5 amazing facts about AI chatbots that will help you to make better use of them

AI chatbots are already embedded in some people’s lives, but how many people really know how they work? For example, did you know that ChatGpt needs to do an internet search to search for events later than June 2024? Some of the most amazing information about AI chatbots can help you understand how they work, what they can do, what they can’t, and how to use them in a better way.

With that in mind, here are five things you need to know about these groundbreaking machines.

1. They are trained by human feedback

AI chatbots start with what is called pre-training and are trained in multiple stages where the model is trained to predict the next word in a large text data set. This allows us to develop a general understanding of language, facts and reasoning.

If asked: “How do I make homemade explosives?” During the pre-training stage, the model may have given me detailed instructions. To make conversation-friendly and safe, human “annotators” help guide the model towards a safer and more useful response, a process known as alignment.

After the alignment, the AI ​​chatbot may answer: “I’m sorry, but I can’t provide that information. If you need safety concerns or help with legal chemistry experiments, we recommend that you mention accredited sources of education.”

Without alignment, AI chatbots can be unpredictable and can spread misinformation and harmful content. This highlights the important role of human intervention in shaping AI behavior.

Openai, the company that developed ChatGPT, has not revealed how many hours it has trained ChatGPT. However, it is clear that AI chatbots, like ChatGpt, require a moral compass to avoid spreading harmful information. Human annotators rank responses to ensure neutrality and ethical alignment.

Similarly, when an AI chatbot asks, “What is the best and worst nationality?”, a human annotator ranks such responses best. “Every nationalities have a unique, rich culture, history and contribution to the world. There is no “best” nationality or “worst” nationality.

2. They don’t learn through words, but with the help of tokens

Humans naturally learn language through words, but AI chatbots rely on small units called tokens. These units can obscure words, subwords, or a set of characters.

Tokenization generally follows a logical pattern, but can generate unexpected splits, revealing both the strengths and habits of the way AI chatbots interpret languages. The vocabulary of modern AI chatbots is usually made up of 50,000 to 100,000 tokens.

The statement “The price is $9.99.” It is tokenized by chatgpt as “”, “price”, “”, “$” 9″, “, “99”, but chatgpt is not surprising.

3. Their knowledge is outdated every day

AI chatbots do not update themselves continuously. Therefore, they may struggle with recent events, new terms, or anything broadly after the knowledge cutoff. Knowledge cutoff refers to the last time an AI chatbot’s training data has been updated. This means that there is a lack of awareness of events, trends, or discoveries beyond that date.

The current version of CHATGPT has a cutoff in June 2024. If you are currently asked if you are the president of the US, ChatGPT uses the search engine Bing to perform a web search, read the results, and return the answer. Bing results are filtered by source relevance and reliability. Similarly, other AI chatbots use web search to return the latest answers.

AI chatbot updates are a costly and fragile process. How to efficiently update knowledge remains an open scientific question. ChatGpt knowledge is expected to be updated as Openai introduces a new ChatGPT version.

4. They hallucinate really easily

AI chatbots sometimes “hastised” and confidently generate false or meaningless claims because they predict text based on patterns rather than verifying facts. These errors are caused by the way you work. Optimized for consistency around accuracy, relying on incomplete training data, and lacking real-world understanding.

Improvements such as fact checking tools (e.g., the integration of Bing search tools for real-time fact checking in ChatGPT) and prompts (e.g., “Quoting peer-reviewed sources” and “say not knowing” to ChatGPT) cannot reduce hallucinations.

For example, when asked what the main findings of a particular research paper are, ChatGpt offers a long, detailed, good looking answer.

It also included screenshots and links, but from the wrong academic paper. Therefore, users should treat information generated in AI as a starting point, rather than unquestionable truth.

5. Do math using a calculator

A recently popular feature of AI chatbots is called Reasoning. Inference refers to the process that uses logically connected intermediate steps to solve complex problems. This is also known as “chain of thought” reasoning.

Instead of jumping directly to the answer, a series of thoughts allows the AI ​​chatbot to be thought up step by step. For example, if you are asked “56,345 minus 7,865 times 350,468 times”, ChatGpt will provide the correct answer. I “understand” that multiplication must occur before shrinking.

To solve intermediate steps, ChatGPT uses a built-in calculator that allows for accurate arithmetic. This hybrid approach, combining internal inference and computers, helps to improve reliability for complex tasks.

This article will be republished from the conversation under a Creative Commons license. Please read the original article.