Oh how quickly the goalposts shift... If the test was done correctly and there was no data leakage, this absolutely does not compare to someone equipped with the answer key.

There's even a specific term for shifting AI goalposts: https://en.wikipedia.org/wiki/AI_effect

https://en.m.wikipedia.org/wiki/Moving_the_goalposts reviewing the 'see also' section of the goalposts wikipage is like looking at a syllabus for PR. People are terrible.

I also think there is a novel new effect happening which is that if one question anything about AI. For example: I tired to use Dall-E for something useful and it sucked. You get labelled...Anti-AI.

Being critical about certain things isn't allowed at the moment.

Kind of like the 2023 replacement for Anti-vaxxer if you don't want to take medicine every 3 months.

My honest opinion, is that while a lot of this technology is cool and novel, it's not perfect, it has uses and then it doesn't. What it does have is a lot of good marketing and PR behind it and so you're up against a lot of that too.

One example I like is the Dall-e homepage shoes the curated "best" images generated, which might be showing only 1 in 1000 or more of the best images. I was excited when I first tried it based on the front page. But now it's drawing people with feet for hands for me and it's super super fun and entertaining to play with, I don't feel it works quite as advertised most of the time.

Disclaimer: Not anti-ai and I took my Covid19 vaccines (in the beginning, no boosters)

Do you know that the answer key wasn't part of the train set of ChatGPT?

The author said that he comes up with new questions every year.

Still questions that I would assume are semantically similar to the questions you can find in exam prep material all over the internet. My point is that exams are a crutch we use to determine how well a person studied a subject. A crutch we use because we seem to lack better measuring devices. It's very much possible to ace an exam while at the same time being horrible at actually applying/working on a subject. I'd argue therefore that measuring how well LLMs perform on exams designed for humans is simply a more complicated Turing test, with all its shortcomings.

>Argument from ignorance, also known as appeal to ignorance (in which ignorance represents "a lack of contrary evidence"), is a fallacy in informal logic. It asserts that a proposition is true because it has not yet been proven false or a proposition is false because it has not yet been proven true.[0]

Why should we assume that the training set includes the answer key in the absence of evidence?

0. https://en.wikipedia.org/wiki/Argument_from_ignorance

We know that the training set is large and non public, and that answer keys are generally available and ML likes to output exact answers when it's been trained on the questions.

It's significantly more likely that a model outputting the right answer was trained in that answer than not

It compares to someone with a textbook and/or access to the internet.

...and the ability to synthetize both into a response to a question it has never encountered before

How many of these comments will get too and what is the actual use of the comment.

If you think about it critically, maybe we will find there isn't actually much use fo ChatGPT outside of it being a pretty cool party trick. Even if it gets "better".

I think it pains people to think about it this way but if we have bullshit jobs, and now have bullshit bots to do bullshit jobs, what's really changed? Likely we'll just end up with more bullshit.

I'm sorry but I think it wouldn't hurt for you to also be objective yourself when evaluating the system your defending.

When you think about it this way, if ChatGPT has the answer to all answerable questions loaded into it's memory, is any of this really that impressive? Initially yes, in the long term maybe so.

My feeling is we think these machines are impressive, because we think we're impressive, and ChatGPT is doing something that looks like what we do.

I guess we all thought steam trains were amazing and unfathomable at one point in time, is this really much different, it will be fun to find out :)

It does compare to inputting a math question into a calculator though.

Do calculators mean that we shouldn't teach multiplication anymore? That we shouldn't have tests for multiplication? The calculator will always get the right answer.

The bigger question imo is what humans are doing with these new tools. We no longer need the human computers of old now that we have calculators, and almost certainly we can innovate over automated MBAs now

Absolutely. You can Google the MBA questions and get directed to tons of pages that contain passing answers.

IIRC it got a 1300 on the SAT

That seems like the same situation -- it got 1300 on a practice test where a bunch of Internet sources discuss the exact questions and answers.

I tried feeding it an LSAT logic game which has multiple explanations on the internet. It got 2/5 correct, and for one answer it said “E, none of the above”

E does not say none of the above on this logic game. Though many other different standardized tests do use E as none of the above. So it was pattern matching to that and not the text in front of it. And it wasn’t using explanations

Probably a lot more? Part of it is we don’t have a definition of intelligence, so the best we can do is measure machines against our tests since passing tests is a thing intelligent people can do. This kind of performance (even if it’s less glamorous upon closer inspection) puts symbolic AI folks from the 80s to shame.

That’s the whole point of Attention is All You Need, the paper that opened the door to web-scale language models. No, you don’t need a brittle and cumbersome knowledge base, at least not for this. MBA professors are perfectly happy accepting an autocomplete on steroids.

* ChatGPT wins the Nobel Peace Prize

* ChatGPT wins the Nobel Prize for Literature

* ChatGPT wins a Ford Foundation grant for the arts

To be fair, the peace price has a pretty low bar.

Marshall ChatGPT leads Allied Armies to victory! [all ranks between Army General and Captain, non inclusive, have been declared redundant and eliminated with ChatGPT absorbing those ranks]

* ChatGPT creates 50 new Startups who fill all seats for the new YC 2023 batch...