It can hardly have escaped your attention: big tech companies love big words. They either promise AI is the magic wand that will revolutionize everything. Or they spread apocalyptic messages on how AI is going to wreck our world if it gets the chance to free itself from our reign and lead a life of its own. In both cases, the big tech guys present themselves as the solution, so it's best to take both positions with a grain of salt.
The general opinion on AI is also more and more polarized; many companies still bet on AI to help them become more efficient, while the ones who deeply care about copyright, creativity, and the environment oftentimes refuse AI altogether.
This raises the question though, if a down-to-earth view on AI is even possible, and if yes, what it means in practice.
If you think this is an interesting topic, I can highly recommend the book 'AI Snake Oil'. It's all about what AI can actually deliver, while debunking hysterical opinions from Silicon Valley.
The authors, Arvind Narayanan and Sayash Kapoor, have been studying AI for a long time, so they clearly know what they are talking about. They offer a very nuanced view of AI, showing the real potential while staying away from generalizations and false promises. They aim to give foundational knowledge that will help you separate genuine advances from hype. And that's it.
First of all, they point out that even using the abbreviation AI is problematic, as it describes a whole area of very different technologies. They compare it to using the word vehicle for all forms of transportation we know: cars, buses, spacecraft, bikes. This does not help for clear conversations: 'there are furious debates about whether or not vehicles are environmentally friendly, even though no one realizes that one side of the debate is talking about bikes and the other side is talking about trucks'.
The book helps you to distinguish valuable types of AI, that deliver on their promises, from what they call snake oil.
They go on to explain the difficulties with the types of AI that attract such unprecedented investments nowadays, by giving insight in how these technologies work. Take chatbots. These are trained to learn statistical patterns from training data (mostly taken from the web), and then generate text based on those patterns. They warn that chatbots don't necessarily remember what's in their training data.
'When a chatbot begins to respond to you, it has no coherent picture of the overall response it's about to produce. It instead performs an absurdly large number of calculations to determine what the first word in the response should be. After it has output - say, a hundred words - it decides what word would make the most sense given your prompt together with the first hundred words that it has generated so far.'
This does not sound like something that is threatening human existence to me, rather like a very difficult way to reproduce human speech. This also explains why GenAI uses so much energy and water. Narayanan and Kapoor call it 'computationally expensive': 'to generate a single token - a part of a word - ChatGPT had to perform roughly a trillion arithmetic operations'.
Given the nature of a chatbot as a mere prediction of the next word, it becomes clear that it takes effort and practice to use chatbots in the right way. Inappropriate uses are much easier and faster, so that's what we usually end up with. Likewise, it will take work to use generative AI in the right way and appropriately integrate it into our lives.
Understanding what AI is and how it works really helps to better judge big claims from big tech, and it surely takes away a lot of the fear. Their down-to-earth approach shows when they discuss how AI is depicted in the media, where it is often presented as robots. But in fact, they say, 'most AI today is used to detect patterns in data. It's more like Microsoft Excel than the Terminator'.
Instead of discussing whether AI will end humanity or transform it, we prefer to use it for very practical purposes: unburdening sustainability teams who drown in work.