Why is everyone using GPT incorrectly?

4 points by MagicMoonlight ↗ HN
These language models are not intelligence. They are not a store of knowledge. You've got companies like Google with a complete record of the entirety of human knowledge... and they're replacing it with Bob. Bob doesn't know everything and he can't do math but by god will he give it a go. He'll always give an answer even if he doesn't know. Is that a great idea? No.

With GPT they have a model that can understand human language. It's not an AI, it's not thinking, it can just understand language. But that's something huge in itself. I just asked ChatGPT to write me a love poem in a Shakespearean style, from a man to a lobster. And it did it. It did it well. That's a ridiculously good piece of software.

But now everyone is going crazy. They're ripping up functioning software and replacing it with the model because it can talk. Just because it can talk doesn't mean it can do things. It doesn't even know what a number is. Even if they were correct and it was now a sentient organism that can think and feel... that doesn't make it better than database. You could be the smartest organism in the universe but a biologically inspired brain is never going to be able to losslessly store 150b pages of content. So why are they trying to replace a database with Bob?

The correct way to use GPT is as a language model. It understands language. So use it to understand. Re-write Siri so that when I ask my phone to find me where I can buy tickets to the london zoo... it understands it. It doesn't try and bing it, it doesn't start playing "London zoo" by "the sheebles" on Spotify, it just does what I asked. If it can't do what I asked then tell me that. Then tell the developers so that they can know what to add next.

Re-write autopilot so that I can talk to my car. "Car, it's too hot in here." "Sorry about that steve, I'll turn the temperature down. By the way, the engine oil is getting a bit low. Do you know which oil you need to buy?" "STEVE WATCH OUT THERE'S A KID RUNNING INTO THE ROAD"

Imagine how good it would be if your existing software actually understood you and you could tell it what to do in normal language. But instead we get Bob, who doesn't know what a child is and keeps telling you that Mexico is a short drive away, take the next left at the cliff.

1 comment

[ 3.1 ms ] story [ 15.8 ms ] thread
> It's not an AI, it's not thinking

The correct term is LLM (Large Language Model). There was an article discussing how ChatGPT is essentially like a 'Blurry JPEG of the web'[0].

If it is AI, it's really just narrow AI, nothing more. We project our greatest hopes and desires onto anything labeled 'AI' because we secretly want strong AI.

[0] https://www.newyorker.com/tech/annals-of-technology/chatgpt-...