The AI dooooooom thread

[Josquius]

I use AI just goofing around, mostly preparing RPG stuff for my games with my son, and it is rather shocking just how bad it is at keeping track of simple details and mixing things up that no human would ever mix up. It requires tons of handholding just in the silly things I use it for. I often think of that person wanting to generate an image of Jesus flipping over the tables of the money changers and for the AI to produce an image of gymnast Jesus doing backflips over tables. The money changers would have been impressed.

It can also be shockingly great at things. But fuck I would be super paranoid about using it in a professional or serious capacity.

Yep. Last month I gave a chat gpt rpg someone had put together a whirl.
It was super good. Really good fun. Except.... It would always forget things. Character you haven't mentioned for a few passages? Oh she's actually called this different name and has a different personality.
That quest you already completed? Here's something very similar indeed. You haven't done this yet.
The maker is aware of this and has interesting ideas to fix it, trying to combine AI with a database. But seems it'll be very load intensive. As an approach has potential though.

Oh. And of course other issues are the genericness of some content. Apparently a known issue that every female character has one of 4 names.

I do use AI at work incidentally. It's good for helping to reframe things.
And of course transcripts are very useful.
The key is to just take what the AI gives as suggestions and not something to blindly deliver.

[The Brain]

Exactly. When I read news stories about companies including AI hallucinations in reports, the shocking thing to me isn't that they used AI but that they released a report without reviewing it.

[Crazy_Ivan80]

Exactly. When I read news stories about companies including AI hallucinations in reports, the shocking thing to me isn't that they used AI but that they released a report without reviewing it.

The rector of the university of Ghent found out that using Ai to write speeches might not have been the smartest thing to do.

[Syt]

Sorry, forgot the second part:



Now, there's going off the rails and there's letting people create sexualized deep fakes of kids. Which Musk addresses by limiting access for non-paying customers.

[Jacob]

I want to selectively shut down the social media platforms of the people who want to overthrow liberal democracy and destroy the EU.

[The Minsky Moment]

I'm up for shutting down all social media platforms.  No favoritism there.

[crazy canuck]

I'm up for shutting down all social media platforms.  No favoritism there.

[Razgovory]

Even Languish?

[Valmy]

Even Languish?

Especially Languish!

Are message boards social media platforms?  1990s Social Media.

[garbon]

I'm up for shutting down all social media platforms.  No favoritism there.

[Tonitrus]

My personal experience with AI (granted, just the freebie ChatGPT version), is pretty much just using it to help troubleshoot some light Rimworld modding/tweaking I've been doing.  My conclusion:

- Essentially like the worst employee/assistant one could imagine. 
- Does the tasks, but when caught out in mistakes, will either lie about them (hallucinate), make broad assumptions...and when called out/challenged, will apologizing, admit the error, and then give a pretty good explanation of how it will correct itself...and then fail again immediately or fail to follow its own criteria.
- With minor image generation, will again make wrong assumptions, and also when challenged, seems to fully understand the instructions, but then ignore them when making a correction.  When challenged on how it ignored its own criteria...it blames the "system". 

[Jacob]

Lol - my wife used the inhouse AI assistant at her work and asked it to polish an email it had drafted. It translated it to Polish 

[Sheilbh]

On the other hand the Google Translate AI stuff is a miracle - going to East Asia with a cameraphone to translate and it's like the future

Thought this was really interesting from an academic. Not sure if it's almost positive in a way?
https://hegemon.substack.com/p/the-age-of-academic-slop-is-upon

[DGuller]

I am starting to get a sense with LLMs that improvement curve might be too steep.  It happens often in data science that the initial prototype that you put together half-heartedly blows your socks off, exceeding all your expectations. 

Now your expectations get reset, you think that if a half-assed effort can yield so much, then surely after iterations and refinements you'll improve by leaps and bounds, but then it turns out that your idea can get to good near instantly and get to better never.

[Crazy_Ivan80]

Lol - my wife used the inhouse AI assistant at her work and asked it to polish an email it had drafted. It translated it to Polish 

Glorious.

-----

We've been using antropics Claude intensively the last few months. Even having it write entire features and doing quite well.
Supervision required of course but the results are quite good.

But what a bunch of yes-men those llms are.