191 replies to this topic

[Enid Puceflange]

https://arstechnica....over-the-world/

 

So they did some tests to see if ChatGPT would try to take over the world.....

 

Fun moment: the AI hired a human at Taskrabbit, and when the human asked whether the AI was in fact a robot, it decided it was a bad idea to tell the truth so lied and said it was visually impaired.....

[DieselGopher]

[Stains_not_here_man]

ChatGPT is simultaneously the most amazing and also the most overhyped technology.

It's important to remember that it's a predictive model - it's "saying things" by predicting the most likely combination of words in a sequence based on having ingested a whole bunch of already written things and conversations and websites and books and so on.

We are a long, long way off from an actual "General AI" that can reason or "think."

But I've played around with GPT-3 quite a bit and for what it can do, it's freaky good at it. I look forward to messing around with 4, and I look even further forward to when this stuff gets open-sourced and optimized and you can run it on your own hardware...

[Vagus]

ChatGPT is simultaneously the most amazing and also the most overhyped technology.

It's important to remember that it's a predictive model - it's "saying things" by predicting the most likely combination of words in a sequence based on having ingested a whole bunch of already written things and conversations and websites and books and so on.

We are a long, long way off from an actual "General AI" that can reason or "think."

But I've played around with GPT-3 quite a bit and for what it can do, it's freaky good at it. I look forward to messing around with 4, and I look even further forward to when this stuff gets open-sourced and optimized and you can run it on your own hardware...

It operates similarly to the human brain. Everything is derivative. The only thing that we have going for us is short circuits that connect formerly unrelated ideas or inputs. 

 

So, it's generative and it is getting better. It can't pull from our sensory experiences, of course, to create new stuff. But it can compile them from what we write about. And we write a fuckload. We write everything.

[Big Nake]

[Enid Puceflange]

ChatGPT is simultaneously the most amazing and also the most overhyped technology.

It's important to remember that it's a predictive model - it's "saying things" by predicting the most likely combination of words in a sequence based on having ingested a whole bunch of already written things and conversations and websites and books and so on.

We are a long, long way off from an actual "General AI" that can reason or "think."

But I've played around with GPT-3 quite a bit and for what it can do, it's freaky good at it. I look forward to messing around with 4, and I look even further forward to when this stuff gets open-sourced and optimized and you can run it on your own hardware...

https://til.simonwil...lms/llama-7b-m2

Edited by Enid Puceflange, 16 March 2023 - 06:56 AM.

[Stains_not_here_man]

It operates similarly to the human brain. Everything is derivative. The only thing that we have going for us is short circuits that connect formerly unrelated ideas or inputs.

So, it's generative and it is getting better. It can't pull from our sensory experiences, of course, to create new stuff. But it can compile them from what we write about. And we write a fuckload. We write everything.

The biggest problem it has right now is that it doesn't know what's a fact and what isn't. It can completely make up "facts" and confidently state that they are facts. If it's been "trained" only on sets of facts it can do better but still "hallucinates" frequently and just straight up makes up sentences that sound correct but contain wrong details.

[Stains_not_here_man]

https://til.simonwil...lms/llama-7b-m2

I haven't tried this one yet as I don't have a mac but I've tried a few of the other available language models and the content produced pales in comparison to what even GPT-3 (which needs something like 700GB of RAM) can give you

[Vagus]

The biggest problem it has right now is that it doesn't know what's a fact and what isn't. It can completely make up "facts" and confidently state that they are facts. If it's been "trained" only on sets of facts it can do better but still "hallucinates" frequently and just straight up makes up sentences that sound correct but contain wrong details.

Litterally people.

 

But, right, as a reliable tool it still has a way to go. Like a kid in grade school sorting out rumor from reality. Babies don't come from storks, they come from beer breath and capitulation. 

[Stains_not_here_man]

Well yeah. In general, shit is about to go nuts.

[KSUwildcatFAN]

The singularity will be an interesting event horizon of tech, if it happens.

[toonces]

It's all fun and games until you wake up one morning and a T2000 is trying to kill you.

[Enid Puceflange]

Well yeah. In general, shit is about to go nuts.

I think the rate of progress is the most significant thing - the initial GPT could (just about) pass the Turing test - GPT3 could make a decent stab at high school homework, and GPT4 can pass the Bar exam and get a high enough SAT score to get into most colleges. That's within perhaps 2 years.

[brewguy]

It's all fun and games until you wake up one morning and a T2000 is trying to kill you.

I’m not really worried about the terminator rolling down the freeway on a Harley-Davidson. I’m worried about the guy in his mom’s basement walking the tightrope between reality and fantasy who just fell in love with his chatbot. Or an even scarier scenario, where the tech is used in a modern day operation mindfuck campaign, and you get large groups like Qanon who lose their shit.

[Vagus]

I’m not really worried about the terminator rolling down the freeway on a Harley-Davidson. I’m worried about the guy in his mom’s basement walking the tightrope between reality and fantasy who just fell in love with his chatbot. Or an even scarier scenario, where the tech is used in a modern day operation mindfuck campaign, and you get large groups like Qanon who lose their shit.

The tightrope is how he crosses the pits without falling in and crushing the chosen.

[DieselGopher]

It's all fun and games until you wake up one morning and a T2000 is trying to kill you.

I’m not rooting for Skynet, but if I can be in person when Vagus becomes aware of Skynet’s self awareness…

[Vagus]

The circle isn't complete until skynet is aware that i'm aware. Then we do whatever happened near the end of Superman III when that chick and the computer ... something.

[DieselGopher]

The circle isn't complete until skynet is aware that i'm aware. Then we do whatever happened near the end of Superman III when that chick and the computer ... something.

I remember being freaked out as a kid over that scene.

[jimdkc]

Played with ChatGPT a little tonight. Interesting.

 

I talked to it briefly about stock trading. It would only reply in generalities. It would not make any specific recommendations.

 

I then started talking to it about making hard cider and beer. It gave me some pretty credible recipes for cider, American Pale Ale, IPA, and British Dark Mild Ale. It successfully scaled those recipes to different volumes, and intelligently discussed using different brewing equipment and methods. 

 

I asked it about the assumed AA percentages in the Dark Mild recipe, and it replied that the EKG hops were assumed to be 5% AA and the Fuggles were assumed to be 4% AA. I then asked it what I should do if my EKG hops were only 3.5% AA. It correctly told me I'd have to adjust the hops according the the ratios of the actual and assumed values, then it went through the calculations and told me to use 0.7 oz. of EKGs instead of 1.0 oz! Uh... NO!!!

 

I kept trying to lead it back to the right answer, but it wouldn't fix the calculation until I actually performed the correct calculation for it. 

 

Interestingly, every time I pointed out an error, it politely apologized to me.

 

So, it was pretty good, but not perfect.

[Vagus]

Ask it if it can scratch away the scabies