To borrow a phrase from Leon Trotsky, “you may not be interested in artificial intelligence, but artificial intelligence is interested in you.” It’s hard to avoid the subject of artificial intelligence, or AI, today. It’s all over the news, with bold predictions of how it will change everything. AI stocks are super hot, and China and the United States are in a chip race based on AI requirements. States, cities, and companies are building vast server farms to feed AI, spiking energy demand at exactly the same moment we’re supposed to be transitioning to electric vehicles! Much of what you hear is hype (no surprise), and you may be old enough (like me) to think, “well that will be somebody else’s problem with which to deal.” But probably not. However AI plays out, it will affect you within a few years, so it may be worth it to understand a little about AI now. Here goes:
I. AI is artificial, but it is not intelligence.
Some of the results AI programs can produce may look almost magical, but in the end, they result from a simple process. Everything called AI today is based on computer coding of large language models (LLMs). What is that? You already know that computers use “ones” and “zeros” (or digits, hence the digital world) to do everything they do. These large language models take words and turn them into tokens, or groups of ones/zeros. The base language doesn’t matter, which is why LLMs can work wonders with translations. What the program does is take the tokens (words in your view) and predict what the next token should be. The prediction uses a probability model (ok, no more math) that says, “based on what I have been trained on, what should the next word be?”
As a simple example, if you trained a LLM on the works of Shakespeare, then you asked it to describe “love,” the LLM would say things like, “love looks not with the eyes, but with the mind,” or “when love speaks, the voice of all the gods make heaven drowsy with the harmony.” Beautiful, no? Now if you added the lyrics from the J. Geils band to the model’s learning, it might say “love stinks.” Which answer it gives is based on how many times it sees that particular combination of words, and who says it, and how.
But wait a second, we’re falling into a trap by using phrases like “it says” or “it sees.” It does nothing of the sort. The model searches its repository of data and pulls out all the tokens which represent “love” then looks at the tokens surrounding that token, then puts them together to respond to you. There is no thought involved. But how does it get such amazing textual answers?
The key is feeding the model more and more data. Companies started using things like Wikipedia, but that isn’t enough. Now they’re scraping public social media, ingesting copyrighted material, anything to get more data. Because the more data that your AI model uses to train on, the better the results you get. One problem AI faces is recency bias. In the rush to add ever more sources, developers turn to today’s data. Almost no one is creating original data from the 19th century, while the Kardashians probably provide a megabyte a minute. AI will lean ever more heavily on recent data for learning, which negates one of the lessons of history: the things that are best are those which stand the test of time. AI does not care about “right” and “wrong,” just “fit.”
II. AI is a tool, nothing more, nothing less.
The use of tools is the story of humanity, from fire to the wheel to flight to computers. Tools are neither good nor evil; it is all about how you use them. A steak knife can cut a filet or a jugular vein. Alfred Nobel thought TNT would revolutionize mining and engineering. So there is no reason to fear AI, but there are many reasons to learn about it.
Like any tool, AI requires some initial effort on your part to work properly. If you don’t learn to use it through practice, you’ll find it too hard to use (the story of me and typing, as my two-finger approach can attest!). Also, AI benefits from you, as another source of learning. AI programs learn how you speak, how you think, what you expect from interactions with you, and then can respond better to you. Of course, that also opens the door to manipulation, too. Current AI developers are not consciously trying to program AI to be “addictive,” but that is a real possibility.
As a tool, AI most benefits those who can master it. Regular AI users point out that if you ask AI a stupid or poorly-worded question, it may give you a similar answer! Why? Not even the developers know for sure. There is no way to “open the hood” and look at what the model is doing; it is just “doing.” Another mind-blowing fact: AI developers have noticed the AI programs seem to work more slowly and are less productive during December. The developers think the AI has “internalized” the notion of a holiday season from its data, and answers accordingly.
III. AI can do many things for you, but maybe not better than you.
Since it is ultimately simply a computer program, AI is fantastic at taking over mundane or routine work. It never tires, never sleeps, never asks for a raise. Computer programmers love it, as “coding” often involves reams of simple instructions which are boring to write, but perfect to be automated. AI is great at giving you choices. Remember how great it was when you first googled “synonyms for” and found a ton of great alternatives? Now you can ask AI to take a paragraph you wrote and give you five different and better alternatives, and milliseconds later, you have them. AI assistants can “look over your shoulder” as you work on an email, for example, and tell you that your tone is harsh or you’ve mixed up the dates you’ve asked for the work to be completed.
What do you do better than AI? Probably the things at which you are best. The initial AI efforts wrote about as well as a sixth-grader. Current models are working at the undergraduate-level. AI will continue to advance, but it will never replace human expertise. For one thing, its data is based on human expertise, so it needs humans to continue pushing the limits of science, art, philosophy and the like in order to provide the “tokens” AI uses. So a very good lawyer may use AI to do research, or to hone a closing argument, but she will not be replaced by AI. Because AI also “hallucinates.” (Notice how we anthropomorphize AI? we have to talk about it a living, thinking “thing” just to make sense of it!).
There are great examples of this. Remember how AI “guesses” the next word? In most legal briefs, you’ll find the phrase “according to the case of X vs Y, . . .” AI “knows” it should use these tokens when you ask it a legal question, but if it can’t find the right citation, it will simply invent one! To AI, these are the right “tokens” to use, but to us, it’s just plain making things up. Since it has no free will, the AI developers call it hallucinating, since what AI does makes sense to it, but not in the real world.
AI can also be manipulated. Perhaps you heard about Google’s AI called Gemini, which started generating images of black Nazis, Asian Vikings, and Indian Romans. How did that happen. AI generated tokens to create images, but a developer put a filter in the program to make the images it created “more diverse.” So AI promptly ignored reality and made black Nazis. These problems are easy to correct, but demonstrate how AI can go wrong.
Positive counter examples exist in the world of medicine. AI can look at hundreds of millions of CAT scans and never tires, never has a headache or eye-strain. A good technician will still beat AI, but AI can serve as an initial screening tool or a post facto double-check. AI can also look at large bodies of medical data and find correlations which individual doctors might never see. Likewise, AI is looking at the development of thousands of ways to fold protein molecules into amino acids and thereby make treatments for various illnesses and conditions. Such work, even with modern supercomputers, would have taken decades.
IV. AI is coming, like it or not.
AI is part of the ongoing digital revolution, which means it happens at a pace with which we humans just aren’t prepared to deal. We have gone from the beta (testing) version of AI to level 4.0 in about ten years, and the trend is accelerating. Unlike the flying cars we were all promised in the 1960s, or the electric vehicle revolution which is always just around the corner, AI is already creeping into many things, whether you realize it or not.
Computers were supposed to replace mundane tasks, and they did put typists out of a job, but then coders became a thing. Now coders are in danger, and even so-called “white collar” workers are being reviewed. If you’re average or below average at your white-collar job (and half of people are, by definition), your boss will be considering whether an AI program could easily replace you. AI is seeping into service centers, the places where someone from India tried to convince you they were in Cincinnati and really wanted to help you. AI can make travel recommendations, edit papers, even “teach.” As it does so, there will be examples of really excellent AI efforts that are astoundingly successful, and others which are complete busts.
What cannot be denied is that the level of investment going on in the AI field, the development of the data sources and the research into better ways for AI to learn mean it will continue to affect you, personally, whether you realize it or not.
V. AI will exacerbate inequality.
As much as humans are a learning species, it is amazing that we always convince ourselves this next tool will be all to the good. Mark Zuckerberg thought Facebook would be a place for people to connect and become “friends.” This despite his original intent was to create a site for guys to rate girls on their “attractiveness.” Larry Page and Sergey Brin at Google worked under the motto, “Don’t be evil,” despite the fact illegal child porn was immediately one of the top search functions.
The internet writ large didn’t make humans smarter, despite making all the world’s knowledge instantly searchable. It did allow neo-Nazis, perverts, and fraudsters a chance to meet and grow. All that information devolves into cat videos, Tindr, and scams. AI will be no different. A good writer will perfect an even better paragraph or story. A scam artist, a better pitch, customized just to you. A good priest, a better sermon. A crank, a more inviting screed against whatever. For every medical breakthrough, there will be fake-news causing unnecessary death or illness.
The power of AI will enable people to do great and terrible things. Those who better understand AI will be better at using it, and far better at avoiding it using them.