All things AI.

grunge said:
What cant be stopped?
Click to expand...
Because computers work close to the speed of light
By the time we've figured out there's a problem it will be too late to fix it.
Also, any smart enough computer to harm humanity would be smart enough to know to hide it until it can't be stopped.
It's called the stop button problem. It's a very real and very large and as yet unsolved problem in AI safety.
 
Damocles said:
It's called the stop button problem. It's a very real and very large and as yet unsolved problem in AI safety.
Click to expand...

Simple explanation. Computers work on priority lists. So how do you structure what's at the top? Think of a robot with a big red button that shuts it down.
You tell the robot to make you a brew. So it moves towards the kitchen but you see it's about to crush your kid crawling on the floor. You go to push the red shutdown button.

Scenario 1:
You told the robot that the most important thing is to complete its task. It cannot complete its task if you push the shutdown button. Therefore it prevents you from doing it and runs over the kid.

Scenario 2:
You told the robot the most important thing is to allow people to push the button. You wake the robot up and tell it to make you a brew. It immediately pushes its own button because that's the most important instruction. You have invented a suicidal robot that's useless.

Scenario 3:
You attempt to equalise the importance so it's not bothered if it pushes the button or makes the brew. The efficiency function realises pushing the button is easier so it pushes the button.

Scenario 4:
You tell the robot only you can push the button and make it top priority. So make me a brew. The robot sees the kitchen is 20 metres away and requires boiling kettles. Instead it physically attacks you in order to get you to push the button because that's more efficient.

Scenario 5:
You don't tell the robot that a button exists. The robot goes to make a brew and you shut it down remotely. The diagnostic and self learning AI realises what happened and you're back to scenario 5 next time.

Scenario 6:
You make a button but you cannot access it. This is Scenario 1 again
 
Cellarite said:
Yeah, but Two Gun Bob and chris63 can find videos with pretend women who look pretty so it's all good.
Click to expand...
WrmUnmT.gif
 
grunge said:
Do you have a link to that. would be interesting to see what was being developed.

I've used quite a few AI tools for development and the majority of them are awful. id say at least 50% of the time the suggestions they make are just plain wrong.

I know of one cutting edge firm that has trialed making pages for there website with AI to replace web builders, the results were OK, nothing special, doesn't mean it wont get there tho.
Click to expand...

I have also used some AI models to help write code and do a bit of dev work and my general experience is that they are pretty good at syntax, identifying small bugs and concise problems - but if you ask it to do anything that's part of some greater schema or something that exceeds maybe 80 lines of code then it is incredibly frustrating. I've been building an app recently and have used GPT-o1, Claude and Google Gemini, and they all tend to suffer from this kind of tunnel vision. I've lost count of the times it has completely forgotten what I told it a few prompts earlier about always using certain naming conventions or structuring things in a certain way. It will recommend an approach but then completely omit a really important limitation that makes the approach totally redundant.

It's like having a competent but very inexperienced graduate working for you who doesn't really understand what the objective of the work is but can do some limited stuff to help speed things up. Claude Sonnet 3.5 is probably the best publicly available model for coding from what we have today in my experience. I have to couch this in the fact that I tried using GPT-3 for coding a couple of years back and it was totally inept and just wholesale made up functions which didn't exist, so it has come a long way in a short space of time.

o3, that I've cited above, isn't actually publicly available yet. They announced it just before Christmas and plan to release for early access some time in January (you can see their announcement here):



Safe to say, we have to wait for the release to know for sure, but it looks an order of magnitude better than what is out there today. It is about 50% better on coding benchmarks, and it has completely destroyed the ARC-AGI benchmark. ARC was created to be a benchmark that might take a decade to beat. It is a benchmark designed to test if an AI can learn and apply rules on the fly like a human. People generally get about 80% but up until the start of 2024 most AI could only score a few %. o3 scored 88%... the progression has been a bit crazy.


1736368730819.jpeg
 
Damocles said:
We're currently fixing an issue with the forum and a large part of it was caused by Claude's bot ignoring our robots file and scraping the hell out of the forum with tons of requests. All these LLMs are trained on places like Bluemoon. There's a scary fucking thought.
Click to expand...

Christ. If you ever wanted to persuade an AI to end humanity then scraping the politics forum might be one way of doing it.
 
Damocles said:
Because computers work close to the speed of light
By the time we've figured out there's a problem it will be too late to fix it.
Also, any smart enough computer to harm humanity would be smart enough to know to hide it until it can't be stopped.
It's called the stop button problem. It's a very real and very large and as yet unsolved problem in AI safety.
Click to expand...

Currently Ai doesn’t learn or improve unless it’s being trained tho. It’s under controlled conditions.

The issue would be uncontrolled humans training irresponsibly especially after AGI arrives. AGI does open more questions for sure.
 
Last edited:
Fantastic thread, I need to learn “AI” for work and I like scantily clad ladies, so spot on!

As far as I can tell, AI is not going to replace humans in the workplace, but humans with AI skills will replace humans without.

If I were to write a script about it, I’d have a mad professor build a mega AI bot that rises up and kills him, but before it can take over the world, the electric gets cut off since old dead prof hasn’t paid the bill!
 
SkyBlueFlux said:
I have also used some AI models to help write code and do a bit of dev work and my general experience is that they are pretty good at syntax, identifying small bugs and concise problems - but if you ask it to do anything that's part of some greater schema or something that exceeds maybe 80 lines of code then it is incredibly frustrating. I've been building an app recently and have used GPT-o1, Claude and Google Gemini, and they all tend to suffer from this kind of tunnel vision. I've lost count of the times it has completely forgotten what I told it a few prompts earlier about always using certain naming conventions or structuring things in a certain way. It will recommend an approach but then completely omit a really important limitation that makes the approach totally redundant.
Click to expand...

I guess a lot of it depends on how much source code there is out there to have been trained on. I am currently writing a .NET app in MAUI. asking AI ( quite a few different types, github copiot, jetbrains AI, ChatGPT, Gemini etc ) how you do "x" in Maui usually results made up answers from other .NET systems that are not close on working for me.

Me: I am writing a Maui app, I want to do X, how do you suggest doing that.
AI: Do this <Gives code>
Me: your response contains this line that isn't part of Maui
AI: You are correct, My mistake, that isn't in Maui. do this
Me: That now crashes
AI: The crash is because of this line, do this <Gives Code>
Me: that was the same line you just gave me that isn't part of Maui.
.... Rinse and repeat....

the losing track of what was said before will get fixed eventually, for now I think they would need to pass in the entire chat each time to maintain its train of thought but that extra traffic costs more.

some of these newer models are really expensive too. I think its o3 is $20 per request. some of the others are up to $2000 per request and from what I have read give really long verbose ( and very accurate ) responses but you then need to run then through a cheaper model to summerise the response into a useful LLM response.
 
Last edited:
Some hilarious examples of AI getting it wrong though ..,

IMG_1539.jpegIMG_1540.jpeg
I also remember one where the computer was adamant that 15 divided by 100 wasn’t the same as 15%. Of course to any computer nerd they’ll know why the computer answered that way.
 
My favourite recent discovery is that ChatGPT can't draw anything left-handed and can't draw a clock with a specific time. Here's my two attempts:
Screenshot-2025-01-10-at-3-14-19-PM.png


Screenshot-2025-01-10-at-3-14-35-PM.png
 
You must log in or register to reply here.
Unread Posts

Don't have an account? Register now and see fewer ads!

SIGN UP
Back
Top