ChatGPT

[Ari]

Unfortunately at the moment there’s no way to prevent shorts from showing up in my subscription feed, and YouTube has been heavily pushing Shorts, so creators who want to stay at the top of the algorithm need to be making shorts, which means my Subscription feeds is becoming more often than not a wall of shorts with actual videos interspersed like gems I must wade through the shorts to find.

(This does seem like a very Google way to annoyingly push a feature they will probably drop with little notice.)

[Crumb]

Oh yes, the subscription page has a lot of usability issues because they want you to use the algorithm so they don't make any improvements to it.

[Ensign Steve]

I was at my team stand-up this morning and my update was that I spent yesterday on a journey of discovery trying to figure out how to accomplish a specific task. A day of Googling and StackingOverflow had not gotten me there, but I had plans to meet with someone later who could probably be of help.

So I met with that person, and she was a huge help, and she gave me exactly what I needed. At the same time, one of my teammates asked ChatGPT my question. He sent it to me after, and it was the exact same solution. Down to the same code.

I literally can't even. Is this the singularity? Is it happening right now?

[Ensign Steve]

I mean in the interest of transparency, I don't know if it produced compilable, runnable code, and I kind of doubt it? But which library to import, which function calls to make, that was all I needed to get unblocked. I think it's still in a place where it can be a valuable tool for me, but not quite ready to replace me. For now, anyway. Still, it's fucking uncanny.

I've seen python generated by predictive models a few years ago, and it was gibberish. A lot of the familiar buzzwords and symbols, but unpaired braces, return statements in the middle of a function, things like that. This latest is so much closer to the real thing.

[Ensign Steve]

I watched this one last night, too:

ChatGPT does Physics - Sixty Symbols - YouTube

At about 13:30, he gives it a physics problem to solve in python, and it involves the Fourier transform, and I'm thinking that if it knows anything about anything, I'm going to be seeing np.fft in its code. And then it fucking did! And then at that moment in the video, the guy that's grading it says "It's scarily good" and I was like OMG, I know, right?!

I'm seriously down the rabbit hole here, you guys.

As far as using it to do your homework, depending on where you are in your academic career, the goal isn't necessarily to get correct, working code, but just to generate something that was convincingly a valiant attempt. Then you turn in it and get a B or C for at least having some idea of the general concepts.

[JoeP]

Originally Posted by Ensign Steve

I mean in the interest of transparency, I don't know if it produced compilable, runnable code, and I kind of doubt it? But which library to import, which function calls to make, that was all I needed to get unblocked. I think it's still in a place where it can be a valuable tool for me, but not quite ready to replace me. For now, anyway. Still, it's fucking uncanny.

I've seen python generated by predictive models a few years ago, and it was gibberish. A lot of the familiar buzzwords and symbols, but unpaired braces, return statements in the middle of a function, things like that. This latest is so much closer to the real thing.

What ChatGPT lacks is any attempt to check or validate or justify its output - because it hasn't been set up to do this - it's just trying to predict "what would come next" given a question.

But programming is one example of areas where it might be very useful, even profitable, to augment GPT stuff with testing and iterative improvement, possibly with reasoning. So it can say OK you gave me an English language problem, and you gave me some test cases. Here's some code that runs and passes those tests - and here's my reasoning for why this is a good solution.

This could already be possible in the labs.

[viscousmemories]

I've had it write a couple Python scripts for me this past couple weeks to solve specific tasks that were just "using data in this file get more data from this api and add it to the file" and it worked spectacularly well. Not perfect because as it turns out writing requirements is hard, but with a few tweaks totally usable code. One issue I had was saying (effectively) "solve for this error condition" and then it did, but then when I said "okay now solve for *this* error condition", the code it produced had allowed for the second error condition but no longer had the change it made for the first one. Does any of this make sense? It's very early here.

Also, we had quite a spat yesterday because no matter how hard I tried I could not get it to spit out an accurate formula to solve something in Excel I needed to do. And when I ask how to hide certain rows or columns conditionally, I do not appreciate being given step by step instructions on how to find and delete those rows or columns, even though it ended its advice with a warning that this was a permanent hiding method.

[Crumb]

An unending AI generated sitcom: Twitch

[Ensign Steve]

I finally bit the bullet and signed up for my own account. "ChatGPT, do the boring parts of my fucking job for me, please"

[Ensign Steve]

Okay, so here's one thing. I can look at those answers and know that they are good (and/or how they could be improved/fine-tuned) because I am already an expert in the domain.

If I asked it for some legal advice or medical advice (which I know we're not supposed to do), I would have no idea how to vet the answers. We're still going to need human experts on this, so that's something.

[viscousmemories]

I'm a paying subscriber to Dave Farley's "Continuous Delivery" channel because I learn something from every single thing he writes or says, but with that caveat this was interesting:

DON’T TRUST ChatGPT’s CODE! - YouTube

[Dragar]

I'm very impressed with chatGPT - it's pretty amazing how much you can do with a simple 'predict next token' model of language, and I think the results of just scaling this up and up and up have surprised a lot of people.

That said...

These language models are not thinking, or reasoning, or doing any of the same things we (seem to) be doing when we talk to one another. (Good discussion on this point: https://arxiv.org/abs/2212.03551). So I'm sort of sick of people on twitter suddenly surprised when GPT makes errors that simply illuminate this aspect.

These models also seem a lot better at generating new content that reliably reproducing information. I've chatted with former colleagues who have ran projects with LLMs to summarise literature, and their main finding was that these things tend to bullshit, confidently and convincingly. In a sense, that's exactly what they're designed to do: producing the most likely 'what comes next' value, irrespective of truth or logical/mathematical/physical consistency, is almost the definition of bullshitting!

So my predictions is that this sort of technology is mostly going to be useful for processes that don't need to be correct (so ruling out a lot of things like writing code, or answering factual questions, or being an AI assistant to 'do things'), but do need to be creative and/or can have a high failure rate.

So things like online automated help services, content generation (do you write bullshit for a living? Be worried!), procedural generation in video games (or even online RPGs!), idea generation...exciting times to come in the creative industries, and that's before we talk about image and SFX models (also very impressive).

[viscousmemories]

Yeah, maybe what makes a great creative writer is the ability to produce unpredictable combinations of words in a way that still embeds meaningful expressions of human experience that humans can decipher but which AI is a long way from understanding. I can imagine AI being just as good at assembling improbable combinations of words as it is at producing predictable patterns, but I suspect its a long way off from being able to imbue unpredictable text with relatable subtext. Then again we humans are so good at spotting patterns and deriving meaning from the meaningless, maybe we would revere word salad from AI as great poetic genius.

[Kamilah Hauptmann]

image.png

[-FX-]

That is simply awful. It's not awful like a badly written piece by a struggling human, it's awful because it is just stupid, unimaginative, and meaningless.

[beyelzu]

On creativity:

Perhaps, I have an overly romantic idea of what constitutes creativity, but it seems to be that when AI like ChatGPT verges into seemingly creative behavior it’s slippage or mistake on the AI’s part. The better the program performs the more banal and obvious to everyone the answer should be.

A user of the program certainly could look through query answers for interesting slippages and combine them in creative ways, but that wouldn’t mean the program was creative itself.

Like I said, maybe I have an idealized view of creativity.

[ceptimus]

Is it even a sonnet? A sonnet should have 14 lines with (simplistically) 10 syllables per line, and some form of rhyme scheme connecting groups of lines - often three groups of four lines, followed by a final rhyming couplet.

It does put a capital letter after (roughly) every ten syllables, so let's write it out in the usual format, but with line numbers and syllable counts:

01 Beneath the moonlit stream the beaver toils (10)
02 With gnawing jaws and sturdy webbed feet (9)
03 He builds his dam with branches and with soils (10)
04 A fortress strong to keep his kin complete (10)
05 With fur as dark as night and eyes that shine (10)
06 He works with purpose driven by his will (10)
07 To shape the land and make it truly thine (10)
08 A home for him and all his kin to fill (10)
09 But man with greed and malice in his heart (10)
10 Has sought to bring the beaver to its end (10)
11 With traps and guns he tears the dam apart (10)
12 And leaves the beaver with no home to tend (10)
13 But still the beaver builds with hope and grace (10)
14 A symbol of our Earth's enduring place (10)

I guess we're supposed to say 'webb-ed' on line two, to get the meter right
... and the rhymes are pretty good.

I give it for being an impressive example of a state-of-the-art language prediction model.

But only as a poet.

[Ensign Steve]

In the Rob Miles video, he talks about that specific misalignment. Because of how the reward function is implemented, the goal isn't to write a sonnet, the goal is to get a "this is a good answer" reinforcement from a human, who may not even know what a sonnet is. The simulation might actually "know" the rules of constructing a sonnet, but the simulacrum has "learned" that it gets rewarded by making a nice poem that isn't a sonnet.

[beyelzu]

We will reach the lowest common denominator with the greatest of efficiencies. You thought human executive producers were exploitive, ripped off and recycled ideas, wait til you see some Fall’s AI produced lineup

[LarsMac]

Originally Posted by -FX-

That is simply awful. It's not awful like a badly written piece by a struggling human, it's awful because it is just stupid, unimaginative, and meaningless.

Not that it's Nobel material, but, It's not THAT bad.

[Kamilah Hauptmann]

He’s the Mikey of the internet, he hates everything.

[JoeP]

Originally Posted by Kamilah Hauptmann

He’s the Mikey of the internet, he hates everything.

*pretends* to hate everything, desperately chasing that high

[Kamilah Hauptmann]

Next stop: Managing Ibuprofen levels to shitpost through the arthritis flareups.

Time, it comes for us all.

[Kevlar]

Originally Posted by LarsMac

Not that it's Nobel material, but, It's not THAT bad.

He's projecting

[-FX-]

"With traps and guns he tears the dam apart"

This is where an AI just fails,
on a level no human would
A child can grasp with ease the fault
The AI misunderstood

The subtle things we never state
We know them from our youth
The computer misses out on this
And stumbles over truths

Simple measures, a rule of thumb
Our basic humanity
Deeply known but seldom spoke
The AI can never see