Hacker News Reader

Comment by fullsend | Hacker News Reader

Have you used it? Start working on anything. Next time you go to Google how to do something, or want to name something, or want to convert formats, or a dozen other things, just ask GPT. Say “give me 5 cool names for this”. Any task. “Show me a garden plot given plants from this time of year.” “Tell me about different types of LLCs and the tradeoffs.” “This code isn’t working what’s wrong.” You’ll quickly be able to feel how it can do your work for you, or shortcut hours of digging through garbage blogs and ad-infested pages served by Google. Go read one of the Twitter threads with 100 different insane, mind blowing examples of use cases.

View on HN

That's not search though. Those features will thrive when users pay for the usage.

General search for anything that involves competing businesses will begin getting gamed immediately and any providers need to be better than Google in preventing it or suffer the same fate but with with full sentences.

Example:

How can the AI get enough info to answer: "Is Frank's Garage or Auto Joe better value for money if I need my radiator fixed".

A generic explanation of what "good value" means is utterly useless. So is just restating their ads.

It might beat google because of the interface but it won't revolutionize search on it's own.

I did.

GDP can get very self repeating. If you ask it for 5 cool names you soon realize it uses the same scheme every time. It doesn't convert formats at all as far as I can tell. If I ask GPT to repeat something in metric it makes up facts about metric instead of repeating what it said. From what I can tell it has zero concept for measurements. And if you phrase your questions awkward enough it would be easy to generate highly false information if you ask clear questions like the LLC example and I highly doubt it can actually differ between country laws and stuff.

Yes we are getting there, but GPT is not it. Yet

> just ask GPT.

And then return to Google to verify that every fact presented by GPT is correct, and there no errors in numbers, names, definitions, etc. etc?

What we are seeing now is a proof of concept, what a highly generic language model can do. I'm sure there is work right now to add a new layer for confidence and truthfulness of answers. I'm not in the field but I guess that can be a another layer that could weight GPT choices in its probability space. Overall we could even have a confidence score for some answers.

I`m very very curious about these approach. In my understanding most of AI models: GPT, StableDiffusion, etc uses statistics approach for learning.

I`m not sure is it really possible to: - add new layer of confidence, - ability to provide correct source for claims, - ability to fine-tune one answer, in the nearest future.

Example: can we teach GPT to learn that Sun color is cyan without retraining it from zero? So on any question (no matter how sophisticated it is) it always "know" that sun color is cyan? Can we teach GPT anything without retraining it?

How about this softball question I just tried?

Q: "What was the most energy efficient electric vehicle available in 2022?"

A: I'm sorry, but I don't have any information about specific electric vehicles that will be available in 2022. In general, the most energy efficient electric vehicles are those that have a large battery pack and a powerful electric motor, which allows them to travel long distances on a single charge. Some of the most energy efficient electric vehicles on the market today include the Tesla Model S, the Nissan Leaf, and the Chevrolet Bolt. It is possible that even more efficient electric vehicles will be available in 2022, but I don't have any specific information about them.

Yeah, mind blowing. I especially chuckle at "This code isn't working what's wrong". Maybe it can help with CS 101 homework and hobby projects. Even if I had permission to upload my company's code, that consists of hundreds of thousands of lines of Java, Python, Perl, C#, C++, and SQL that runs on wildfly server clusters and oracle databases, spread over dozens of repos, all knitted together with thousands of lines of build scripts and gitlab pipelines. Oh and its only partially documented. Oh and there are contractural requirements, so the program will have to understand the contracts and the accepted interpretations of the requirements before making recommendations. This thing is not going to solve even the relatively trivial error I fixed today because it wouldn't be visible from any syntactical analysis. It would have to have sophisticated understanding of the software architecture itself, and stand up a duplicate system and run specific tests to verify that the problem is what it is. It would also lack context as to program objectives, because the problem wasn't strictly a malfunction, it was a vestigial feature that was supposed to have been removed but was still erroneously being deployed and interfering with its replacement.

And my company is hardly unique in that regard, nor is our codebase the largest. ChatGPT is a consumer demo; which is fine. But if I showed you a newer, faster, better speedboat would you be talking about how it's going to alter container shipping and naval warfare?

Maybe this is the start of something huge, but we have no idea what the backend of this thing looks like to judge what its potential actually is. Everyone's just slapping their wildest AI sci-fi fantasies on it the same way every new harebrained green tech with an impressive CG render is going to save the world from climate change and Crypto is totally going to replace all traditional finance and even the dollar itself. A few months ago I literally had someone counter my argument against crypto with "but your argument doesn't account for FTX, they solve a lot of what you're pointing out!", another situation where people were projecting what they wanted to see onto a black box. Even rich people and companies with lots of money. Don't make the same mistake here, wait until it does something real.

“ Everyone's just slapping their wildest AI sci-fi fantasies on it”

Well said. I agree there is good reason to be a little skeptical of the hype.

You’re not asking it to rearchitect your whole company and software stack. I think you’d be surprised by how much value is there, instead of writing HN essays about what we all know isn’t there. Same goes for crypto.

Judging by many of the comments, clearly we don't "all know" what isn't there. In fact, unless you work for OpenAI or the Microsoft team acquiring it, we barely know the surface of what is there.

ChatGPT has value, but it isn't world-changing in this state, and we have no information as to how much potential it actually has. I'm reminded of Google's AI vaporware from several years ago where they had conversational AI that could call humans and make appointments for you without hitting the uncanny valley. Was a sweet demo, public release was supposedly just around the corner. People were talking about how it was going to revolutionize how we interact with technology. Still waiting.

Reply on news.ycombinator.com