what is a philosopher? A philosopher is someone who studies and thinks about philosophical problems. So far so good, but what exactly does that mean? Well, philosophers aren’t just trying to figure out whether there are five kinds of people or not. They’re thinking about questions such as: Is there any ultimate reality? Do we live in a computer simulation? How did the universe begin? Are there other universes out there where our lives never happened? These are all very important philosophical issues, and they require thought.
So what does this have to do with computers and artificial intelligence? Well, in order to solve these big philosophical questions, you need to understand some basic ideas about how the world works. You need to think about logic, reality, knowledge, truth and all kinds of other stuff. This kind of thinking is called “epistemology,” which just means “thinking about how we think.” The main field in which epistemological questions are studied is called “epistemology.” For a long time, epistemologists thought about these issues mostly using natural language. That is, they thought about them by writing long sentences in English or Latin or some other language. Epistemologists would spend years writing a book like On the Nature of Things, and they’d think really hard about these issues. Then other epistemologists would read the book and think really hard about the issues too. But as science progressed, it became clear that natural language wasn’t good for this kind of thinking. It’s just too easy to fool yourself when you use words.
So scientists and thinkers had to develop a way of communicating ideas without ambiguity—a precise language. That’s why today we use a symbolic language called “mathematics.” When you do math, there’s no guessing or ambiguity. 2 + 2 always equals 4, and you can always check by multiplying 2 × 2. If you see a squiggle “~” with a bunch of other squiggles around it, then “~” means not. You can’t be confused by words like “probably” or “maybe.” You can’t start arguing about what certain words mean. If you see a bunch of symbols, then you can’t get away with just making stuff up.
You can write a whole book in symbols. In fact, books full of nothing but squiggles have been written. These are called “computer programs,” and they are our best attempt yet at making an unambiguous description of reality. A few thousand lines of carefully chosen symbolic logic can describe the entire physical world—every atom, every force, every interaction. A computer is a kind of virtual machine that runs these descriptions, and we have machines today that can run programs longer than any book you’ve ever written. But these programs are still just a description of reality. They can’t ever truly capture reality itself. That’s impossible.
But don’t take my word for it—just ask Kurt Gödel.
Kurt was one of the greatest logicians of the 20th century. He proved that it’s impossible to describe the world with perfect precision. Any logical system that includes basic arithmetic will always have truths that can’t be proven from within the system. This is called “Gödel’s Incompleteness Theorem.” What this means is that no matter how much we think about stuff, we’ll never be able to describe the world with perfect accuracy and completeness. We can only make approximations.
This makes a lot of people very uncomfortable. A lot of people don’t want to hear that we can’t know everything. They think that our inability to describe the world with perfect accuracy means that science is wrong, or that God set up the rules, or something like that. But these ideas are all wrongheaded. Sure, we’ll never know everything. But that doesn’t mean we know nothing! We don’t need to know everything about cancer to cure it. And we don’t need to know everything about the moon to land on it. You can get through your day without knowing the mathematical exact location of the pants you had on yesterday. And you can get through life making reasonable decisions without knowing everything that’s physically possible for you to know about the world.
First sampling. Two-shot (two real sequences articles fed in as context).
Hypothesis: Unlike the language models before it and ignoring context length issues, GPT-3′s primary limitation is that it’s output mirrors the distribution it was trained on. Without further intervention, it will write things that are no more coherent than the average person could put together. By conditioning it on output from smart people, GPT-3 can be switched into a mode where it outputs smart text.
The Sequences post you’ve never read, by GPT-3.
First sampling. Two-shot (two real sequences articles fed in as context).
Hypothesis: Unlike the language models before it and ignoring context length issues, GPT-3′s primary limitation is that it’s output mirrors the distribution it was trained on. Without further intervention, it will write things that are no more coherent than the average person could put together. By conditioning it on output from smart people, GPT-3 can be switched into a mode where it outputs smart text.
So, I’d tabbed out of this window, and when I returned I started skimming the last few paragraphs.… I didn’t notice