Wednesday, January 08, 2020

Human Compatible by Stuart Russell - go to guy on AI - a must read..

Unlike Bostrum’s Superintelligence, this book is eminently readable. Russell is not only an AI guy, he is a renowned educator and has won awards as an educator. Along with Peter Norvik, he wrote the classic, globally recognised textbook AI Artificial Intelligence: A Modern Approach and has been in AI long before most of the arrivistes currently bashing out books on AI and ethics.
He anchors his ethical arguments on the problem of ‘control’, machines that achieve our not their objectives. Unlike many of the popular books on ethics in AI, from contrarians like Cathy O’Neil, Safiya Noble and Shoshana Zuboff, who tend to throw the babies and bathwater out with the bath, he is level-headed and draws upon his decades of experience as a pioneer in the field to explain the flaws, limitations and dangers, as well as the power, promise and potential of AI.
He takes on his opposition directly and unlike Kevin Kelly, Andrew Ng and Stephen Pinker, he thinks that Superintelligence is a possibility. This is a brilliant summary of the main positions with Musk, Hawking and Gates at one end of a spectrum and Pinker at the other. Russell is in the middle, scared but offering a change of direction and solutions.
AI and learning
He points towards significant advances through AI in health, finance and education. The Global Learning XPRIZE is cited (you never hear about this but it’s important) and gives us 80 years before human-level AI arrives. Note that super-competent teaching AI may arrive decades before this. Understanding language is essential as is understanding and building hierarchies of abstraction. He scotches the idea of robots carrying around their own brains (therefore robot teachers) in favour of connected computing, EaaS, Everything as a Service. The low hanging fruit may well be in accelerating learning in humans. AI tutors, he thinks, will deliver learning to all, at negligible cost.
Unlike most treatments of ‘bias in AI’ Russell remains cool and objective in surfacing the problems but also patiently explaining the solutions. He avoids the usual teeth gnashing about capitalism, ethicswashing and ethicsbashing, for a typically balanced analysis of the problems but also the solutions – pre-and post processing, embedding fairness criteria and so on. 
Beneficial AI
Provably beneficial AI systems will need clear principles around altruism, humility and reference to human preferences. Beneficial AI is AI where we know what the purpose is. This is not as easy as it sounds and he uses AlphaGo to show how tricky it is -  is the goal 'winning' or 'reinforcement learning where winning is the reward'?
Preference Utilitarianism
The complication in all this is we humans. His take on moral philosophy is utilitarian, or rather Harsanyi’s preference utilitarianism where AIs only purpose is to manifest human preferences. His defence of this position is impressive and if you want to have a position on AI and ethics this is well worth reading as he rightly takes ethics (moral philosophy) seriously – many don’t which leads to fantasy castles of certainty built on quicksand.
Everyone and their uncle
The old model of technology that meets an objective is no longer valid. We need technology that defers to humans and can be switched off. Rather critical that ‘everyone and their uncle’ is setting up an AI and ethics board, department, framework etc., he proposes we give the top institutions a chance here – AAAI, IEEE, Partnership on AI and EU. It is not that others do not matter, only that there is too much low-level noise.
The strength of the book is that it is neither dystopian nor utopian. His line is realism, both about the dangers and opportunities. This is a book of some depth as he faces up to the possibility of an existential threat not with platitudes, anthropomorphism or manifestos butvreal solution in terms of moral philosophy and practical steps.
Of all the books I’ve read on AI and ethics, this, along with pieces from Dennett and Pinker have been the best. These three have truly shaped my own opinions on ethics in AI. I lean perhaps more to Dennett and Pinker in thinking the problem of superintelligence is not as likely, extreme or unmanageable as many suggest but this book is as good as you get on facing up to the perils, even if they are improbable. AI poses risks he but remains optimistic that we can retain control and harness the technology for good. 

No comments: