Faulty Reward Functions in the Wild
openai.com
openai.com
> other measures of success.
This is also true for humans in poorly designed systems. For example, kids become experts at passing tests irrespective of mastering the material. In the workplace, employees become skilled at clocking extra time without finishing additional work. It's reasonable to say that this would eventually emerge in systems which approximate human behavior.
The video shown in the article could just as easily have been a human who just discovered the bug, and wants to troll a bit. The key difference is that a human would soon get bored. Our algorithms don't know about boredom outside of the domain of the reward function.
After playing with a bugged state, A human would lose just enough interest so as to keep playing the game, but without any further interest in the "bugged" state. A human is smart enough to know that there are various microstates of such a "bugged" state, and to ignore those instances as well.
The algorithm is smart enough to find the hack, but it's not smart enough to say "Hey, this is a non-solution, and I am not very happy about that". What is it that makes a human decide to lose interest in such a bugged state? Are these factors locally contained or are they due to external influences?
Repetition, the human brain has a reward function that is interested in finding new patterns. Using the same pattern to gain rewards has diminishing returns in the human brain, eventually we don't get enough reward and we try to find a new pattern. When this breaks down and the same pattern continues to get the same reward you can potentially fall into an addiction.
So in the case of this AI, simply diminishing its reward if it uses the same route every time to get that reward would prevent it from getting stuck in a loop.
If you want it to actually finish the race though, you might want to reward it a little for following the direction of the course. And it would make much more sense if it was rewarded for finishing the race first, humans are also a competitive bunch after all.
By not rewarding the AI for those things, they just did a very bad job at explaining the goals of the game.
Any publicly traded corporation (save a small handful with a non-traditional governance model) are metric-driven companies.
Modern corporations are paperclip maximizer functions executing on a network general-purpose biological computational engines tied together with powerpoint and email and excel spreadsheets.
Want to know what the AI of the future will look like? It will be a lot like Comcast, because it will be built by Comcast and harnessed to the corporate goals of Comcast and thus will have the same value system as Comcast.
The only thing it will lack is Comcast's institutional incompetence, as it will be Comcast's goals executing on dedicated hardware and not semi-autonomous employees. And it will build a dedicated model of every man and woman on the planet, and use that information to build a personalized profile that will determine exactly how many illegitimate charges it can cram on your bill before you'll suffer through a customized cancellation service that is calibrated to your personality and mental state to be just painful enough to drive you to the brink of suicide. And the only reason it's merely to be brink, is because a dead customer is an unprofitable one. (and if you think that's hyperbole, you have a far brighter view of the future than I do)
The code in the example isn't faulty. The goals are faulty in an non-obvious way, and drives the AI to optimize a solution to a fitness function in a way that humans might consider pathological behavior.
> to Comcast paving the future of the AI?
Comcast's goals are faulty in a non-obvious way. Comcast's goals are to maximize shareholder profit, and in doing so creates a culture of fraudulent billing and a horrifying gauntlet of a cancellation process that humans might consider pathological behavior. And AI is the next logical step to optimizing those corporate values.
It's nitpicking, but no, no big company has the goal of maximizing shareholder profit.
Most have the goal of maximizing board members profit, what is mostly aligned, but different in a non-obvious way from maximizing shareholder profit.
I think the core point is that corporations are fairly general-purpose machines, designed to do things too "big" for individual people. (But smaller and easier to analyze than "civilization.")
Their hardware stack is different, but lessons about their computational structure and design pitfalls may be applicable.
There’s something they used to call ‘the Myth of Shareholder Value.’ It goes, basically, like this: everything any corporate officer does, at any level, must be dedicated to only one thing, which is to increase shareholder value.
...
The Weyland Consortium are the high priests of the cult of Shareholder Value in today’s economy. The Consortium is less a corporation and more an algorithm, buying and selling corporations and extending its tendrils out through every sector of industry: research, transportation, you name it. The Consortium moves and acts like a living thing. No one executive or committee steers that ship. Shareholder Value is its only captain, and every decision made by its chief officers is pre-destined, an inevitable result of gears turning for decades, of market forces filtered through AIs running on corporate servers.
They have faith that they are doing the right thing. Research tells them they are not, and they either do not care, or, in my opinion, desperately guard their self image as a 'leader' over and above any concern for profit or long-term viability of the business. An AI would be permitted to take a strong role in making decisions for a company only insofar as it plays the role of toady, having been tweaked and misconfigured to ignore all facts which could result in it telling the executives that they are running 'their company' in the wrong way.
Why new companies run by competent people don't displace current companies today?
No, the AI won't be in charge. It won't set the overall goals. But it will be the truck drivers, the call center workers, the customer service specialists, the personal banking advisers.
When humans are told by the CEO of Wells Fargo, "you need to open 8 new banking accounts per day (because 8 rhymes with great) and we totally don't want you to cram unwanted products into existing customer portfolios, because that would be wrong wink-wink-nudge-nudge", most Wells Fargo employees are going understand what's being said (especially when they are disciplined for not acting illegally), and do the thing that their CEO tells them to do.
An AI given a fitness function to "open the maximum number of accounts, with a minimum probably of being caught performing illegal activity" is going to fuck over banking customers with a far more ruthless efficiency.
1) They're measuring incorrectly
2) Their "world view" is incorrect
In either case as long as they continually re-assess the above 2, their metrics will change and (ideally) optimize toward a more accurate reflection of reality.
I'd rather have evidence that my world view is incorrect, and learn this as quickly as possible so I can adapt to what the data is telling me.
It's certainly not easy to do this right (and may be impossible to perfect), but I think it's an objectively better option than, say, "going with your gut".
"When a measure becomes a target, it ceases to be a good measure."