For a long time, specialists have worked without care in the world. They developed construction software, and published them smoothly. Then he struck the era of isolation, and suddenly, they felt bored (of course, this is fun in actual events). People wanted to create something that could deal with their work while staying at home: answering routine questions, creating cold personal photos, and analyzing huge amounts of data in minutes. They dreamed of traveling to a wonderful place, and so on, I guessed that, they revolutionized artificial intelligence.
\ AI now works, provides answers and improves lives. As a skilled assistant as it is, artificial intelligence is really effective only when used in the right context.
\ We are witnessing rapid progress in artificial intelligence applications, from generating images and videos to predicting the stock market and analyzing encrypted currency. yet, Artificial intelligence may provide information that we do not request Or provide wrong answers in starkly. His behavior is very similar to the behavior of home cats – as you know, the type that sits quietly and then decreases on you suddenly?
\
\ Our cats, as well as artificial intelligence, enjoy unpredictable:
\
- You give them the same food (or data) – sometimes they eat, and sometimes they ignore it.
- You train them to respond, but sometimes they interact when contacting them.
- The greater, wild, or larger the model of artificial intelligence, the more difficult it is to predict its behavior.
- In the morning, cats may be calm; By evening, they turn over activity (just like dynamic data).
- Cats may be friendly (inevitable), but they can scratch you without warning (random).
\ You may wonder about the inevitable and attached meaning – let’s discover.
Inevitation and stocag
The inevitable system always produces the same result by looking at the same inputs – Think Expression If you Davus engineer. An example of the real world is your cat that eats the same amount of food that you put in a bowl every time-this Inevitability. But when the cat is inhaled and only half, it is no longer decisive.
\
A / a Randomness The process includes an element of randomness: with the same inputs, the result can vary. For example, the ATM model often uses random algorithms, such as Randomness SGD decliningWhich is trained by the form by choosing random pieces of data instead of the entire data set.
\ These definitions do not fully explain the reason for the presence of AIS we have sometimes and behavior. There are other contributing factors, including the following:
\
- Inevitability
- Randomness
- Errors ’errors and the floating point account
- Multiple and parallel reading operations
- Update data constantly
- Chaos and “butterfly effect”
\ If we look at a little closer, we will see other mechanisms that affect the unpredictable behavior of artificial intelligence models.
A glimpse of nerve networks
You may know that AIS uses everyone Dependence on various neurological network algorithms. Here are some types of nerve networks:
\
- Fully connected nerve networks (FCNN): Classic structure, where each neuron connects to each nerve cell in the next layer.
\
- CNNS: CNNS: These networks use notes or filters that highlight images features such as edges, textures and shapes.
\
- Repeated nerve networks (RNNS): These networks contain observations episodes that allow them to remember the previous steps (that is, they remember the sequences).
\
- Long -term long -term memory (LSTM): Augmented version of RNNS with mechanisms to forget important data and remember them selectively.
\
- transformers: The most powerful category for the processing of the text. They use multi -head attention, allowing them to consider the entire context simultaneously.
\
- Gans: Gans: It consists of two networks, one of which creates data and the other establishes its quality. Their competitor leads to better results.
\
- Car tools: Information designed networks (cipher) networks and then rebuild (decoder).
\
- GNNS: GNNS: They work with charts (contract and edges) instead of regular data.
\ We need all this context to understand why it is the most common model, ChatGPT, often cheering.
How Hallucinations happen Amnesty International
Chatgpt works on adapter Architecture, was presented for the first time in Paper 2017, “attention is all you need.” This is the same mechanism that revolutionized the text processing. Transformers work on the autonomy mechanism, which allows them to consider the global context instead of just the closest words such as old repeated nerve networks (LSTM and GRU). The model belongs to the GPT series (pre -training transformer), which means:
\
- Prepared by: It was initially trained in huge amounts of text (books, articles, websites and symbols).
- generative: Its mission is to create a text, not only classifying or extracting facts.
Chatgpt answers are caused by a random process instead of a rigid base. It does not memorize or reproduce texts but generates responses using a likely model.
Word predictions as a possibility
When ChatGPT responds, it does not choose the correct individual word but it calculates the distribution of the possibility.
\
P (wi | W1, W2, …, Wi-1), where:
- “Wi” – the following word in the sentence
W1, W2, …, Wi-1- Previous words
- P (wi | W1, …, Wi-1)-The possibility of “Wi” is the following word
\ For example, if you ask, “What is today?” Chatgpt may have different possibilities:
\
- “Monday” – p = 0.7
- “Wed” – p = 0.2
- “42” – p = 0.0001
\ You often choose the word with the highest possibility, but due to the temperature of the generation (a teacher who controls randomness), you may sometimes choose a less likely option based on the context.
The impact of context and forgetting information
Chatgpt works with a limited context window, which means that it “remembers” the recent symbols NN. For GPT-4, the context window is about 128 kilos (about 300 pages of the text). If the important information is outside this context, then::
\
- Forget the details (the effect of context cutting)
- Make -up information (Stochastic)
\ However, ChatGPT can often correct his answer after you ask if it’s sure. However, Chatgpt can often correct her answer if you ask if it’s sure.
Amnesty International sometimes corrects itself, but why?
When you ask Chatgpt, “Are you sure?” It is re -analyzing his answer using a new context where there is doubt. This results in:
\
- Repeat the possibilities of answer.
- Choose a more logical option if there is one.
\ This process can be explained by the possibility of Baysi.
\
P (a | b) = p (b | a) p (a) / p (b), where:
\
-
P (a | b)-the possibility of answering a correct, taking into account the follow-up question b.
-
P (B | A) – Possibility to ask if Chatgpt is right at first.
-
P (a) – The initial possibility of answering ChatGPT.
-
P (b) – The general probability you will ask.
\
A lot of information for you? High brain temperature? Imagine that AIS also overcomes large amounts of information.
Mistakes caused by loud and loud data
Huge quantities of text data flow into ChatGPT training, including noise or contradictory information, such as:
\
- Some sources say that the Earth is round, while others claim it is flat.
\
- AI cannot always determine the correct information when it appears with varying possibilities.
\
\ These are examples of typical hallucinations, which occur because Chatgpt weights are trained in probability words instead of strict logic.
The bottom line
This is what we can learn from this. Halosa has been seen since then:
\
-
It predicts a possibility, not inevitable.
-
It has a limited memory (context window).
-
Repeat the possibilities when interrogating it.
-
It has training data that includes noise and contradictions.
\
This is clear and direct. I hope you don’t get tired. If you do that, this is a good sign because it means that you are thinking critically, and this is exactly what we should do when working with artificial intelligence.
