An event of truly historic dimensions offers the inspiration for this instalment of SnaS. The event already occurred last October, but was announced only yesterday. For the first time a computer programme, AlphaGo, created by a Google team, has been able to beat a human professional player in the game of Go. The news was announced on the Google Blog (https://googleblog.blogspot.de/…/alphago-machine-learning-g…), coinciding with the appearance of a paper in Nature (https://storage.googleapis.com/…/…/deepmind-mastering-go.pdf). Up until very recently, even experts had estimated that it would take at least another decade until computer systems would be advanced enough to challenge human dominance in this game. More conservative people had even thought that human go professionals would be unbeatable for computers for at least another century. But now this very significant event has happened and it sends out an unmistakable signal about the state of development of artificial intelligence. Because of the complexity not only of the electronic systems involved, but also of the methodical approach taken and the algorithms used, this is of greater significance than the victory of IBM’s DeepBlue over Gary Kasparov in 1997.
The game of go was invented in China at least 2500 years ago. In traditional Chinese, it is written 圍棋 (simplified Chinese 围棋), which reads wéiqí; in Japanese it is 囲碁 igo ‘encircling boardgame’. The modern Western name go is actually a misnomer because Jap. 碁 go simply means ‘boardgame’. When we turn to medieval Ireland, we find two boardgames regularly mentioned, brandub and, more prominently, fidchell (cf. Jan Niehues, ‘All the king’s men? On Celtic board-games and their identification’, Festschrift Erich Poppe, 2011, 45–60). The rules of both are notoriously unknown. Brandub means ‘raven-black’ and probably refers to the colour of the stones or figures used. In the case of fidchell it is known that two groups of men of different colour were used, like in chess or in go. For the only surviving example of a fidchell board, see the illustration in SnaS 14b. Fidchell (dil.ie/22014) has an exact correspondence in Welsh gwyddbwyll which allows to project it back to Insular Celtic or even Proto-Celtic *u̯idu-kʷei̯llā ‘wood-sense’. *kʷei̯llā clearly continues a pre-Celtic *kʷei̯s-leh2, a derivative with the instrumental suffix -lo/eh2- from the verbal root *√kʷei̯s- ‘to see’ (LIV2 381–2) that is found in OIr. ad·cí ‘to see’ and Gaul. pissíiumi ‘I will see’ and appisetu ‘let see’ < *kʷis-e/o-. PC *kʷei̯llā is directly continued in OIr. cíall ‘sense, intelligence, mind, reason, etc.’, W pwyll ‘prudence, wisdom, patience, understanding, intelligence, mind, reason, (common) sense, etc.’, and Bret. poell ‘lien (de ficelle, arrêt d’écheveau), liaison logique, bon sens’. Words for ‘seeing’ are often used for concepts of ‘understanding’ or ‘mind’, cf. only PIE *√u̯ei̯d- which in some languages serves as the basis for verbs for ‘to see’, in others for ‘to know’; or cf. Engl. insight or Germ. Einsicht.
The first element of the compound *u̯idu-kʷei̯llā is PC *u̯idu-, continued in OIr. fid ‘tree, wood, timber’ (dil.ie/21999), but also ‘letter’, OW guid, W gwŷdd ‘trees, forest, timber’, OCorn. guiden ‘tree’, MCorn. gueyth ‘trees’, OBret. guid, Bret. gwez ‘trees’, but it has also a cognate in Engl. wood. eDIL (https://www.facebook.com/eDIL-1559061804342736/?fref=nf) informed us yesterday that in compounds, fid- indicates that things are made of ‘wood’. While this is correct in many instances, it doesn’t work for fidchell, but in this and similar cases it must refer to something ‘that has to do something with trees or wood’. In the case of fidchell we can only guess that the original meaning was ‘(a game/an application of the) mind that has to do with (a gameboard/gaming figures made of) wood’. In fidḟogal ‘plunder of timber or woods’ or fidbreth ‘judgement on woods’, fid furnishes the object of the verbal action expressed by the second element. If understood that way, fidchell would be ‘(the application of the) mind on (objects made of) wood’.
So, why is this event, the first victory of a computer over a human pro player in go, so significant? Unlike previous attempts at the problem, the first part of the used algorithm does not simply rely on the enormous computing power of the machine to mindlessly calculate all possible positions and variations (‘brute force’). While this is possible in chess, the unimaginably gigantic number of possible positions in go cannot be computed due to limitations of capacity. Instead, what the Google team did is that they created a kind of artificial neural network that, as far as I understand it, is able to learn on its own, improve itself by learning and make judgements about reasonable and unreasonable approaches. With this, the method comes closer to how the scientific mind operates. Only at a late stage of the problem solving process do previously used strategies of so-called Monte-Carlo tree searches come into play where probabilities are being assessed by creating statistics of a large number of random moves. I estimate that this achievement is on a par with the landing on the moon.
When we make judgements about problems, we make educated guesses informed by our professional expertise, and taking account of scores of accompanying factors, such as ‘world knowledge’, that we may not even be conscious of. This is the human ‘gut feeling’ or ‘instinct’; the quality of a scholar fundamentally is a reflection of how accurate this instinct is. If I understand it correctly, AlphaGo is kind of emulating this part of the human gut feeling, but with the added advantage that it is strictly quantifiable. Going further in this direction, I can see that programmes operating on neural networks may in the future create things like etymologies (which require not only a good formal knowledge of sound laws, but also an enormous amount of world knowledge). And obviously, this approach is also highly significant for one of the aims of ChronHib, the quantification and estimation of dates of texts.
PS: Artificial Intelligence can then be regarded as being on a level with human intelligence when a computer succeeds in parsing correctly any given Old Irish verbal form found in any manuscript or state of transmission.