Richard Sutton is one of the foremost authorities in Artificial Intelligence. His dedicated Wikipedia page credits him as one of the founders of modern computational reinforcement learning, which refers to fully adaptive AI systems that can improve through trial-and-error interaction. In 2019, Sutton wrote a short post on his ‘incomplete ideas’ blog, pondering the ‘bitter lesson’ of more than 70 years of AI research: large scale computation is more effective than leveraging human knowledge. He was referring to two distinct paradigms of artificial intelligence: the more conservative and ‘tame’ approach that relies on knowledge representations – human expertise and wisdom are modelled and built into agents – and the more aggressive (‘brute force’) computational approach that relies on scaling through computation and big data. Sutton captures something quite important in this dynamic: the successes of AI are often tinged with bitterness, because they systematically confirm that a less refined and blunt approach is more effective than a human-centric one. This bitterness was already apparent in one of the earliest and most well-documented battlegrounds of AI research: computer chess. Here is how Sutton recounts it:
The methods that defeated the world champion, Kasparov, in 1997, were based on massive, deep search. At the time, this was looked upon with dismay by the majority of computer-chess researchers who had pursued methods that leveraged human understanding of the special structure of chess. When a simpler, search-based approach with special hardware and software proved vastly more effective, these human-knowledge-based chess researchers were not good losers. They said that `brute force’ search may have won this time, but it was not a general strategy, and anyway it was not how people played chess. These researchers wanted methods based on human input to win and were disappointed when they did not.
However, the trajectory in AI research over the past 20 years – and its growing relevance in the political economy of platformisation – proved the champions of human knowledge wrong time and time again. In domain after domain human-centric models that tried to valorise human expertise and values were outperformed by opaque statistical approaches that could quickly mobilise parallel computational muscle (multiple processing units at the same time) and the seemingly unlimited data pool of the internet. The intoxicating promise of statistical learning was to do away with labour-intensive human expertise and to reframe knowledge as a ‘discovery’ – the fortuitous conclusion of automated journeys into the wilderness of relational data, guided by the notion that proceeding inductively (and indeed automatically) from proxies and left-behind traces offers a more efficient and performant route to ‘truth’. This promise encapsulates the immediate horizon of the platform logic in society and, it follows, in education. Indeed, we are at a point where large language models built on the premise outlined above (statistical learning benefiting from internet-extracted trace data) have reached unprecedented levels of performance, enabling AI to replicate human communicative competence and creativity without any scientific or philosophical insight into the actual dynamics of ‘minds in culture’. This may well be a point of no return – statistical learning is here to stay in one form or another – and the challenge at this juncture is no longer one of tweaking these systems to make them more ethical, but a more pragmatic one concerned with their regulation and governance. This is what Frank Pasquale called the ‘second wave’ of algorithmic accountability, which seeks to move on from the technical goal of algorithmic improvement to ask more political questions concerned with the responsible integration of these systems in society – one of such questions is whether some of these algorithms ‘should be used at all and – if so – who gets to govern them’. The answer won’t be easy to find, but it will have to take into account the risks of our democratic societies developing an overreliance on proprietary AI ecosystems which are:
- driven by monopolistic practices;
- led by unpredictable and volatile owners;
- prone to ‘enshittification’;
- deeply embedded in a business model that demands the extraction of rents from end-users.
automatED 
Leave a comment