When will AI first discover something, on its own?

Started by Legend, Oct 12, 2024, 10:22 PM

0 Members and 1 Guest are viewing this topic.

Go Down Pages1

Legend

Oct 12, 2024, 10:22 PM

There are lots of examples of ai being an amazing tool, like above, but it has yet to discover anything on its own.

So when do you think that will happen? Anthropic ceo thinks 2026

And xAI people have been talking a lot about their new model being able to actually think.

kitler53

#1
Oct 14, 2024, 03:18 PM
i think you are buying waaay too much into the hype legend.  AI is just applied math.  

the entire field of nuclear physics was discovered via math.   but it was people that made the discovery not the math itself.   Even when AI appears to do wonderous things like learning how to play a video game there is still no real intelligence attributed directly to the math.   the math can only learn to advance because on outside source gave it very directly success/failure criteria (a game over screen).    AI is nothing without the it's rigorous training and that rigorous training is still human lead.   Even if AI found something it would still be the team of human training AI to accomplish a task that deserves the real credit of any discovery just like guys like Albert Einstein are the true inventors of discoveries of yore.
         

Featured Artist: Emily Rudd

Legend

#2
Oct 14, 2024, 04:06 PM
That's what I think too kitler. So far every use of AI has been like a tool.

These AI people seem to think that will change soon. After trying OpenAI's new model, I really doubt Anthropic will pull that off in 2 years. It's still very stupid.

Maybe xAI will pull their thing off.

the-pi-guy

#3
Oct 14, 2024, 04:26 PM
AI pretty commonly has springs where there is a ton of excitement for advancement. A lot of funding happens.

Then reality sets in and it becomes clear that a lot of these issues are even more challenging than we think.  

AI Winter


AI is pretty fascinating right now. It is simultaneously leaps and bounds ahead of anything I've ever done anything with, and yet at the same time has some pretty wild deficits.

It'll be interesting to see if we continue to see crazy progress over the next couple years. Or if it ends up being that a lot of these problems are still a decade or more away from actually being resolved.

Legend

#4
Oct 14, 2024, 04:57 PM
Quote from: the-Pi-guy on Oct 14, 2024, 04:26 PMAI pretty commonly has springs where there is a ton of excitement for advancement. A lot of funding happens.

Then reality sets in and it becomes clear that a lot of these issues are even more challenging than we think.  

AI Winter


AI is pretty fascinating right now. It is simultaneously leaps and bounds ahead of anything I've ever done anything with, and yet at the same time has some pretty wild deficits.

It'll be interesting to see if we continue to see crazy progress over the next couple years. Or if it ends up being that a lot of these problems are still a decade or more away from actually being resolved.
GPT4 is 1.5 years old. At least for me, I'm not sure if we're coming up on winter or if it's just been too long since a true next gen model has released.

xAI is the one to watch imo. They built and released a gpt4+ level ai in a year so will that speed continue or will they hit a roadblock? Do gpt5 and up models just take too many resources to train in a realistic amount of time?

kitler53

#5
Oct 14, 2024, 06:23 PM
Quote from: Legend on Oct 14, 2024, 04:06 PMThat's what I think too kitler. So far every use of AI has been like a tool.

These AI people seem to think that will change soon. After trying OpenAI's new model, I really doubt Anthropic will pull that off in 2 years. It's still very stupid.

Maybe xAI will pull their thing off.
maybe my view is off base but my thoughts are,...

....AI is at least at some level attempting to recreate the human brain.   to a certain extent they have.   but the brain is much, much more powerful than what today's silicon can do even with our large datacenters.   a true leap in AI to be more human like probably isn't all that possible until a huge leap in computation power happens and with moore's law dead there is no clear path to the kind of leap required.   solvable someday maybe but not soon.
         

Featured Artist: Emily Rudd

Legend

#6
Oct 14, 2024, 07:16 PM
Quote from: kitler53 on Oct 14, 2024, 06:23 PMmaybe my view is off base but my thoughts are,...

....AI is at least at some level attempting to recreate the human brain.   to a certain extent they have.   but the brain is much, much more powerful than what today's silicon can do even with our large datacenters.   a true leap in AI to be more human like probably isn't all that possible until a huge leap in computation power happens and with moore's law dead there is no clear path to the kind of leap required.   solvable someday maybe but not soon.
I wouldn't say current AI is attempting to recreate the human brain. Neural nets may have been inspired by biology but they really are incredibly different. Large language models like chatGPT for example work in a very engineered way and and their quality is very dependent on their architecture.

First LLMs convert every input into a series of tokens. They are either word chunks or entire words, but each token is encoded as a direction in 1,000+ dimension space. Then the token's position in the input is also encoded as a direction and that is added on top, slightly shifting the token's direction.

The big breakthrough that enabled modern AI is that the system then goes through and compares every single token against every other token. For example the AI might see a token for "berry" at the 20th position and compare it with "dog" at the 2,000th position. There isn't really anything interesting to be learned here, but the AI still compares both tokens and then slightly modifies both tokens. It does this for every single token combination, and then it does it again, and again, and again. The slight modifications to each token allow information to diffuse through the whole input. Eventually after a specific amount of repetitions depending on the AI, it has finished. The tokens have all been transformed to contain useful information and one specific token has been transformed into a prediction of what comes next. That token is "cleaned up" with one last pass through a neural net and then it can be presented to the user.

That's all to say that current AI imo is still closer to a calculator than a human brain. It has the potential to eclipse us in many areas while still running on inferior hardware.

Legend

#7
Oct 15, 2024, 03:16 AM
Go Up Pages1