The social currency of any self-respecting futurist-thought-leader type is to marvel publicly at how we live in an age where robots can beat us at Go, Chess and DoTA. “We live in a special time”, he repeats, for the documentary-makers in the crowd. Surely if a robot is better than the human at these, we must be close to the holy grail : artificial general intelligence.
Of course, it is also true that window cleaning robots clean windows better than humans and Roombas are absolutely superhuman at vacuuming the floor. If only the Roomba inventor had a better PR team.
But no, he says, it just doesn’t play complex-sounding board games – it can also recognize dogs & cats, which is important, translate from Russian to English, and caption photos of things it has seen before. It’s the same neural network everywhere! Doing all these different things!
Of course if you did create a baby with a neural network inside its head, and it became an expert at recognising cats from dogs, you would have to kill it and give birth to it again to teach it to translate Russian. The thought leader calls it “retraining with different hyper-parameters”, but we might see how a better name is “baby-killing“. Mind you, not just killing babies, but also changing their sense organs to make sense of different inputs and outputs every time – like that evil kid “Sid” from Toy Story with the Frankenstein toys in his house. Given the lack of mass baby-killings it is unlikely a baby learns by this method of force feeding the same image of a cat/dog repeatedly down it’s retina (although the education system wouldn’t mind that). Do you really need to kill the baby every time? If the paper submission deadline is near, why not.
But we could scale it up with more compute, and more data, the thought leader further opines. After all, it took 45000 years of training on millions of dollars of compute to beat the best human at DoTA!
Empirically, “victory” is an arbitrarily chosen metric that happens to tilt towards one side. It took the human David just a bowl of ramen, a cup of coffee and a few years of playing it daily to wake up and be able to evenly compete with this 45000-year superhuman Goliath. It’s a wonder he isn’t giving TED talks (yet). Should we be aiming for more compute and data or less of it? The victors write their own history.
Then the final bogey comes : the bigger quest is to make “ethical, safe and moral” AI – the thought leader gesticulates. We must transfer all these “human” qualities to our cat-dog categorisers before it is too late. The assumption being that human beings somehow possess overflowing portions of ethics, safety and morality, in a way that tigers, polar bears and other degenerates do not. It is a way for us to feel better – this robot is better than me at everything, but surely I’m winning at morality & ethics (both invented by serious-looking men without any vocation, to endlessly pontificate over trolley problems)
The hands are thrown up – the thought leader has run out of parlour tricks. Well, atleast they’re trying! – he whines. Anyone trying to change the world will get hate. If we can solve this, we can solve famine, and drought, cancer and the big whopper – global warming.
These were all problems earlier delegated to a certain bearded Man in the clouds. One word that doesn’t come up a lot is how AGI pursuit will be immediately “useful”. How dare one make such utilitarian claims of something that holds such obvious humanitarian benefit? Just like the bearded Man, this preordained AGI providence is essentially useless until it is useful. Things that aren’t useful in the short term either die a slow painful death in the long term, or live forever as gods.
String theorists know this very well. There is no way to experimentally verify the theory’s claims. No practical applications. Which means there will never be a polite way to declare it hogwash. And so on and so forth the deliberate cycle of fund granting and thesis writing continues. Similar is deep reinforcement learning, the current shiny thing, which is without any practical use. Although that is not completely true – it makes for great headlines.
The reality of most well-funded AGI pursuits is bleak. Real environments are made artificially simpler for dumb algorithms. Horizontal ‘generalised’ mastery across tasks is sacrificed at the altar of superhuman vertical machoism that generates press. Billions of dollars of talent and compute are spent on methods completely insulated from the risks of being wrong, which any plumber/truck-driver/handyman will tell you is called bureaucracy.
Deep learning is brittle yet useful, although efforts to generalise it are not. Learning is reasoning from cause to effect. As long as you’re computationally brute-forcing cause backwards from effect, which is what deep learning does, it is hard to make an intelligence that is ‘general’ and can learn on it’s own – because you cannot program all possible effect-cause pairs into it. Hence you have to retrain it from scratch to learn new things, instead of relying on it knowing a few primitive causes to figure out all other effects from.
There is no easy solution to this, and hence it is worth working on. But not the way the thought-leaders would want it. Rather than idolising any one technology, perhaps it is better to create systems where usefulness and generalisation is explicitly incentivised by the market to improve over time.
“I know that most men can very seldom discern even the simplest and most obvious truth if it be such as to oblige them to admit the falsity of conclusions they have formed, perhaps with much difficulty—conclusions of which they are proud, which they have taught to others, and on which they have built their lives.”
Tolstoy is right, and Russian. Most thought leaders will never abandon their ways. But they can surely condemn baby-killing. It is unethical, amoral, and not very safe.