What is AI and how does it work?
Understanding how AI works is key to getting good results. Using AI is a lot like talking to a person: the more precise your instructions and context, the more helpful the answer.
Large Language Models (LLMs)
Generative AI systems like AI CTRL are built on Large Language Models, or LLMs: very sophisticated text prediction engines, trained on enormous amounts of text and using the patterns in it to generate a response when you ask a question.
Different models have different strengths, depending on what they were trained on. Some are better at creative writing, others at technical analysis or complex reasoning. That's why AI CTRL offers multiple models from multiple vendors.
Prompt Engineering
Because these models predict answers based on their training, the way you ask matters a great deal. There's a whole discipline around this called prompt engineering, and it comes down to four habits:
- Be specific about the outcome you want
- Provide relevant context or examples
- Break complex tasks into steps
- Adjust and retry until you get what you're after
What the model doesn't know
Anything recent. Every model has a knowledge cutoff date and knows nothing that happened after its training ended. Ask about last week's market trends and you'll get an answer, often a confident one, built on outdated information. Verify critical facts, and turn on web search when you need current sources attached to the conversation.
Anything about your organization. A model may have general knowledge from public sources, but it knows nothing about your internal projects, processes, or documentation unless you provide them. That's what attachments, knowledge collections, and connectors are for, see Attaching Documents & Data.
How context works
Context is everything the model can see while answering: your question, the conversation so far, and anything you've attached. Two things about it are worth knowing.
Context is per conversation
Each conversation stands alone. Everything discussed in a chat stays available within it, so you can ask follow-ups without repeating yourself, but none of it carries into your next chat. Every new conversation starts cold.
That's deliberate. AI CTRL doesn't quietly build a picture of you over time; if you want something remembered across conversations, you say so, with a memory or a personal system prompt.
Context has a size limit
How much fits depends on the model. Very long conversations start to degrade, the model loses track of earlier details or drifts toward generic answers. They also cost more, because each message re-sends everything discussed so far.
So: new topic, new chat. And when a conversation is getting unwieldy, ask the AI to summarize it and start fresh with that summary.