First reads in AI: three short reviews
This article is originally published at https://robertgrantstats.wordpress.com
I had cause a few years ago to read up quickly on AI. I don’t mean the same as people who do a little data analysis and call it AI, I mean “strong AI”, the sort that is a serious attempt to build towards human-like levels of cognitive capacity. It really wasn’t something I’d thought about, so I got three entry-level books, and I’ll give a quick review here.
The pessimistic
Our Final Invention by James Barrat is a straight up pop science book by a journalist. It’s well written, the skill that only comes with years of practice, and although at first you might feel it is too drawn towards drama and doom, the scenarios of explosive learning and human extinction are reasoned and justified. The idea in the title is that ASI (artificial super-intelligence, which is more capable of thinking and learning than us) will be better at problem solving and inventions than us, so it can just take care of everything from now on. Hence, final invention. Or, it will categorise humans as an unacceptably unpredictable risk to its goals and will remove us from contention forthwith. Hence… you get the idea. It’s a fun read. ML/AI enthusiasts of course hate it. I gave it to a friend who had previously opted out of all tech since the telex, and I rather wish I hadn’t done that now.
The optimistic
The Master Algorithm by Pedro Domingos is not quite the counterpart to Barrat’s book. It’s about algorithms, stats and machine learning, not so much about strong program AI. But that makes it a rather more satisfying read long-term. Domingos makes one of the best efforts at explaining these various tools since Nate Silver. His writing is easy going and enthusiastic. Sure, as an expert, I felt there were some bits (my favourite bits!) omitted, but he could never have covered everything. How this book does relate to ASI is that he is very keen on the notion of unifying theories and meta-algorithms that learn how to learn and can draw on all the other little algorithms in the toolbox. Like Hegel, he has no doubt we will inevitably learn how to combine all our algorithms — the probabilistic, the heuristic and the racialistic — into one mash-up that will tackle all problems. Like Sauron, he has put all his chips on it. I don’t agree, but this is still a good read — if you don’t know about contemporary data science and want to know more than a business magazine will tell you, and less than Trevor Hastie will tell you. You won’t really learn about AI here. This little book went down the charity shop.
The learned
Throughout Barrat’s book, the disturbing scenarios generally have a footnote telling you that this or that idea comes from Superintelligence by Nic Bostrom. And that is the most serious and sober of the three books, though it still belongs to the pessimistic side of the road. Of course, ML/AI people hate it. It has the best cover, an owl in a (random?) forest made of the sort of alphanumeric strings that graphic designers take for code. Bostrom is pretty convinced that we’re doomed, although recent articles I’ve read give the impression he has lightened up. Everyone from Bill Gates to Nelson Mandela is quoted on the dust jacket. Alan Turing says “I wish I’d written this”. Ada Lovelace writes “I wish I’d read this”. But once you get past this publishers’ love-in, it is unique enough to have warranted some hype. Scenarios have clearly been played over and over in Bostrom’s mind over a long time. In the manner of Dungeons and Dragons, though, it’s hard to invent really comprehensive and realistic scenarios by yourself; you end up repeating your own tropes. In Bostrom’s case, these are von Neumann probes and nanotechnology. He has a dread fear of microscopic villains, possibly colonising space at the same time. In contrast to the other two books, this is tough going, not because it is more highbrow, as much as because it is not well written, or at least not well edited. He is sometimes given to languid, nay indolent, sentences and sonorous wordplay, which are pursuits I too enjoy, as you may have grasped already — not unlike a nettle — but I look about and find few fellow travellers, so, alas! reader, I am unable to count these as mitigating features. If you can’t be bothered, you can read Barrat and learn the same ideas without the detailed scenarios. This one I have kept, mostly because I want to see how it ages.
Thanks for visiting r-craft.org
This article is originally published at https://robertgrantstats.wordpress.com
Please visit source website for post related comments.