Live data from Hacker News

What is GPT-3? written in layman's terms

tinkeredthinking.com

61–70 of 70 posts

Re: What is GPT-3? written in layman's terms

#61

Let me guess. GPT-3 wrote this article??

That was my thought as well. It has a lot of the idiosyncrasies I've noticed in other GPT-3-generated articles: coherent, but meandering, with a lot of cliches and rhetorical questions that don't really add to the article. Of course, you also see that in human writing. :-)

Re: What is GPT-3? written in layman's terms

#62

I have some issues with how this article describes GPT and its impact. The scenarios that are described as using GPT-3 remains the stuff of science fiction as far as I know. They would require major technological breakthroughs in areas of autonomous agents, combining language models with 'concept-based' models, and more mundane things like labelling (can you imagine how difficult it would take to accurately label all…

A much more realistic and accurate take on the tech than the article. One combination that can work is to have GPT3 with an unskilled human judge in the loop. The human can discriminate with common sense while the machine generates good jargon with correct grammar.

That seems even worse!

Jargon (sometimes) isn't just obfuscation; it carries important shades of meaning that would be tedious to spell out every time. 'Homicide' and 'murder', for example, are sometimes used interchangeably, but are not actually legally interchangeable. You would not want your involuntary manslaughter plea upgraded to murder, a more serious--and therefore more severely punished offense.

You could, of course, try to train "unskilled labor" to detect things like that. By the time you are done, the labor won't be unskilled and you will have reinvented the paralegal.

Re: What is GPT-3? written in layman's terms

#63

How long until someone builds GPT-4 with several trillion parameters? How much would that cost?

Training GPT-3 costs only a few million dollars. Scaling up is still pretty cheap. I wouldn't be surprised if we see a quadrillion parameter model in 5 years, given the potential value.

Google has actually already trained a trillion parameter model IIUC [1], though that was a Mixture of Experts so was way cheaper to train.

[1] https://twitter.com/lepikhin/status/1278174444528132098

Re: What is GPT-3? written in layman's terms

#64

Earlier quoted context omitted.

I'm not convinced that "capable of reasoning, but not consistently" is a meaningful claim. The examples seem to primarily consist of people spending hours trying things, until eventually GPT-3 outputs a chunk of reasoning they could personally do in seconds. Does that mean that GPT-3 is doing the reasoning, or does it mean that GPT-3 is an English-based lookup table and they managed to find a clever sequence of searc…

Have you seen the database prompt? https://www.gwern.net/GPT-3#the-database-prompt

As I mentioned, I don't think any single prompt can demonstrate the presence of true reasoning. If the prompt isn't shown to broadly generalize, it might just be doing a text match to something that was said before on the depths of the internet. You can see this in the next section; Kevin Lacker gets GPT-3 to demonstrate it knows some basic trivia questions, but it "knows" any prompt with the same textual structure as a basic trivia question, even if the prompt is nonsense. This strongly suggests that it's parsing out key words and doing a lookup on them rather than accessing a consistent internal model.

Re: What is GPT-3? written in layman's terms

#65
> If prompted correctly it responds identically a human.

It appears to be grammatically incorrect. It made me think "ok, probably written by a human." But then also realized that typos and grammar issues are probably prevalent in the data set. How would they manifest? Will they reinforce emerging changes in language? (eg, think of how the meanings of words have slowly mutated). And how much of the scraped content is itself written by a AI?

Re: What is GPT-3? written in layman's terms

#66
post #44

To me, what GPT-n really tells is how redundant human text in general is. It doesn't provide a lot of new information (if it does, they are accidental), but more or less wandering around the given topic. It is the redundancy that gives the system a space to churn.

The reduntancy is not low. GPT-* are basic "pattern generators" for our language organ. We have pattern generators for our gait for example -- all those synchronized movements of so many muscles can be effectively controlled by a well understood circuit, to the point that even a deafferented animal can still be be made to recognizably walk.

GPT is the equivalent, but for language: it can be used to phrase thoughts. Real language however, has thoughts, GPT doesnt have any thoughts. People are impressed by how many responses it has learned, but forget that it is contains a lot of gigabytes of "compressed" text associations. It needs "something else" to become actually useful.

Re: What is GPT-3? written in layman's terms

#67

I have seen comments on Reddit that smelled as though something like this was used to create them. I don't see how, in future, we are going to be able to meet up and discuss constructively with strangers online, if this thing can be used to emulate people.

Voting systems exist to promote insightful comments and suppress the noise. Those systems may or may not need to adapt.

Would it matter if on occasion some value is gleaned from generated content?

Re: What is GPT-3? written in layman's terms

#68

I have some issues with how this article describes GPT and its impact. The scenarios that are described as using GPT-3 remains the stuff of science fiction as far as I know. They would require major technological breakthroughs in areas of autonomous agents, combining language models with 'concept-based' models, and more mundane things like labelling (can you imagine how difficult it would take to accurately label all…

A much more realistic and accurate take on the tech than the article. One combination that can work is to have GPT3 with an unskilled human judge in the loop. The human can discriminate with common sense while the machine generates good jargon with correct grammar.

Yeah, I think those will be some of the first wave of language generation applications. Stuff like Buzzfeed but without as many writers because the language model is generating potential articles and the humans pick which are good (and maybe clean them up). Similarly, disinformation or noise attacks on social media - with just a few people overseeing the language model as it generates potential posts/replies.

Re: What is GPT-3? written in layman's terms

#69
post #14
post #9

Earlier quoted context omitted.

> Does that mean GPT-3 was trained on an arbitrary, huge database of text? Yes, 500 billion words, for a model with model 175 billion parameters.

So that continues my question: how much of that text would be on the private domain? If there is, is it all legally purchased?

This is an excellent question which applies to a lot of machine learning datasets. AFAIK there is no specific licensing of much of it (licenses you see are generally attached to the labels: the images/text/etc are often not even part of the download and you need to go scrape them yourself) and it's often claimed that the results of the network are free from the copyright of the training data used to create it, but this is contentious and has definitely not been tested in court.
Post reply on HN