Artificial Intelligence and Sheepdogs
Large language models (LLMs) represented by ChatGPT have been around for several years now, from initially amazing the world to people slowly recognizing their true capabilities. This wave of "generative artificial intelligence" bubble is gradually heading toward bursting.
People who have never owned a dog will similarly be amazed when they first get one, especially compared to other pets like cats or rabbits. A dog's intelligence and interaction with humans can leave a deep impression on first-time dog owners, with sheepdogs among the smartest.
Border Collies often astonish humans with their intelligence, and today's large language models are equivalent to Border Collies, initially shocking humans with their "intelligence." Given time, people discover that their intelligence has limitations - no matter how smart a Border Collie is, it will never understand quadratic equations.
What venture capital rushing into the AI field is trying to do now is train a Border Collie to become a world Go champion. Their idea is that with enough money and training, they can develop it into a Go champion capable of competing with Ke Jie.
Even if it doesn't work, it doesn't matter as long as there's a bigger fool willing to believe this story and buy in at a higher valuation.
Currently, that bigger fool appears to be Masayoshi Son.
Rather than being a successful venture capitalist, Masayoshi Son is more of a gambler. His most successful bet was investing in Alibaba, confirming Lei Jun's saying that "even pigs can fly when standing in the wind." Occasionally, betting correctly can make a gambling addict seem like a genius. Having made enough money from investing in Alibaba allowed him to casually lose money on projects like WeWork and Uber.
Gamblers find it hard to quit, and Masayoshi Son is no exception. Two years ago, he proposed "All in AI," making a big bet by investing $40 billion in OpenAI. If this bet fails, he will not only lose previous profits but also end up owing quite a bit.
Today's AI field has reached the "Emperor's New Clothes" stage, but no one is pointing it out. Everyone pretends the emperor wears clothes as long as fools are willing to buy their chips. As with previous bubbles, participants in the related industry chain will indeed make some money, but ultimately, all that remains is a mess.
From practical experience, anyone with some AI usage experience can easily conclude that it is completely unreliable in areas slightly beyond its capabilities.
A sheepdog can effectively herd sheep, but ask it to write a leadership speech, and it will be at a complete loss.
Ask an AI language model to produce leadership speeches, and it can excel; ask it to introduce academic content, and it starts making things up.
This is because large language models are trained on text. Internet text that appears frequently, such as leadership speech clichés, is easier for them to generate, while in obscure fields, they simply cannot generate content and can only fabricate.
It doesn't even know it's fabricating - the principles of large language models determine this. A large language model attempts to predict what the next most likely character should be. For example, in this article, after the word sheep, the most likely word is dog. Yes, the seemingly profound artificial intelligence is simply a more accurate "auto-suggestion" function.
OpenAI and numerous AI companies, as participants, must keep inflating the bubble, so they often present things like Ghibli-style artwork to confuse the public, which are merely trivial tricks.
The emergence of Deepseek has already proven that accurate large models can be trained through distillation without requiring massive financial investment. Still, the participants in this gamble are unwilling to admit defeat and leave the table.
All the text humans can use to train artificial intelligence has already been exhausted. A significant portion of "new" text being produced today is generated by artificial intelligence, which brings us back to the sheepdog metaphor - spending more money training a sheepdog to become a Go champion only results in it chasing its own tail, not only failing to improve its intelligence but also making it dizzy in the process.