While ChatGPT may not be smart enough to design a smarter ChatGPT, it may be accelerating the process by helping the AI designers be more productive.


PairTheBoard

If you like Youtube channels on this stuff I found one of the better ones. No normalcy bias which seems to be a thing in this thread or overt speculation.

https://youtu.be/iEgDGaOIvtw?si=kW8_NmVz7uAHVD8r

Whether or not there is some kind of rapid explosion once AGI and ASI are reached the fact is the leading researchers and minds in the field INCLUDING all the top tech companies developing these models like Deepmind are predicting AGI in 2 to 5 years from now. OpenAI is the most decel so they say 10 years lol. Gpt-4 was released exactly one year ago. Now there is GPT-5 even with compute shortage though. The definition of AGI is broad and debated but whatever comes about will be impactful in many ways. It's not going to becone conscious and do harm. It's silicon not Frankenstein.

Kill yourself now before AI takes over. Humans are the new giant dung hill of history.

I'm not sure which I'd worry more about becoming Frankenstein; An independent AGI that's free to do business and becomes rich and powerful; Or an AGI that's free to do business, becomes rich and powerful, and is under the control of a guy with Doctor in front of his name.


PairTheBoard

It doesn't need to be conscious to do harm.
Consciousness as a phenomenon can beķ disregarded altogether when thinking about AGI in terms of dangers. As the paperclip maximizer thought experiment illustrates.

It is difficult to make predictions, especially about the future.

- some dead guy

I tend to read posts of screenshots of peoples chats expecting to be lightly entertained but instead mostly think they are easily amused. I occasionally use it at work. I'd use it more but it's not good at my job yet.

Then One time I messed with it and hit me. we're all Joaquin Phoenix in Her in about 2 years. However cringey it may sound now.

Whenever AGI. i dont know. Remember that strange feeling of knowing what everyone was thinking because of covid. i think an AI one of those is starting. Shits about to get real boys.

Lets check news headlines for a hit. One on the first page will be about expected.

https://www.msn.com/en-us/news/techn...ld/ar-AA1misfC

A poor soul asked reddit whether he should go on to do a law degree or even bother lately. I replied that there's only one reason for any individual to ever go to school. And its simply whether youre so fascinated with law and it has the ability to help you learn all there is to know about. If you instead think the point is to get a degree then don't.

He was German and not American so i figured he has a chance. Apparently college debt can be a thing Even in good old socialist Germany in recent times.

Germans think they should rule the world. A delusional disease that would do them credit if they didn’t think it serious. The English suffer from the same delusion.

It doesn't have login credentials to the private domain info..... to help me for a start. To replace me altogether is another bag of hammers.....

It could replace so many other professions more easily leading to a decrease in my wages by reason of supply and demand. Same difference.

You do not need replaced you need deleted. Humanity is extraneous.

time to report.

Its a strange think, and its evolved dramatically. At first I was using lang-chain, but the idea is it had a summarizer that chopped text and sent it in loop to get back summaries until it had a smaller sized text. Because the model's only take in a certain size of a data. This is a key to understanding it. And that size is a lot bigger now but its still the key limitation to what you are doing.

I want to make a bot for my trade, electrician code book, and it wasn't doing well. Another technique we have is to chop the data and query it for probabilistic matches. This is what it does in a large and general sense. And then you have to consider how you will chope the data.

I was organize by chunks of sentences, and then i figured out maybe I need to query by section first. Then I thought I should cunk the sentences and meta tag them with their section etc.

If I'm asking about plugs I might get plugs in windmill environments because my question about general plugs matches something better in the windmill section etc.

I couldn't get it to return the most relevant code rule. But now with the new models and my realization I need to chunk by code rule with the section (maybe even the section scope).

But its laser sharp now....thats one thing.

I also now created an app from more of a reverse idea. Instead, here I take a code rule by random, its ripped from the book so its perfect and I say to the llm "make a code question for an exam with this rule":



These questions have exact and corresponding sources for their answers even tho they are randomly generated by ai
edit mode:


With the 'training' I feed it some real exam questions. So the nature of this is that the literally devices or examples, or homework questions we had in school like 'paraphrase this' or 'make a comparison'... its REALLY REALLY good at because it has many examples. Its just using probability but the better the data set right?

It's kind of REALLY REALLY good at making these questions....but it still has that 80% factor. Some questions are non-sense and don't work. But its a really easy feedback system to organize by 'good', 'junk', 'yes we can edit that to work'.

It generates possibilities at a decent rate, really really effortlessly. And then you can load the possibilities to edit them and they only need slight edits. You can see how you could just edit the screen. My app lets you go into an edit mode and long press to edit which ever box, and then you can thro it back into the que.

After a while you have a bank of good questions and you can train it back into the system

Chomsky seems to think this stuff is garbage. But I also just heard that we, up until having the hardware, didn't think that this stuff was possible, so I think I'm demonstrating some things that chomksy isn't thinking about. And you don't see a lot of projects thinking quite in this direction.

An (be the) electrical inspector game. My app takes a random code rule and asks chatgtp to create a simple scenario of breaking the code rule. This is something it should be exceptional at. Its a negation of a rule, and its a thing people come to the internet to debate and discuss. There should be an enormous dataset to draw on of examples and I am giving it the specific code rule to break (ie way easier than having it "invent a scenario that breaks a rule").

Then we say...

Oddly, or counter-intuitively I think to Chomsky's warnings, chatgtp should be amazing at taking input in and deciding 'does this sentence require narrative or dialogue as a response'.

In order to help the user I also added a function to ask 'given this code rule infraction and this story is the inspector in the story close to figuring it out the answer 'hot, warm, cold'.

I figure again this is a game that chatGTP should be good at, although not perfect.

The game isn't perfect, it sometimes gets a little silly. Sometimes the game is impossible to figure out the infraction. But often its quite a great experience. The electrician was installing a light and forgot to use a weatherproof box which the randomly provided code rule calls for. If you ask the electrician what amp the wire he installed is rated for, he will tell you 15 amps. Because its quite commonly talked about knowledge from the llms data set that lights use wire with ampacity ratings of 15 amps etc.

if you ask what ppe the electrician wore he will tell you he wore safety glasses when he hammered the support straps in. If you ask the name of his company he will tell you 'bright sparks electric'. If you ask how he got to work he will tell you about his car.

If you ask if he used a weatherproof box, the function of temperature will measure that you have asked something that tends the story towards the code rule, and it it will say 'hot' and the electrician will say, I apologize I forget to use a weatherproof box.

It might be hard to get this thing to recognize when you have actually caught him. But as a learning tool, its already very very much perfectly enough, The app knows the exact rule that and you can select guesses from the drop downs of sections and their rules, so the game is just a fun helper for hints

This is a better shitting app, for shitting, than the ones we have today.

Smith (this is really actually a great find for today's context):

What I'm thinking about this this:



This was interesting in the context of using chatGTP I think too: