HomeBlogPapersTechWolf

Second Order Intelligence

13 November, 2020 - 6 min read

Every once in a while, a new development within artificial intelligence sends the world into a flurry of excitement -- did we crack the code of actual artificial general intelligence this time? While the media might forget about this every so often, the answer from experts is pretty clear: we most definitely have not, and it might take us decades more if not longer. In fact, there's some active discussion ongoing about whether it is even possible for AI to gain true understanding of the world, rather than just being able to repeat and extrapolate patterns. Nevertheless, even when we're armed with a solid dose of cynicism, AI models can still surprise us with their intelligence. In my experience, this feeling of delight often originates from a clever setup, rather than raw processing power.

Losing the Imitation Game

In 1936, Alan Turing described the imitation game -- a test (nowadays called turing test) in which a person enters a chatroom and has to figure out through conversation whether their counterpart is a human or a computer. It's a simple enough concept, but a very effective one at that: give a computer and a real person the same communication means and see whether you can still notice the difference. More than 80 years later, the approach still stands. For example, after NLP model GPT-3 started making waves this summer, many were quick to give the experiment a try. Impressively, this model doesn't just seem to understand things, it has knowledge about the world, a reasonable amount of common sense and sufficient awareness of context to keep a conversation going. For example, this blog post by Kevin Lacker describes the following exchange, in which the model responds to questions asked by a human:

Q: What is your favorite animal?
A: My favorite animal is a dog.

Q: Why?
A: Because dogs are loyal and friendly.

Q: What are two reasons that a dog might be in a bad mood?
A: Two reasons that a dog might be in a bad mood are if it is hungry or if it is hot.

As impressive as these responses are, the same piece also finds a dozen ways to push the model towards failure:

Q: Which is heavier, a toaster or a pencil?
A: A pencil is heavier than a toaster.

Q: How many eyes does a blade of grass have?
A: A blade of grass has one eye.

Q: How many rainbows does it take to jump from Hawaii to seventeen?
A: It takes two rainbows to jump from Hawaii to seventeen.

In essence, the model breaks because it is asked to do something it was never trained for: indicating that it doesn't know or understand something. The discovery of these flaws is where AI loses its shine of intelligence -- it's no longer an actor, but rather a tool with some clever tricks. In most current day cases, the AI is bound to lose the imitation game when directly opposing a human, simply because it can't keep up with our trickery. In essence, humans are skilled at thinking outside of the box, but our models are either trained specifically to stay inside of it, or the computational cost exhibits them from going very far.

Moving Goalposts

When we aren't actively trying to outsmart AI models, it turns out that they do a much better job at convincing us: only last month, the same model described above managed to pass as a normal reddit user for more than a week. Furthermore, when our incentives align with the goal of a model (rather than just not being at odds with it), a new opportunity arises: the model can now prove itself to be directly valuable to us. This became very clear when GPT-3 was applied to generate code. Suddenly, thousands of people saw a future in which you can casually tell a machine what to do, rather than writing code. All of that convenience and productivity gain was imagined because of the tweet below:

The goalposts of machine intelligence are constantly moving forwards: what blows our mind today, might be considered somewhat neat in a couple of months, then utterly boring within years. It seems that any tool, no matter how advanced, is doomed to end up characterised by its limitations, rather than its possibilities. For example, think about the last time you used Google: did you wonder about the intelligence required to search the web and return exactly what you were looking for? Or did you just see it as a piece of code going through another loop?

Show and Tell

From my personal experience, the impression a machine learning system leaves on a person without a technical background boils down to two things: does it solve a problem that keeps them up at night? And importantly, if it does, is it a piece of the puzzle or rather the entire picture at once? In the latter case, you've got yourself a winner: the mere idea that a computer will take an actual problem off your mind entirely is enough to give it an aura of intelligence. While very appealing, this mystical quality has a downside as well: to preserve it entirely, model transparency has to stay on the sidelines, almost the same way a magician can't both keep their mystery and explain their tricks. Yet this very same transparency is crucial to allow a user to trust the machine they work with.

Finding the right balance between perceived intelligence and transparency is a crucial part of any AI-based product -- the former can often get you a foot in the door, while the latter will help you build a more sustainable user base. Over time, the excitement from both of these aspects will typically fade for any user. With daily exposure, even the most cutting-edge tool can grow boring. But if you're bored by a system rather than frustrated, it might just be solving your problem exactly as you expected, and isn't that what intelligent systems are all about?

Want to stay up to date?

Leave your email address below to receive my newsletter!