The ChatGPT is garbage when you ask it anything that is not videly available on the internet. It outputs a very verbose and superficial answer. Usually with many mistakes.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Maybe you refer to the way they are trained, i.e. predicting next word in a sentence.
But training system and final result are 2 different concepts.
If you see a boxer jumping the rope and you know nothing about boxing, maybe you think that he trains to jump ropes....instead he trains to punch people.
At the moment we have still limited deep understanding of these models, and even less of how our brain works....maybe one day scientists will discover that also our brain works as a very sophisticated stochastic parrot...who knows?
This is not the case. It is already verified they build an internal representation of the world called "latent space", and generate text / images / videos out of this latent space. If you google for it, there is plenty of source out there.
broadsword
Brigadier
You mean they were not generated from a vast cloud library of images? If every pixel was machine-generated, the amount of GPUs must be incredible.
That is the case with all generative models and no, simple generative models could be run even on your CPU - output resolution and overall image quality would be crap obviously but every pixel would be machine-generated. No neural network works as some kind of matching platform that picks out "similar" images out of some stored library - this is a pretty severe misunderstanding of how NNs operate. In fact, it is hard for the network to even produce identity of its input if there are no global skip connections (i.e. adding the input after the whole network) in the architecture. You could argue that the latent space is a compressed cloud library but in that case human mind can also be interpreted as a compressed library since we are not born with the knowledge we have today, our mind, much like the latent space of the neural network, is formed through the learning process which is constant for humans.
However, it is important not to overrate modern NNs and put them on the same level as a human brain - NNs in their current state are crude simplified mathematical models of our brains. Their latent space is far more limited in real capacity (it's true that some models have trillions of parameters but when you dig further, you will find that these models are extremely sparse in their representation - i.e. only a small portion of those trillions carries important information) and the current learning methods are extremely inefficient - they need billions upon billions of examples sampled from different distributions/domains in order to learn how to generalize the training information, which leads to the problem of "overfitting" where a dataset without enough data diversity leads to model merely learning how to perform well on the data from the same distribution as its training data. Humans, on the other hand, have a "few shot" learning where only a couple of examples could already create some basic understanding of the task.
Overall, I'd say modern NNs are very convenient and powerful tools that can be used to further increase human productivity rather than real "intelligence".
Last edited:
How good is Ernie bot compared to ChatGPT 4.0?
Is China really lagging behind the US in AI as western media keeps producing articles about this? How big is the gap and how fast can China reduce the gap?
Is China really lagging behind the US in AI as western media keeps producing articles about this? How big is the gap and how fast can China reduce the gap?
From my casual observations, it seems that research on AI is quickly accelerating in the US. These are my rankings
LLM Ranking:
- OpenAI
- Everyone else
- OpenAI
- Everyone else
Areas to watch for in 2024:
- World Simulation
- Agents
- Robotics
- Context length
- Multimodalities
Last edited:
let's see what the end result is.
in China there's medical AI that aids radiography diagnostics.
It has already been put into field usage during COVID.
so what is more valuable, unemploying a few thousand fake news journalists with mediocre quality but almost 0 cost fake news, or saving lives?
Western media also reports little about Chinese AI, period. Comparing the amount of articles being generated about the likes of OpenAI, Google etc. there's extremely little presence of Chinese AI in English language journalism.
Last edited:
New Sora videos dropped. Quite a few things need fixing but you can see the potential