Cataloging the problems facing AI researchers is a cross between a parenting manual and a management book
The comparison to parenting is interesting since a loving bond between parent and child is needed. We talk about AI like it’s something alien. The significance of a personally identified safe, nurturing relationship is worth considering if we want baby Skynet to learn how to love the other baby AIs … people too.
The real question is is there a single ideal prosocial behavioral schema we want all robots to adhere to?
It seems intuitive that we don’t want all of our robots to behave like Ed Gein or Martin Shrekli, but … maybe some of them should?
Do we want half our robots to be introverts and half extroverts?
Do we calibrate their risk tolerances uniformly?
Brave New World indeed.
I think that’s an interesting focus on prosocial behaviors. What about the question of whether we first need teach and practice for ourselves more prosocial behaviors?
How do we know when someone has learned enough about humane practice toward one another, toward other species, to do a competent job teaching AI devices?
Isn’t it tough to teach what we don’t adequately know and practice ourselves?
Having just read the Wired article on Google’s DeepMind (here’s an Engadget writeup b/c Wired’s anti-adblocker pisses me off), one of the takeaways appears to be that after Lee Sedol was beaten by DeepMind, he went on to win further matches because of playing DeepMind. That is, DeepMind’s own…alien(?) intuition, or at least the results of it, heightened Sedol’s game-playing intuition. So if we can further our own intuition/learning from playing games against an AI, perhaps with the right AI we’ll learn a little more…humanity
In the near term, designing robot agents will have more in common with training a high school drop-out to work on a warehouse floor: it won’t know much, so it will need some task training, some general safety prohibitions, some baby-proofing to the environment, and a whole lot of procedural infrastructure to lean on when things go wrong.
But that’s also going to be very different from the sort of AI that would be of use to, say, an architecture or engineering firm, or a sales department, or a customer service line. For those you benefit from much more advanced learning, but you also have a lot less risk to permanently breaking something.
It’s important to recognize that when we speak of AI as a single body of expertise, we’re tricking ourselves into applying white-collar algorithms to blue-collar robots. A number of the issues they highlight specifically arise when you have a much smarter algorithm than is really necessary.
For example, it’s true that “an agent optimizing [some] objective function might thus engage in major disruptions of the broader environment if doing so provides even a tiny advantage for the task at hand.” But that assumes you reward or seek out such tiny advantages, which is the sort of thing you care about a lot in the computer science of complex algorithms, but probably should take a back seat in real-world robotics.
For example, suppose you have a box-moving robot. Now let’s suppose we’re moving boxes from a moving truck to a customer’s apartment. Why would you even attempt to optimize the problem, when all you need is a “good enough” plan? If you’re doing a rare task in a complex environment (a new home), you should avoid the distraction of optimization because the one-time gains are negligible. On the other hand, if it were in a warehouse, it would be doing a frequent task, so there’s an opportunity to aggregate tiny time savings into a real benefit. But if that’s the case, the operator is probably also inclined to specializing the environment to make it as safe, simple, and distraction-free as possible, which as an upshot allows the agent to optimize safely.
Remove the goal of optimization, and your diminish the benefit of gaming the system and better protect the complex environment. Remove the environmental hazards, and you simplify the task and limit the risk of optimization.
Teaching any person about any topic is nothing like coding instructions for a computing device. And if we’re teaching sci-fi AI, then it would be especially important to emphasize the humanity of the relationship . . . trust, rapport and love are first.
OTHOH, I understand there’s such a thing as “smarter” code, and no argument there.
Code that lets a camerabox device distinguish more marketable cherries from less marketable cherries on a transport and activate air jets for sorting at some point gets iterated and iterated into something the marketing department will call AI.
Yes, that’s different.
As I mentioned teaching a worker is only one small part of creating a solution to solve some problem. Adapting a problem to solution partially based on a human worker has a lot of the same challenges that a solution involving a robot worker would have. And because of that the total solution has to be more than just the worker. For example, robots are basically safety-blind: it’s a concept they don’t understand at all. Humans generally have some idea of safety, but eventually you’re going to run into a really naïve human who is similarly gets himself hurt and ruins it for the rest of us.
To avoid that, you need we defined procedures for workers, and a slightly sanitized environment. That way you reduce the hazards to something you can easily enumerate articulate – stay away from that saw, don’t get that line wet, fire is hot – either to a human or to a robot. In either case, the naïve worker is free to do his/her/its job as trained because you streamlined the training process to include just the job, and not a broad notion of safety. Procedural infrastructure and an engineered environment take some of the burden off training… and much of the freedom to do dangerous things
Now, if there were no naïve humans (or at least no laws protecting them), then training pretty much every (surviving) human would be radically different from programming a machine. No human is as dense as any robot.
But in any given task, there probably exists at least one human who approaches the density of a robot. Writing procedures to protecting all the humans from their lapses in judgment approaches constructing a plan to protect the robots from their lack of judgment.
And so, the processes become very similar: remove everything you can control that can go wrong; alert the worker to the ways in which remaining things are known to be able to go wrong; give the worker a fallback plan, such as “call a manger over and wait”; give the work basic safety training on how to avoid the dangers of the dangerous things he/she/it must work with.
But that is so far away from possible that it’s pretty much pointless worrying about. You may as well be writing a manual on the care and feeding of dragons. We’re not building animals or children, we’re building tools that can manipulate matter, the way we already have tools to manipulate data. Whatever we come up with, we will probably be able to simulate basic job competence long before we can simulate love.
These quotes just keep on being useful.
We are no longer particularly in the business of writing software to perform specific tasks. We now teach the software how to learn, and in the primary bonding process it molds itself around the task to be performed. The feedback loop never really ends, so a tenth year polysentience can be a priceless jewel or a psychotic wreck, but it is
the primary bonding process–the childhood, if you will–that has the most far-reaching repercussions.
- Bad’l Ron, Wakener, Morgan Polysoft
Took me a moment to recognize (ironic, because I’ve been playing Civ:BE for the past week). I’m still amazed at how well they did the philosophy in that game, both overt and covert, and how it was an excellent jumping-off point for those who wanted to learn more on their own regarding those concepts.
Thank you for sharing this detailed description of the differences in teaching humans and robots. The notion of safety-blindness in robots is a helpful way to show how the solutions need to engage context. It’s interesting, and I don’t often get the chance to to learn about the subject from someone knowledgeable.
I’m also interested in how social relationships inform humane interaction which is a different though not a pressing question for many engineers working on AI. Effective teaching requires a rapport, a trusting relationship. What does that relationship look like with non-humans? What if model relationships among humans are in short supply?
If we were attending Pres. Obama’s beer summit, I’d quibble with you about this point.
It’s not that I think engineers are closer to making HAL than they seem to be. It’s that I’m not sure that humans aren’t closer to making themselves into HAL.
I think that humans-teaching-each-other-to-behave-like-inhumane-robots is a concern even if we won’t have the opportunity — or responsibility — to teach AI to machines for another 1,000 years.
Just stick to sexbots and all will be well.
There’s no wrong way to apply that.
Nope. No way at all.
Offered without comment, as I’ve not read through all the posts below (most, but not all):
But now that I think about it, Clark’s statement, “need more diversity”, points to @hello_friends thought about AIs needing some human emotional training, too.
This topic was automatically closed after 5 days. New replies are no longer allowed.