Let me give you an example

[sadiksojib35]

The main problem with neural networks is hallucinations and over-performance. The neural network is unable to distinguish truth from lies and will not cope with the search for a specific answer or fact on its own. It is critically important for the network to give an answer. Wrong? It does not matter. The main thing is that the user is satisfied.


Why did we get such answers? When a neural network is trained, terabytes of data are run through its neurons, which are converted into neuron weights. In other words, the data is not saved, but transformed into weight numbers, with the help of which the neural network will then generate token after token. The information is transformed irrev india telegram database ersibly. There is no way to restore it to its original form.

The second problem is that neural networks are not profitable to use in tasks that are faster to do "by hand". Sometimes writing a good prompt takes more time than solving it yourself. Also, neural networks are not very helpful with complex multi-step tasks that first have to be divided into simple steps, and then you have to ask yourself again whether it is faster to do everything manually.

A neural network is a tool of limited applicability. A neural network is like a hammer: it is great for hammering nails, but not so great for tightening bolts.

The third problem is the limit of short-term memory, or forgetfulness. All text in the neural network is converted into tokens, and these tokens have a limit. If you have a dialogue with the neural network, then with each new message from you, it must “remember” (i.e. run through itself) the entire history of correspondence. But if you go beyond the token limit, then it has no choice but to delete the oldest messages. And it is there, most likely, that the description of the original request was. That is, after some time of conversation, the neural network simply forgets how it all began.