Human Compatible

Human Compatible : Artificial Intelligence and the Problem of Control (2019) by Stuart Russell is a look at the threat posed by Artificial Intelligence (AI) and how it could be ameliorated. Unlike many other books on the threat of AI Russell is truly and expert. The AI textbook he wrote with Peter Norvig is probably the most used book on the field.

In Human Compatible Russell first goes into detail about the potential scenarios in which AI causes disasters. These are excellently expounded and are genuinely worrying. AI used to control fleets of explosive drones to perform mass murder or used by authoritarian regimes to gain more control are even worrying without ponder Artificial General Intelligence (AGI) and an Intelligence Explosion.

Russell is even good enough to mention those true AI experts, notable Andrew Ng and Yann LeCun who are either circumspect or sceptical about AGI. He even has Andrew Ng’s quote that ‘Worrying about AI evil superintelligence today is like worrying about overpopulation on the planet Mars.”

In roughly the second half of the book Russell goes into how carefully teaching well grounded moral philosophy to AI’s can make the threat of AIs far less. This part of the book is interesting and Russell has both read quite a lot of serious philosophy and is using it to think about the problems that AI may pose.

This book, along with ‘You Look Like a Thing and I Love You’ are both excellent books to read and think about AI with. Russell is truly a very concerned expert and writes very well about the threat that he believes AI may well pose to humanity in the next decades. He also writes lucidly about how it can be dealt with.

A World Without Work

A World Without Work (2020) by Daniel Susskind looks at how the author things that with advances in AI work will become less available for many people. Susskind looks are fairly well explored territory. Quite a number of other people seem to believe that the job apocalypse is about to arrive.

The book does make some good points. Susskind nicely says that highly advanced narrow AI, as opposed to general AI, will start to take jobs. Susskind is impressed by AI’s ability to play Go better than any human. He does, nicely take into account the way that IBM’s Watson, after winning Jeopardy, has failed to actually generate much income for the company. This doesn’t rattle his faith in AI though. It’s not clear if Susskind has ever had a go at trying AI himself. He would be well advised to read some books that are a bit skeptical of AI such as ‘You look like a Thing and I love you’.

Susskind does go through false alarms about the end of jobs that have occurred previously and he does not that we have gone from a world where 95% of people worked in agriculture to a world where ~2% do and from where 30% of the workforce worked in manufacturing to a world where less than 10% do. But Susskind believes that because AI will take over non-routine work this time is different. He doesn’t make the case well enough that it actually will be though. Critically Susskind barely seems to look at the many jobs that are not just cognitive. It’s hard to work out how many people work at a desk, but according to the US Bureau of Labor Statistics only 40% of people work sitting down. This indicates a lot of people are moving around and interacting a lot with the environment. For a book written in 2020 Susskind also doesn’t look enough at the relative failure of self-driving vehicles. They will, probably come in the 2020s, but even that is far from certain. And really, driving is a task that many people accomplish along with other tasks they perform. It is not as varied as a job like, say, plumbing, home repair or gardening. If jobs like these still exist it’s hard to believe that work is going to rapidly decline over the next two decades or so at least.

Susskind does, nicely consider the statistics that despite a lot of AI being around unemployment hasn’t increased. Instead he points to participation rates that have slightly decreased and have in particular decreased for men and particularly young men who now have the internet and amazing video games to provide a low cost alternative to the often less emotionally rewarding world of work.

Susskind then goes on to suggest that the answer to these issues is unlikely to be education because we’ve probably educated ourselves to the point of diminishing returns and that sending everyone to university probably isn’t worth it. So instead Susskind suggests that Big Government is needed to redistribute even more wealth. Susskind doesn’t pause to ponder how the redistributive state is already the biggest that it’s been in history. He then goes on to suggest that big tech needs to be reined in by committees of the worthy. He does, nicely point out that in many ways their economic power isn’t that great and also points out how Microsoft were seen as the great computing monopoly of the 1990s but who are now seen as an underdog against Google and that their domination of desktop operating systems is no longer seen as a great threat. Instead he believes that the political power of big tech through social media and search needs to be reined in. Here Susskind is not on strong ground and it’s also worth noting that this strays from his book about the future of work to being ‘why government should direct big tech’.

A World Without Work looks at a genuinely interesting question, that of whether AI is going to take all our jobs. Susskind doesn’t make the case that it really will strongly enough, but he does make a worthwhile effort in looking at important issue. He’s clearly a very smart guy and writes well and has read widely about the topic. However it’s hard to imagine many people who were not convinced prior to reading this book would be convinced by his case. Still, the book is worth reading for anyone interested in the topic.

The AI Delusion

The AI Delusion (2018) by Gary Smith is an attempt to show how overblown the hype about Artificial Intelligence is. In part it succeeds with some interesting tales of how various AI systems have failed and also by pointing out flaws in just picking correlations. But it also fails by seriously considering the way modern Machine Learning (ML) systems have solved problems that people thought intractable 20 years ago and to consider if there is more progress where it might lead. It is something of an antidote to people who think ML is about to put everyone out of work and lead to the Singularity but it goes far to far the other way.

There are some fascinating stories in the book. In 2008 and 2012 Barack Obama’s data analytics team was regarded as great and crucial in helping him win. Few people know that in 2016 Hilary Clinton had an ML system that instructed her that it wasn’t worth campaigning in Michigan and a number of states where Obama had done well but she’d done poorly against Bernie Sanders. The book alleges that Bill Clinton thought she was in trouble in these states and was furious but was over ruled by the computer.

The book has other stories like that about trading and other things.  But also many that are repeated in other books like the one about where to armor aircraft in WWII.

Smith puts forward the case about how ML can learn specific things but can be easily fooled. He also spends far too many chapters on ML that is over fitted. Introduction to ML courses teach students the dangers and how to avoid that.

The chapters on finance go to far. Correlation and chartism are well known. I’ve also heard quants  on podcasts calmly say quite happily that ML doesn’t really work on markets because of false positives. However, something quants do does work. Human traders are being replaced, but no by ML.

The AI Delusion does make a few good points and will have some good stories for most people. It also contains a lot on how statistics can lead people to make mistakes. Smith’s recommendation that we don’t defer to ML is also pretty wise. However, the case it makes is overdone and Smith spends too much time attacking a straw man of what. Whether ML will lead to job losses and the Singularity is highly questionable, but there is more to this than Smith lets on.

 

 

 

Deep Thinking

Deep Thinking: Where Machine Intelligence Ends and Human Creativity Begins (2017) by Garry Kasparov and Mig Greengard is a book that looks at how machines eclipsed people in playing chess and what this means for humanity.

Kasparov is one of the greatest chess grand masters of all time and the last human to be the best chess player on the planet. In 1997 Deep Blue defeated him taking the crown for an activity that was once seen as the epitome of human intelligence.

The book looks at how computers play chess, how they were initially fairly week and how in the late 1980s they began to become as good as the best human chess players and finally beat them.

The story of the actual game shows that Kasparov believes, it appears with good reason, that his loss to IBM’s Deep Blue involved quite a bit of unfair play. The machine was allowed access to his back catalogue and was quite possibly altered during the game. However, he also makes it clear that if he had won that game he would have lost within a few years. Computers had become too good.

Kasparov goes on to reflect on how this has changed chess, how modern grand masters use computers very differently and how the combination of humans and chess computers is, for the moment, better than just computers on their own. He also reflects on how this doesn’t mean that the singularity is near and the book has a great quote from Andrew Ng, the Machine Learning expert from Stanford, Baidu and Google who says that worrying about super-intelligent computers is like worrying about overcrowding on Mars.

The book is very interesting in parts but also quite dull in parts. You’d really have to be a great chess aficionado and someone who is interested in what a smart, informed person has to say about AI to really appreciate it all. And even then you’d probably find it sags in the middle. However it certainly contains insights from someone with a unique perspective on chess and AI.

The Most Human Human

The Most Human Human (2012) by Brian Christian looks at how AI can make us reflect on what our human qualities are and how AI relates to them. Christian was one of the human participants in the 2009 Turing Test, where computers attempt to pass as humans as both chat to judges over written messages.

Christian has degrees in Computer Science and Philosophy and so is ideally placed to write about the subject of humanity and AI. The book is also well crafted. The story of Christian being a confederate in the Turing Test and his discussion of AI and humanity is interleaved skillfully.

Comparing chatbots and how people also follow scripts as well as how computers have improved at chess and other fields and pondering what makes humans different is definitely an interesting topic and the book does this all well.For anyone interested and how humanity is different and similar it’s definitely worth a read.