My wife’s interest in policy and broader societal implications of the development of AI led to her ordering one of, if not the foundational books on general AI: Superintelligence by Nick Bostrom.
My interest in machine learning and data science led to me to borrow said book almost from the moment it landed on our doorstep. It’s been an interesting read and has kept me occupied on trains and in spare moments. I wouldn’t recommend it for bedtime reading as it doesn’t shy away from being technical!
Below is a brief definition of superintelligence and a summary of a few topics offered at a very, very basic level.
What is superintelligence? Superintelligence (strong-AI or general intelligence), is a system that would be able to surpass human brain capacity in every field; imagine the smartest scientist ever born, an unbeatable trivia partner and the most engaging social butterfly all rolled into one.
Sounds complicated? It is. It’s like comparing the smarts of Einstein with those of your slow colleague at work and that’s not even close! Superintelligence would be like comparing Einstein’s cognitive capacity to that of ant, only humans are the ant and, because we’re the ant, there’s a lot of uncertainty on the topic. Articles written about superintelligence have many qualifiers (could's, would's, should's), this article included.
How do we make one? It is imagined that at some point in the future a seed AI will be able to rewrite it's codebase enabling it to recursively self-improve. This approach would allow it to become increasingly more intelligent, leading to an intelligence explosion. And then what?
How quickly the initial AI can improve itself governs our ability to develop and control the AI and our interaction with it. Fast self-improvement and the AI might
be able to out think us - giving it a strategic advantage - slow self-improvement and we might have time to perfect the AI and governance of its use.
I hope someone is monitoring our progress. Good news and bad news here. The short answer is yes, many are aware of the potential downsides. The most vocal lately has been Elon Musk, but many other prominent technologies have raised the alarm. There are now a few institutes looking into these issues, many affiliated with universities where this research is taking place. The bad news is that, unsurprisingly, governments are woefully behind.
Would AI ever be unfriendly? Bostrom warns that superintelligence might
not have anthropomorphic goals, citing an AI run paperclip factory that never knows when to stop or a machine that optimised human ‘happiness’ with matrix style electrodes implanted inside our brains. Brain electrodes, that’s something to avoid. Indeed.
You can’t program common sense. Programming typically has very literal, specific task-based code. This won’t be possible with an agile, self-taught AI. So we’ll have to do better and design ways to guide and control the AI in new ways to steer us away from brain electrode type outcomes.
I’m sure we’re a long way from anything yet. Well, almost sure. Most estimates put us at 20 - 50 years away from true superintelligence, and these are highly speculative timeframes. The truth is that we don’t know. However, applied AI, where we train AI on a specific task using a massive coded data set is very much here today and part of our daily life. Most of the hot new tech that offers responses to a range of user queries involves applied AI. Specific applications include speech recognition, language translation, search, image clustering, and even automated model creation (AutoML).
Phew! Overwhelmed? You should be. AI has the to potential to solve some of society’s most complex and existential challenges from climate change to food scarcity. At the same time, it has the potential to massively disrupt employment as we know it and lead to a world where decisions are made by AI, that we don’t understand, making it impossible to audit them and determine whether they adhere to our values.
Do say: data policy and general AI is something to take seriously and watch out for in the future.
Don’t say: Skynet is coming - run for your lives!