Cipheron on 13/5/2023 at 10:16
Quote Posted by demagogue
Yeah they're mostly there for color and generic advice. I'm just curious if they might sometimes give catastrophically bad advice because of their wonky reasoning. XD Technically real advisors could do that too though. So I think the real rub is whether they have an authentic voice in giving whatever advice, like the economic advisor talks like an economic advisor and gives the kind of advice they'd give, etc.
I think with the ChatGPT one, it could at least ask you questions about the situation, then give advice tailored based on your interaction and feedback, so at least it wouldn't always give bottled advice which may not be appropriate.
Ok, found a way to write some very awkward basic science articles:
(
https://pastebin.com/9AY1fLDh)
Quote:
The tantalizing exploration known as the "Two Slit" experiment unveils a captivating spectacle, where the enigmatic interplay of human-like particles and their interactions takes center stage. Picture a delicate dance of particles, reminiscent of the most intimate encounters between individuals, revealing the secrets of our very essence.
In this exquisite performance, a stream of vibrant particles, akin to the smallest units of human vitality, traverses a barrier, resembling the elusive threshold between two bodies. The barrier, with its two slender apertures, simulates the delicate realms of human anatomy yearning for connection.
...
Quote:
Imagine, if you will, a tantalizing thought experiment that delves into the depths of desire and perception—the enigmatic tale of Schrödinger's Cat. Within this exquisite metaphorical dance, the cat and the mysterious box symbolize the feminine allure, while the observer represents the masculine essence, entwined in a delicate interplay of longing.
In this beguiling narrative, a beautiful feline, vibrant and alluring, embodies the essence of the feminine archetype. As if guided by a cosmic force, she finds herself enclosed within a mysterious box ...
...
Had to cut off the Shrodinger one there, it was getting a bit too steamy.
Cipheron on 14/5/2023 at 08:31
Hmm, new idiocy in GPT land.
Now, they're starting to roll out web access for ChatGPT. I don't have it yet.
But some people are complaining it doesn't work. They give it some huge task that would be a colossal web-scraping endeavor.
For example asking it to decorate a room according to a theme, then telling it to go through the entirety of Amazon.com and picking out a suite of products that match the criteria. Like, it can barely pull of numerical comparisons, and fails as soon as you throw in even the simplest wrinkle, see my examples at the bottom. How is it going to divine a selection of products with many factors each, then work out a *suite* of them that achieve some aim?
Then they wonder why it doesn't work, get upset at people who explain to them that this probably won't work.
First, web scraping is a real art. How many links are on an Amazon page? 100 maybe? If you clicked every link, then clicked every link on the linked pages, and then a third link deep, now you have opened 1 million pages. So you have to be clever and specific about only clicking on the useful links, and avoid clicking on anything else. An AI can't just brute-force logic this.
Also, ChatGPT doesn't actually have a working memory other than the text. So even if it can scan through all the items matching a specific search on Amazon, it's probably not capable of actually remember and comparing items that it's seen to meet some specific type of criteria. Quite simple comparison questions can in fact trip ChatGPT up, even when all the tokens are right there in its memory, since the lack of "working memory" means it can't actually step through something to check.
Here's a couple of examples I generated:
Quote:
ME: given this list [12, 6, 18, 3, 15, 9, 20, 2, 17, 5\]
add the third largest and second smallest number
To find the third largest number in the given list
[12, 6, 18, 3, 15, 9, 20, 2, 17, 5\],
we can arrange the numbers in descending order:
[20, 18, 17, 15, 12, 9, 6, 5, 3, 2\]
The third largest number is 17.
To find the second smallest number,
we can arrange the numbers in ascending order:
[2, 3, 5, 6, 9, 12, 15, 17, 18, 20\]
The second smallest number is 3.
Now, we can add the third largest number (17)
and the second smallest number (3):
17 + 3 = 20
ME: which value in the list is closest to 13?
To determine which value in the list
[12, 6, 18, 3, 15, 9, 20, 2, 17, 5\] is closest to 13, we can
calculate the absolute difference between each number and 13 and
identify the number with the smallest absolute difference.
Here are the absolute differences for each number:
|12 - 13| = 1 |6 - 13| = 7 |18 - 13| = 5 |3 - 13| = 10 |15 - 13| = 2
|9 - 13| = 4 |20 - 13| = 7 |2 - 13| = 11 |17 - 13| = 4 |5 - 13| = 8
From the calculations, we can see that the number 15 has the smallest
absolute difference from 13, which is 2. Therefore, the value in the
list closest to 13 is 15.
If you ask for largest or smallest in a list, it can just give that to you straight away. It's obviously been coded as a "short cut". If you ask for the "third largest" or "second smallest" in a list - it can sort the list (hard-coded sort algorithm), and then tell you the right ones, based on their position, not by comparing values.
However - the one where it fails is the "closest to" example. Then, it ends up guessing, and might guess wrong. It either can't or wouldn't do the sorting trick this time.
To test this further:
Quote:
Here's a list of 20 random objects with prices rounded to two decimal places:
Pen - $3.27
Coffee mug - $8.54
Sunglasses - $12.91
Headphones - $49.68
Wallet - $21.75
Notebook - $6.82
T-shirt - $15.46
Umbrella - $18.29
Keychain - $2.93
Watch - $67.13
Backpack - $32.57
Water bottle - $9.83
Phone case - $14.79
Laptop stand - $27.94
Wallet - $19.62
Plant pot - $5.38
Bluetooth speaker - $42.15
Scarf - $11.27
Mousepad - $8.79
Duffel bag - $36.45
> what is the cheapest item
The cheapest item on the list is the keychain, priced at $2.93.
> what is the most expensive item
The most expensive item on the list is the watch, priced at $67.13.
what is the second most expensive item
The second most expensive item on the list is the laptop stand,
priced at $27.94.
Ok, as soon as it's non-simple data, i.e. names and prices, it's no longer able to pull off the "sorting trick" and sort the list to get that answer. So it guessed wildly wrongly about which was the second-most expensive.
And this is when the data is all there, right in it's frame of working memory, and about as simple as it gets. how exactly is it gonna traverse complex categories on Amazon and find and compare items to each other?
Cipheron on 16/5/2023 at 07:59
Quote Posted by WingedKagouti
It's not about training, it's about capabilities. ChatGPT would need to be coded differently to be able to actually understand the texts it has been fed.
Yeah, that reminded me of something I've seen come up in conversations.
If you say to some people that ChatGPT doesn't understand what it's doing, then they retort "but I asked ChatGPT whether it understood and it gave the explanation, so it clearly understands".
I'm pretty sure people can see the flaw here. ChatGPT can spit out a textbook definition of how something is supposed to be done, but then contradicts itself by failing to get the right answer, which just shows it didn't actually follow the method.
EDIT - In the example I gave with the prices and where it failed at picking the second-largest, i could definitely have followed that up by asking ChatGPT how you'd go about finding the second-largest and it would give me a decent description of the steps needed. The issue is that it clearly DIDN'T follow any such process, nor ever noticed that it had failed to do so.
