The concept of general intelligence does not always gain general acceptance. It seems too general, and thus unable to explain the myriad sparkles of individual minds. Multiple intelligence, some people aver, is a better thing to have: a disparate tool set, not merely a single tool which has to be deployed whatever the circumstances.
Not so. The great utility of general intelligence is its generality. It can solve, or at least partially solve, the very large range of existential problems which beset our ancestors. It can help solve the current problems which beset us now, none of which were pressing on us in that particular form over the many generations in which our brains developed.
Indeed, having specialized skills would be a risky strategy for any life form, because whatever triumph that specialist mind conferred in a very specific niche, it would be hopeless if the niche disappeared, and it had to survive in an unfamiliar environment. For many specialized brains that would be a death sentence. Seen from an evolutionary perspective, a case can be made for prioritizing general problem-solving ability above all else. Specialized skills are limiting: they are too refined for the rough and tumble of ordinary existence. Better a Jeep (General Purpose vehicle) than a low-slung racing car if you want to travel across the rough roads of the world. The latest developments in artificial intelligence are based on making that intelligence general, not specific.
I have talked about artificial intelligence before, in the distant old age of 2016. That was when the best game we had in town was Alpha Go. How ancient that seems now. It was programmed to do things, using the game-winning strategies developed by the programmers.
Now we have Alpha Zero. It has been given an improved but still very simple brain, a few dozen layers deep, rather than the mere 3 layers of former years. Of course, there are not actual neurones or axons. These are concepts which serve to organize the way the programs run, and this is done on whole sets of servers, the way that most complex big-data problems are handled. It is this form of quasi-neural organisation which allows deep learning networks to operate. I see them as correlation accumulators, being conditioned by the reward of winning into deriving the strategies which promote winning, and learning their craft by perpetual competition. Call it speeded intellectual evolution: thousands and thousands of games being won or lost (generations flourishing or perishing) which lead to a well-conditioned, super-smart survivor, ready to take on the world.
These changes in the depth of learning make a big difference. Just give Alpha Zero the rules of a game, and it dominates that game, even though (or perhaps because) it has zero domain knowledge about that game. It has been stripped of human wisdom. It is an ignorant but fast-learning student. And it dominates all games, once it has been given the rules. Zero knowledge, but an ability to learn. Finally, a blank slate.
What will all this mean for us? Some citizens are waiting for the Singularity, also known as the Second Coming. Forget it, says Hassabis. Artificial intelligence will be a tool we use: fine in some settings, not so good in others. For example, artificial intelligence is very good at looking at retinal scans and detecting anomalies which require further investigation. The artificial intelligence programs of old would have flagged up the particular scan for further investigation, and left it at that. Now the program flags up the scan, and also identifies the features which led to it being selected for investigation. The best human expert can look at the suggestion, and decide whether the artificial intelligence program has got it right. The expert now has an even better tool than before.
Can artificial intelligence be mis-used? Yes. So can a filing cabinet. Do you remember those? Code breakers with filing cabinets helped win the Second World War for the Allies. That code-breaking required very crude artificial intelligence tasked with doing just one job: calculating whether an encoded message could have been created by a particular rotor setting of an enemy Enigma machine. By rejecting, say, 240 impossible settings a few possible ones could be studied in more detail. After breaking the code then real intelligence took over, and the knowledge stored in filing cabinets made sense of the messages.
Demis was interviewed by Jim Al-Khalili on The Life Scientific, a BBC radio program.
I hope this link works for you, though it may be UK only.