Do Chatbots Think? Yes. They “Think” Like Mushrooms
Okay, mushrooms don’t think. But then they don't need to. Neither does the chatbotA nearly 15-minute YouTube video offers insights into how a chatbot’s development mimics that of a mushroom, except for this: The bot works with information instead of soil.
This explanation could be helpful to knowm especially if you run into someone who has spent hours online and has come to believe that he is talking to a person:
From the Introduction:
When you press send, ChatGPT does not look anything up. There is no library inside it, no list of facts and no saved copy of the internet. Imagine instead a giant web of mushroom roots under a forest floor, where every thread grew thicker each time two words appeared together. Your question lands on it like a spore, and the answer grows from there one step at a time.
Here are a couple of excerpts from the transcript
Chapter 2: The web of mushroom roots
[0:21seconds] To understand that, you only need one picture in your head. Under a forest floor, mushrooms grow a huge web of roots. Thin white threads spread out in every direction, connecting one spot to another. Some of those threads are thin. Some are thick, because more has travelled along them.
Scientists call this web of roots mycelium. An AI like ChatGPT works a lot like that web. So for the rest of this video, imagine the AI is one giant web of mushroom roots. Before that web can answer anything, its roots have to grow.
Chapter 3: How the roots grow
[0:55 min] That growing stage is called training. During training, the AI is fed an enormous amount of writing. Books, websites, news articles, computer code, recipes, arguments online. Every time two words show up close together, the root thread between them gets a little thicker. Bread and butter show up together all the time, so the thread between them grows thick. Bread and engine almost never do, so that thread stays thin.
Now repeat that across billions of pages. After enough writing, the roots hold more than which words go together. They start to hold grammar, facts, and the usual shape of an explanation.
Chapter 4: What it was fed
[1:32 min] Since the roots grow from whatever writing they are fed, the writing itself matters a lot. And nobody hand picked it. The 2020 version of the AI behind ChatGPT was trained mostly on a huge copy of the open internet. The rest came from pages people had shared online, two big collections of books, and all of English Wikipedia…
That means the thickest threads grew wherever the internet talks the most, not wherever it is most correct.
The video goes on to point out that both words and numbers are broken into pieces and reassembled later. The bot is not reading and understanding anything. It is searchng and reassembling.
When you ask a question
Chapter 8: Your question is a spore
[3:50 min] Think of your question as a spore, the tiny seed a mushroom drops, landing on that web of roots. From the spot where it lands, the AI looks at every thread leading away and picks a thick one. That gives it one piece of a word. That piece is added to the end of your question. Then the AI does the exact same thing again, starting from the new spot. One piece.Then the next. Then the next one after that. A whole page of answer is that single move, repeated a few thousand times.
Chapter 9: It does not know the ending
[4:20 min] Picking one piece at a time has a strange side effect. When the first word of the answer appears, the last word has not been chosen yet. There is no plan. There is no finished answer hidden somewhere, being read out to you.
The AI only ever picks the next step from where it is standing right now. A good sentence is just a string of small choices that happened to fit together. Nothing goes back to check the ending against the start, because when the start was written, the ending did not exist yet.
Why AI models hallucinate
Chapter 14: Why it sounds so sure
[6:46 min] It has no way to check whether a thread is true. It can only measure how thick it is. To the AI, a real path and a made up one look exactly the same. So the wrong answer comes out in the same calm tone as the right one, at the same speed, with the same confidence. People call this hallucination.
That name makes it sound like a bug somebody could fix. It is not a bug. It is the same process that gives you good answers, working the normal way.
Much more at the link.
While the video was not written for this explicit purpose, it makes clear that what underlies a chatbot (large language model) is not a mind. It is a process that automates the previously produced output of minds, in many cases, the minds of people long dead or never heard of, in order to produce answers.
Here’s a link to the open access 2017 paper, “Attention Is All You Need, by eight Google researchers, that kicked it all off.
