Most often prejudices are laid down

[sadiksojib35]

Then the neural network starts to behave like a person who joined someone else's discussion somewhere in the middle and speaks out of turn. But they are trying to combat this. For example, Claude and ChatGPT 4 already acc india telegram database ept 100 thousand tokens. Perhaps in 2024 there will be neural networks capable of processing 200 thousand or a million tokens.

Another disadvantage of neural networks is the reflection of human prejudices. accidentally. A recent study conducted in the United States showed that neural networks can often generate racist or sexist results if they were trained on a data array that already had some bias in one direction or another. For example, the Midjourney neural network can return an image of a black criminal in white clothes when asked for the query “white male robber”.

At this stage, neural networks cannot work independently. They need human control. Like the autopilot in Tesla - the system controls the car, but the driver must keep his hands on the steering wheel.

That is why it is unlikely that the world will be taken over by the profession of "promt engineer" or that neural networks will completely replace specialists. A promt engineer must not only understand how to write promt, but also understand the subject area very well. If a non-expert turns to a neural network for help, he will hardly be able to filter out its hallucinations - they can be extremely plausible.