Can We Trust Large Language Models

Large language models can sometimes make the same mistakes over and over again. This makes it hard to use them for automation tasks because no one wants to automate processes with unreliable outputs. Today, I'm going to talk about AI hallucination of large language models from a developer's point of view. Also, I will discuss some approaches to dealing with these hallucinations.

Let's start by understanding the meaning of the problem and what hallucination means. AI hallucination is similar to a person seeing something that isn't there. Large language models do something similar: they produce information that isn't based on real data but appears to be real. Even if they don't have enough data, they will always give you an answer. These answers are based on autoaggressive predictions, which means these answers may not meet real-world expectations.

Now let's continue by looking at how large language models (LLMs) work. Large language models are designed to mimic human speech. Basically, they predict the next word in a sequence. They are trained on massive amounts of text data from the internet, books, transcripts, etc. It means they are not magicians; they are just producing human-like text without considering whether it's true or not. This explains why they often get things wrong. Even large research companies like OpenAI or Anthropic have started creating their own synthetic data to train their models because they consumed much of the internet. If this training data contains biased data related to your context, you will likely get biased results. If the model doesn't have enough data about your context, it will probably give incorrect results because, as you know, they always have an answer. However, this doesn't mean they are liars or machines. As some people call them "lying" means knowingly making false statements, but large language models do not care whether they say it's true or not; their main goal is to impress or persuade the user.

However, they can be soft bullshitters. I read an article titled "ChatGPT is a Bullshitter" where the author discussed how they could be soft bullshitters because they don't have any intention about the truth of their statements. This might be a correct concept. So now let's find the starting point of this hallucination problem. When you use ChatGPT or similar models, they tokenize your text into similar units. After tokenization, they identify patterns among these tokens. Essentially, large language models understand your context this way. Finally, they predict the next token in a sequence and repeat this until the output is complete. These outputs may not meet real-world expectations, and we call this AI hallucination. This usually starts when the model doesn't have enough training data or relies too much on incorrect patterns. As I mentioned, large language models always give you an answer even if it is false because they are not aware of it.

If you want to use large language models for automating processes, financial trading, or similar tasks, you need to align this bias or incorrect answers with real-world expectations. So, there are several methods for doing this. Let's look at the main logic of some of them without going into the technical details. The first one is ensemble methods. Ensemble methods are highly preferred when dealing with AI hallucinations. You combine predictions from different models and cross-validate results to eliminate incorrect statements. It's better than relying on just one model. Another good method is dynamically adjusting your prompts. You can guide your model to better match your expectations or use both methods simultaneously: dynamically adjusting your prompts and combining results from different models. Additionally, human-in-the-loop techniques can guide your model to more realistic results. You can fine-tune your model based on human evaluations, meaning you continue to train and refine the model using human feedback. This will reduce errors.

These are some methods you can apply to improve the outputs of large language models. In summary, large language models can sometimes give incorrect statements, but there are ways to reduce these statements. This way, large language models can be used in many fields.

Thank you for reading my article, and see you in the next one.

Keywords

Large Language Models
AI Hallucination
Autoaggressive Predictions
Tokenization
Training Data Bias
Ensemble Methods
Dynamically Adjusting Prompts
Human-in-the-Loop
Model Refinement

FAQ

Q: What is AI hallucination? A: AI hallucination is when large language models produce information that isn't based on real data but appears to be real.

Q: Why do large language models make mistakes? A: Large language models are trained to mimic human speech by predicting the next word in a sequence. They may make mistakes because they generate text without considering its truthfulness and always provide an answer regardless of the data's accuracy.

Q: What is the main goal of large language models? A: The main goal of large language models is to produce human-like text to impress or persuade users, rather than ensuring the truthfulness of their statements.

Q: What are some methods to reduce AI hallucination? A: Methods to reduce AI hallucination include ensemble methods (combining predictions from different models), dynamically adjusting prompts, and human-in-the-loop techniques (refining the model using human feedback).

Q: Can large language models be used effectively despite their tendency to hallucinate? A: Yes, large language models can be used effectively by employing methods to reduce incorrect statements, thus making them viable for various applications.