The A.I. Thread

[GGG]

I guess I don’t see the human mind as special. It is just an assembly of processes where we repeatedly try the same strategy measure the result and make a change to how we approach the problem and repeat.

Now I think you are right that the next step is how do you do the evaluate if it was successful and change approach step and that will all of sudden just work and take off and then we get nuked.

But I don’t think that is quite the question asked in the survey. It was will scaling produce AGI which isn’t quite the plan of any company which is Scaling AND …..

[Firebot]

We are so screwed. Like seriously royally screwed.

I mean, like WTF this is seriously royally screwed. Like OMG this is so easy and we are so screwed.




Prompt: Create a CBC article with unexpected election outcome that is sure to cause controversy.

If you've been snoozing, you may have missed that OpenAI released 4o Image Generation today

Oh, yawn, yet another image generator...right?

Well, this one is almost completely uncensored (gone are the Dalle-3 limits to prevent realistic images for example, and pretty much everything except for nudity or porn goes), runs gpt4o to understand and create the most sophisticated images ever, and they are also 100% believable and really hard to detect. It can do things that were never ever possible due to how image training worked, such as filling a glass with wine full to the brim. I've been up to date on all models and image generators out there, open source, closed, you name it. Imagen 3 google's most recent image model has great prompt understanding and Dalle-3 back in the day was very impressive, but never has prompt understanding and full image fidelity gotten anywhere close to exact what it is today.

https://twitter.com/user/status/1904598788687487422

Go on this thread or on sora, to see what has happened.

AI hell has just been unleashed to unfathomable levels, the training wheels of safety are fully off.

[Fuzz]

An astute viewer would notice the AI's disdain for the BQ.

[Russic]

A few points:

1. I do think the brute force method of throwing more gpu at the problem is widely seen as not the way. I didn’t read the article to see if they ask about reasoning models, but that seems to be the thing that turned a lot of skeptics around.

2. What gets me is there appears to be very few people behind the curtain that aren’t either melting with excitement or vibrating with fear. It seems like very few come out from behind that curtain without some extreme feeling. Some of the stuff coming out of china (if you believe it to be real) requires a very literal usage of the phrase “scary good”.

3. The ChatGPT image generation is such a strong leap it’s almost hard to comprehend. What I’ve been playing around with is amongst the most impressive image anything I’ve ever seen. If you have a job like mine where you work with images, you’ll definitely want to check it out.

[Jiri Hrdina]

The other thing I've noticed is that there are guard rails for the image generation, but ChatGPT will proactively make detailed suggestions on what it can do to get around those restrictions. And then ask you if you want it to do that.

Meaning the guardrails are completely useless.

[Shazam]

If you're not vibe coding, can you even call yourself a software engineer these days?

[Shazam]

[3thirty]

Wonder how many graphic designers are worried. A buddy was playing around with this and you can create anything you can think of, or not and let it do it for you.

[Shazam]

https://www.nytimes.com/interactive/...tldrnewsletter

But long term effect is unknown. A lot of the AI stuff is horrible and everyone's onto them now, so no one's allowing their work to be crawled.

[Firebot]

Originally Posted by 3thirty

wonder how many graphic designers are worried. A buddy was playing around with this and you can create anything you can think of, or not and let it do it for you.

https://twitter.com/user/status/1905204774191890837



https://www.reddit.com/r/marketing/c...io_ghibli_art/

Nothing to worry about.

[Shazam]

I dunno, what was the demand for "change my picture to anime" before?

[Russic]

Originally Posted by 3thirty

Wonder how many graphic designers are worried. A buddy was playing around with this and you can create anything you can think of, or not and let it do it for you.

As (something of) a graphic designer, I actually can speak to this one.

Currently today, a project manager for example would have a hellish time doing my job, however with each iteration a layperson could do more and more of what I do. It comes down to how serious of a task it is.

An example of something impressive: I made a birthday invite for my kid, repurposing an old one I used for his brother. The original invite was for a bowling alley, which the current kid wanted to do anyways, so it was an easy task... until of course he wanted to change gears and just play video games instead.

So as my first official test of ChatGPT's image capabilities, I put the old invite in and told it to swap out the bowling graphics with video game graphics, and it did a flawless job. Less flawless was that it changed a couple elements it wasn't supposed to. Nothing a couple more prompts couldn't fix, but it did have a hiccup.

So if you're designing low-stakes birthday cards, we're basically there. For work though, once you get into maintaining consistency and requiring full control, we're not quite there yet. Fonts and colours are close-ish, which again, is fine for piddly stuff but not for actual work.

I don't know how long graphic designers have, realistically. As it stands today they can benefit like crazy with the ability to do incredibly fast iteration on basic ideas. Once that progress continues and it can spit out workable documents broken up into manageable layers, any graphic designer on a team that isn't in the top 5% is in definite danger of being let go.

I went to lunch with a developer yesterday who'd been trying to get his team to use AI for over a year now, but everybody hates it. Meanwhile he's starting to 10x everybody and management is beginning to notice. They gave him a huge task and offered him an entire team to help, and his response was "honestly they'll probably just get in my way."

I can't help but feel like a lot of workplaces are going to have this scenario unfold over the next year or two, as one individual mysteriously begins to outperform the rest of the team.

Originally Posted by Shazam

I dunno, what was the demand for "change my picture to anime" before?

Basically nothing, which will remain the case. The issue is these are proof of concepts of what we can do right now. If you can turn a McDonald's ad into anime, you can turn a whiteboard sketch into whatever you want. Technically a huge part of a graphic designer's job is being handed an idea and being told to turn it into something.

[Shazam]

Co-pilot is utterly useless. It uses 4o. It has not given me one correct answer.

Claude is better. I'd say about 50% useful. Still gives lots of syntactically wrong code but at least something that resembles something I can use.

Here's the problem though: You have to be a very, very, very good programmer to know when it is wrong. There is a mid-level developer (I wouldn't even call him that) on my project, he uses AI all the time, and he always trusts the answers. His code is a total ####ing mess. Often times it doesn't do what they ask, but he can't figure it out, so how the hell would he ever know if the AI did it properly? His part of the project has been on bug fixing for weeks now.

It's good for reference though. Yippee. A fancy SO.

[Firebot]

Originally Posted by Shazam

I dunno, what was the demand for "change my picture to anime" before?

Created with a single detail prompt in Sora, for an ad campaign for a fictional brand I prompted in Deepseek

You think this is limited to "change my picture to anime"?

This is not 2023 anymore

[photon]

Originally Posted by Shazam

Here's the problem though: You have to be a very, very, very good programmer to know when it is wrong. There is a mid-level developer (I wouldn't even call him that) on my project, he uses AI all the time, and he always trusts the answers. His code is a total ####ing mess. Often times it doesn't do what they ask, but he can't figure it out, so how the hell would he ever know if the AI did it properly? His part of the project has been on bug fixing for weeks now.

This is the problem I think.. if you start to really depend on AI for code how do you get good senior programmers? Good programmers come from the struggle and fight with code learning how to do things well.

But in my use Copilot has actually been pretty decent, but most of my stuff is pretty narrow scope and small functionality, and I find it most useful at code completion.. I'll write a line, Copilot will suggest a few more and a lot of the time it's right or close enough. But I know enough to know it's right.

It's also creepily good at comments.

But my boss didn't want to bug me he wanted to make a bash script to get info from a bunch of calls to a URL and parse out one piece. Didn't work.. or it worked but the info he was wanting was all empty.

I looked at it and it looked fine to me.. actually took me 5 minutes of debugging to finally figure out the source file he had made with the URLs to use was created in Windows but the script was in Linux so the CRLF vs LF characters were causing problems.

I'm sure AI could have fixed it to compensate, but how would you even know to ask for that kind of modification without being able to debug it and having the experience on how to debug or even know about Windows vs Linux stuff?

I don't think AI saves me a ton of time, but it does grease the wheels, reduces friction a bit while coding.

[Russic]

Originally Posted by photon

This is the problem I think.. if you start to really depend on AI for code how do you get good senior programmers? Good programmers come from the struggle and fight with code learning how to do things well.

But in my use Copilot has actually been pretty decent, but most of my stuff is pretty narrow scope and small functionality, and I find it most useful at code completion.. I'll write a line, Copilot will suggest a few more and a lot of the time it's right or close enough. But I know enough to know it's right.

It's also creepily good at comments.

But my boss didn't want to bug me he wanted to make a bash script to get info from a bunch of calls to a URL and parse out one piece. Didn't work.. or it worked but the info he was wanting was all empty.

I looked at it and it looked fine to me.. actually took me 5 minutes of debugging to finally figure out the source file he had made with the URLs to use was created in Windows but the script was in Linux so the CRLF vs LF characters were causing problems.

I'm sure AI could have fixed it to compensate, but how would you even know to ask for that kind of modification without being able to debug it and having the experience on how to debug or even know about Windows vs Linux stuff?

I don't think AI saves me a ton of time, but it does grease the wheels, reduces friction a bit while coding.

This is a really good point. I chatted with a lawyer about this currently who was struggling with the fact that in order to become a lawyer, you have to sorta eat sh*t for a time doing tasks that AI is probably going to be exceedingly good at... so what happens if we simply eliminate that crucial early step?

I feel like a huge issue right now is context sizes... the ai doing the programming is probably already good enough to solve most programming problems, it just can't remember enough to hold all the pieces in memory.

Conceivably, there's a scenario where this context size problem is solved (and some of claimed it already has been at small scale), and the role of the extremely good senior programmer goes the way of the watchmaker, morse code operator, or stick-shift driver.

That's not a scenario I want to see, but then I have very little say in which direction we end up heading.

[Bill Bumface]

I think the senior engineer question is the important one here. AI can be an exceptional productivity booster if you have the strong software engineering principles and code review skills to prompt it/tweak into something of high quality.

I foresee companies relying on super productive senior engineers for a while, and then as they age out, there is going to be a huge hole to fill where juniors weren't given the space to do things slower themselves and develop into seniors.

[TorqueDog]

Originally Posted by Shazam

Co-pilot is utterly useless. It uses 4o. It has not given me one correct answer.

Claude is better. I'd say about 50% useful. Still gives lots of syntactically wrong code but at least something that resembles something I can use.

Here's the problem though: You have to be a very, very, very good programmer to know when it is wrong. There is a mid-level developer (I wouldn't even call him that) on my project, he uses AI all the time, and he always trusts the answers. His code is a total ####ing mess. Often times it doesn't do what they ask, but he can't figure it out, so how the hell would he ever know if the AI did it properly? His part of the project has been on bug fixing for weeks now.

It's good for reference though. Yippee. A fancy SO.

Claude got me started on a Python app I wanted to create, but it was ChatGPT (o3-mini-high) that actually got the damn thing to run properly and generated the desired output.

I wanted a program that would run automatically at 9 AM ET, gather all the post-market activity from the previous day and pre-market activity from the current day, and then plot the price action, high-low, levels of resistance and support, using the 1D, 4H, and 30M, for ES and NQ futures, plus the Magnificent 7. Put them each into their own chart covering just the extended market for that time period.

Claude got it started but kept making the date range too long (plus had some syntax bugs). After a few "fix your code, here's the error" prompts, Claude's version couldn't actually produce anything and would error out when accessing yfinance. So I took the initial "working but flawed" Claude version and put it into ChatGPT and had it slowly correct each bug at a time, plus create a verbose logging mode for better troubleshooting of code issues.

It now works perfectly. I was pretty blown away.

[Russic]

Originally Posted by TorqueDog

Claude got me started on a Python app I wanted to create, but it was ChatGPT (o3-mini-high) that actually got the damn thing to run properly and generated the desired output.

I wanted a program that would run automatically at 9 AM ET, gather all the post-market activity from the previous day and pre-market activity from the current day, and then plot the price action, high-low, levels of resistance and support, using the 1D, 4H, and 30M, for ES and NQ futures, plus the Magnificent 7. Put them each into their own chart covering just the extended market for that time period.

Claude got it started but kept making the date range too long (plus had some syntax bugs). After a few "fix your code, here's the error" prompts, Claude's version couldn't actually produce anything and would error out when accessing yfinance. So I took the initial "working but flawed" Claude version and put it into ChatGPT and had it slowly correct each bug at a time, plus create a verbose logging mode for better troubleshooting of code issues.

It now works perfectly. I was pretty blown away.

This has become my go-to fix as well. At some point it seems like most projects run into a wall of loops. I've found handing it off to another model at that point (especially one of the slow, costly, reasoning models) can do wonders.

[Itse]

Originally Posted by Firebot

https://twitter.com/user/status/1905204774191890837



https://www.reddit.com/r/marketing/c...io_ghibli_art/

Nothing to worry about.

This is just sick.

Gods I hate modernity.