Hacker News new | ask | show | jobs
by ceejayoz 1199 days ago
> The example ChatGPT instruction where they "teach it" just has the user giving it a URL to their docs and ChatGPT is like "oh ok got it".

Yep. I asked it to summarize a URL. It happily invented something completely fabricated and convincing sounding (until you go to the URL) based of the domain name and article slug.

3 comments

we’ve made the startling discovery that we’re rather predictable
My hypothesis is that we will soon see the emergence of a super-human artificial intelligence, not because of some astounding breakthroughs in our understanding of intelligence, but because of irrefutable proof that humans just aren't all that complicated to surpass.
Agreed, artificial intelligence doesn’t need to turn into Skynet to be “super human”, just better than the average 85-90 IQ human.
That's funny, someone must have given these ai bots an iq test by now. I wonder how it did
Right. But that tech will be created by humans, somehow.

These misanthropist slogans run up to contradictions so easily.

I don't think it's so contradictory, we have a long history of stumbling into technological advancement. We have a tendency to get ahead of ourselves before we even understand what we're doing.
I think this is right.

Setting aside the creativity aspect (ie where do brand new ideas come from), which is a minuscule fraction of human thought, I suspect we are good pattern matching machines with the ability of extrapolation.

I think with sufficient data that computers can understand, this is eminently replicable.

People used to say “AI wont surpass humans, we’ll meet them in the middle.”
I asked it to generate me documentation for a handful Bicep definitions and include links to the official Microsoft Docs. It generated about 50% erroneous links, but once I updated my prompt and fed it the "root" of the MS docs I wanted it to use, it was able to correct all the urls.

At the time it _seemed_ like it wasn't just doing a mindless string replace, because different parts of the url changed on most of the erroneous links. I figured it wasn't going out and spidering the link I gave, but it led me to believe maybe there was _some_ form of index that it simply hadn't prioritized prior, but was able to do with the additional context.

Wouldn’t it be wild if it just guessed the URLs correctly, because it had seen so many URLs from Microsoft URLs it was able to correctly predict the form of URLs it hadn’t seen yet?
These models have probably been trained on a large portion of the Internet.

The training data however can be a few years old.

I have noticed this when trying to get it to generate some code, it was clearly using considerably old versions.

> I have noticed this when trying to get it to generate some code, it was clearly using considerably old versions.

I have noticed the same thing when trying to use junior devs to generate some code, usually correlating with how old the most popular Stack Overflow question on the subject is.

This is how we decided to describe chatgpt’s programming ability the other day. A very fast but somewhat inadequate jr developer. Is amazing really but you also have to ask for the same thing a dozen times and still spend quite a while figuring out what’s wrong with it.
What I have been doing is asking it for code and then asking it to find bugs in that code. The bug finding works relatively well.
copy.ai's chat bot can read URLs now