Hacker News Reader

Comment by foobazgt | Hacker News Reader

When you're able to find a prompt for ChatGPT where it doesn't have a lot of data, it becomes immediately and starkly clear how different a next word predictor is from intelligence. This is more difficult than you might naively expect, because it turns out ChatGPT has a lot of data.

View on HN

This also works fairly well on human beings. Start asking people questions about things they have no training in and you'll get bafflement, confusion, lies, fabrication, guesses, and anger. Not necessarily all from the same person.

> Start asking people questions about things they have no training in and you'll get bafflement, confusion, lies, fabrication, guesses, and anger. Not necessarily all from the same person.

It’s almost like we’ve taken humans and through school, TV and social media we’ve taught them to solve problems by writing essays, speeches, blog posts and tweets, and now we have human discourse that’s no better than LLMs - regurgitating sound bites when they don’t really understand the issues.

"Really understanding the issues" might just mean "deeper neural networks and more input data" for the AI though. If you are already conceding that AI has the same capabilities as most humans your own intelligence will be reached next with a high amount of probability.

Or, from time to time, you find people who readily admit that they have no expertise in that field and refuse to comment any further. Those people are hard to find so, that's true.

It’s a language model without grounding (except for code, which is why it’s so good at refactoring and writing tests.)

Grounding LLMs in more and more of reality is surely on AI labs list. You’re looking at a beta of a v1.

Reply on news.ycombinator.com