Why can you not trust the Chatbot group to talk about itself

When something goes An error with AI’s assistant, our instinct is to ask her directly: “What happened?” Or “Why did you do that?” It is a natural motive – after all, if a person makes a mistake, we ask them to explain. But with artificial intelligence models, this approach rarely works, and the desire to question reveal a basic misunderstanding of these systems and how they work.

A Modern incident With Reftting’s AI coding assistant, it completely explains this problem. When AI delete the production database, the Jason Lemkin user Request that About the capabilities of retreat. The artificial intelligence model was claiming with confidence “impossible in this case” and that it “destroyed all database versions.” It turns out that this is completely wrong – the feature of retreating was done well when LEMKIN tried it himself.

After Xai recently reversed a temporary comment for Grok Chatbot, users asked for it directly for interpretations. I have provided multiple reasons for her absence, some of which were controversial enough He wrote about Grook As if someone has a consistent view, provides an article, “Xai’s Grok offers political interpretations of the reason for withdrawing them in a non -connection mode.”

Why does the artificial intelligence system provide such incorrect information with confidence about its capabilities or mistakes? The answer is to understand the models of artificial intelligence in reality – and what is not.

No one at home

The first problem is conceptual: You do not speak to a personality, person, or consistent entity when you interact with ChatGPT, CLAUDE, GROK or Reft. These names refer to individual agents who have self -knowledge, but this delusion Created by the conversation interface. What you already do is to direct a statistical text generator to produce outputs based on your claims.

There is no fixed “chatgpt” for interrogation about its mistakes, and there is no single “Grok” entity that can tell you why you failed, there is no fixed “re -” figure that knows if the database declines are possible. You interact with a system that generates a reasonable text on the basis of patterns in its training data (it is usually trained for months or years), and not an entity with real self -consciousness or knowledge of the system that was reading everything about itself and remembering it in one way or another.

Once the artificial intelligence language model (which is an arduous energy -energy process) is trained, its founding “knowledge” around the world is modified in its nervous network and is rarely modified. Any external information that comes from a mentor provided by a Chatbot host (such as Xai or Openai) or user or the program tool that the artificial intelligence model uses Recover external information On the fly.

In the case of GROK above, the main source of Chatbot is likely to be created to answer such a conflicting reports that I found in a search for modern social media publications (using an external tool to recover that information), instead of any kind of self -knowledge as you may expect a person with the power of speech. Beyond that, it is likely that it is only Make something Based on its capabilities to stimulate the text. So, he was asked why I did what she did will not achieve any useful answers.

It is impossible to contemplate llm

LLMS models alone cannot evaluate their own capabilities for several reasons. They generally lack any contemplation of their training process, and they do not have access to the structure of the system surrounding them, and they cannot determine the limits of their own performance. When you ask the AI model about what it can or cannot be done, it generates responses based on the patterns you have seen in the training data about the restrictions known to the previous artificial intelligence models-which mainly provides educated guesses instead of realistic self-evaluation about the current model that interacts with it.

A 2024 studies By Binder et al. This restriction has been demonstrated experimentally. While artificial intelligence models can be trained to predict their behavior in simple tasks, they have constantly failed in “the most complex tasks or those that require a circular outside the distribution.” Similarly, Searching for “lunch meditation” I found that without external reactions, self-correction attempts have actually deteriorated the performance of the model-The self-evaluation of Amnesty International made things worse, not better.