There has been a lot of recent excitement about reinforcement learning, led by DeepMind’s news about AlphaGo zero being able to achieve the best Go-playing rating by learning from scratch. Their system was able to rapidly learn by itself, with no prior training and preparation for Go. This impressive achievement brought forward considerable buzz about all kinds of other possibilities of machines bootstrapping themselves to achieve superior results in other fields.
However it is rather quickly becoming clear that such bootstrapping efforts do not translate to other much more complex problems with unclear rules, constraints and ambiguities (see e.g. why self-taught artificial intelligence has trouble with the real world) . In that respect, games are an almost perfect environment where rules are very clear, even if machines need to figure them out by themselves. In the real world, such situations almost never exist and the levels of complexity and uncertainty are immensely higher.
But there is another very important point that rules are not known in many of the most important human activities. And it is not just that ordinary people may not be aware of some great insights of geniuses; instead even the geniuses themselves are also shrouded in uncertainty and mystery. Their genius is not always in absolute clarity of their great results but in ability to create great things in spite of such obstacles.
Economics is a great example of such a field of tremendous importance (politics is another). It would be difficult to say there has ever been a person on this planet, regardless of their economic genius, who understood with any great degree of clarity how economies really work and what the key concepts of money, capital and wealth are really about. Instead, great economists created grand theories with different sets of rules, that attempted to explain certain economic and financial concepts and behaviors. All of them had only mixed success for there is no, even to this day, an economic theory universally considered to explain a significant part of economics.
Economics is still so unclear and uncertain that it can be viewed as a game, of finance, where rules are incomplete and mysterious. Bootstrapping such a game seems futile and doomed to fail as it keeps changing and is never clearly spelled out.
But all is not lost, and this article is not only about criticism. Regardless of uncertainties, there are well-known examples of top human performers in economics, such as successful traders, hedge-fund managers, financiers, industrialists, tycoons and others who are still able to produce great results. And it is clear that they do not operate in vacuum, devoid of any rules, observations and insights. On the opposite, their success comes from their own views and understandings of economy and finance, encapsulated in their proprietary rules, correlations and systems.
And it is laso not that there are no known rules, laws, observations and insights in economics. In fact there are many of them in public. The difficulty comes from figuring which ones to apply and when. So in that respect, rules of the game of economics are out there for all to see. Trying to bootstrap them from nothing would be pointless, not to mention impossible. Is a supposed AlphaFinance Zero system to learn all of what we have known in economy and finance from nothing? And how could it even begin to try that?
The answer is that it could not and it is pointless to even try. Instead, superior AGI will leverage all of what we know about very complex fields and topics, and its great intelligence will come reasoning and choosing what to apply, where and when.
It is at this point that we can start seeing that all of these reasoning rules, across all categories and languages of human knowledge are what Reasoning Graph is all about. Its immense scale is what will enable truly great AGI, not tryingg to bootstrap it from nothing.