Internet

3) US military creates a powerful AI which takes control of this arsenal (whether it was put in charge or hacks in seems to vary across the movies)

4) AI "becomes self-aware"

5) AI tries to wipe out humanity

Almost all of the discussion around this focuses on step 4, either by asking if/when an AI will "become self aware", or by trying to explain why that's meaningless and/or unlikely.

Meanwhile I think the real dangers are steps 1 and 2, which seem to be proceeding without much public outcry.

Yes, there are rogue AGI scenarios which end badly for everyone; but there are also issues of hacking (state-sponsored or otherwise), and/or terrorism (homegrown or otherwise).

It may have made political sense to build up ever-larger nuclear arsenals during the cold war, but these days it seems like that's just increasing the risk of accident or misuse.