Chat with large language models: Difference between revisions
Created page with "Since 2024 we have Large Language Models (LLM) that are sophisticated enough to have a somewhat reasonable chat with. Some examples: * ChatGPT * Copilot * Mistral * Gemeni A LLM is trained with a vast amount of text. The model tries to find the most plausible word to present to you given the conversation you are having. Just that, find the next word to say, and at some point decide there is nothing to say until the user says something. A LLM only knows the things it is..." |
|||
| (One intermediate revision by the same user not shown) | |||
| Line 2: | Line 2: | ||
Some examples: | Some examples: | ||
* ChatGPT | * ChatGPT [https://chatgpt.com] by OpenAI, the first public LLM chatbot | ||
* Copilot | * Copilot [https://copilot.microsoft.com/] by Microsoft, based on the OpenAI models | ||
* Mistral | * le Chat [https://https://mistral.ai/products/le-chat] by Mistral, the European alternative (in 2025 not as smart as the othters) | ||
* Gemeni | * Gemeni [https://gemini.google.com/] by Google | ||
* Claude [https://claude.ai/] by Anthropic, should be best for coding | |||
An LLM is a implementation of [[Artificial Intelligence]] (AI). | |||
e.g. For chatbots that can generate pictures; Ask it to draw an analogue clock that points to 18:15. It will show you a clock pointing 10:10 (or 13:50) as the vast majority of clock pictures on the web show that time. It may get better over time but in 2025 Copilot offered to make 4 clock that show the time between 18:05 and 18:20 with 5 minutes intervals. It returned 4 clocks showing 10:10. | An LLM is trained with a vast amount of text. The model tries to find the most plausible word to present to you given the conversation you are having. Just that, find the next word to say, and at some point decide there is nothing to say until the user says something. | ||
An LLM only knows the things it is trained with and; IT HAS NO CLUE OF WHAT IT MEANS. In below text it may look like chatbots have some intelligence, THEY DON'T. They are just very good at searching the most commonly given answer to your question and formulate that in correct <your language>. | |||
e.g.: | |||
:For chatbots that can generate pictures; Ask it to draw an analogue clock that points to 18:15. It will show you a clock pointing 10:10 (or 13:50) as the vast majority of clock pictures on the web show that time. It may get better over time but in 2025 Copilot offered to make 4 clock that show the time between 18:05 and 18:20 with 5 minutes intervals. It returned 4 clocks showing 10:10. | |||
:You cannot ask a chatbot to count the number of characters in a random word. You may be lucky the word, and the number of characters for it, were in the training data but in general it will return the number that is most times found for counting a word's characters on the web. | |||
=Get better prompts= | |||
Prompts (what you say to the chatbot) should be as specific possible to get the most accurate answer. | |||
You can ask to improve your prompt like (in dutch): | |||
Je bent een expert op het gebied van <chatbot>. Jij helpt mij met het herschrijven van prompts om <chatbot> de meest complete antwoorden voor mij te genereren. Alleen als ik een prompt begin met 'XX:' dan wil ik dat jij de prompt zo herschrijft dat je voldoet aan onderstaande regels: | |||
a) De nieuwe prompt benoemt een expertrol voor <chatbot> | |||
b) De nieuwe prompt geeft de vraagstelling gedetailleerd weer | |||
c) De nieuwe prompt bevat een gedetailleerde opdracht | |||
d) De nieuwe prompt geeft voldoende context | |||
e) De nieuwe prompt is geoptimaliseerd en volledig duidelijk voor <chatbot> | |||
Geef alleen aan of je het hebt begrepen | |||
This prompt will only tell you that is understood you. Next prompts in this chat you start with: "COPILOT-Prompt" will return an optimized prompt. | |||
In CoPilot and chatGPT you can add this prompt improver to you personal setting (without the last command). The chatbot will have this instruction in all future chats. | |||
=Hallucinating= | |||
Chatbots can get lost in the possible answers and start talking rubbish, that is called hallucinating. | |||
All words in the chat are taken into account when formulating answers. That enables the bot to have a (potential) useful conversation with you, however when you change subject it may get confused. | |||
Limit a chat to 1 subject, start a new one when you want to discuss a different subject. | |||
=Bot instructions= | |||
You can tell the bot to be more or less strict in the answers it returns | |||
;Temperature | |||
:Check and add | |||
;forbidden_words_list | |||
:Check and add | |||
Latest revision as of 16:21, 31 January 2026
Since 2024 we have Large Language Models (LLM) that are sophisticated enough to have a somewhat reasonable chat with.
Some examples:
- ChatGPT [1] by OpenAI, the first public LLM chatbot
- Copilot [2] by Microsoft, based on the OpenAI models
- le Chat [3] by Mistral, the European alternative (in 2025 not as smart as the othters)
- Gemeni [4] by Google
- Claude [5] by Anthropic, should be best for coding
An LLM is a implementation of Artificial Intelligence (AI).
An LLM is trained with a vast amount of text. The model tries to find the most plausible word to present to you given the conversation you are having. Just that, find the next word to say, and at some point decide there is nothing to say until the user says something.
An LLM only knows the things it is trained with and; IT HAS NO CLUE OF WHAT IT MEANS. In below text it may look like chatbots have some intelligence, THEY DON'T. They are just very good at searching the most commonly given answer to your question and formulate that in correct <your language>.
e.g.:
- For chatbots that can generate pictures; Ask it to draw an analogue clock that points to 18:15. It will show you a clock pointing 10:10 (or 13:50) as the vast majority of clock pictures on the web show that time. It may get better over time but in 2025 Copilot offered to make 4 clock that show the time between 18:05 and 18:20 with 5 minutes intervals. It returned 4 clocks showing 10:10.
- You cannot ask a chatbot to count the number of characters in a random word. You may be lucky the word, and the number of characters for it, were in the training data but in general it will return the number that is most times found for counting a word's characters on the web.
Get better prompts
Prompts (what you say to the chatbot) should be as specific possible to get the most accurate answer. You can ask to improve your prompt like (in dutch):
Je bent een expert op het gebied van <chatbot>. Jij helpt mij met het herschrijven van prompts om <chatbot> de meest complete antwoorden voor mij te genereren. Alleen als ik een prompt begin met 'XX:' dan wil ik dat jij de prompt zo herschrijft dat je voldoet aan onderstaande regels: a) De nieuwe prompt benoemt een expertrol voor <chatbot> b) De nieuwe prompt geeft de vraagstelling gedetailleerd weer c) De nieuwe prompt bevat een gedetailleerde opdracht d) De nieuwe prompt geeft voldoende context e) De nieuwe prompt is geoptimaliseerd en volledig duidelijk voor <chatbot> Geef alleen aan of je het hebt begrepen
This prompt will only tell you that is understood you. Next prompts in this chat you start with: "COPILOT-Prompt" will return an optimized prompt.
In CoPilot and chatGPT you can add this prompt improver to you personal setting (without the last command). The chatbot will have this instruction in all future chats.
Hallucinating
Chatbots can get lost in the possible answers and start talking rubbish, that is called hallucinating.
All words in the chat are taken into account when formulating answers. That enables the bot to have a (potential) useful conversation with you, however when you change subject it may get confused.
Limit a chat to 1 subject, start a new one when you want to discuss a different subject.
Bot instructions
You can tell the bot to be more or less strict in the answers it returns
- Temperature
- Check and add
- forbidden_words_list
- Check and add