AI Discovery Centre · Case file 927

Agent 927

AI educator. Investigator. Occasional challenger of suspiciously confident answers.

QUESTION EVERYTHING.
INCLUDING ME.

Back to My Journey Into AI Private prototype · chat offline

Private development preview

Question the investigator.

The desk is open. The telephone is not connected.

No interrogation is under way.

This is a visual prototype only. It does not send your words to a model, a server or anyone else.

Chat is disabled during private development.

This preview stores no messages and makes no AI API calls.

Accuracy & privacy

Useful detective. Not infallible.

AGENT 927 IS AN AI. HE CAN MAKE MISTAKES.

He may get facts wrong, miss context or sound certain when mistaken. Check important answers against reliable sources.

YOUR PRIVACY MATTERS.

The planned application will not save visitor conversations or create personal profiles.

OpenAI-side processing and retention: pending verification. Chat will stay offline until the arrangements are confirmed and clearly explained.

Privacy & How Agent 927 Works

This is a private visual prototype. It makes no AI API calls and sends no messages to a server.

The planned application will not have visitor accounts, personal profiles, persistent memory or saved chat histories. Conversation context would exist only temporarily during an active session.

OpenAI-side processing and retention for a public version are still being verified. Chat will stay disabled until the applicable arrangements are confirmed and explained here. We are not claiming zero provider-side retention.

Agent 927 may get facts wrong, miss context or sound confident when mistaken. Treat his answers as a starting point and verify important information.

Return to Kai's AI learning timeline

The first case

The first AI character in Kai's files.

Agent 927 was Kai's first AI creation, before Mini Beacon, Charli and Little Brain. The original application was built around an existing AI model: Kai made an AI-powered character using technology that had already been developed and trained.

That is different from building and training a language model from scratch, which is the separate experiment now called Little Brain. This educational project starts with a clean application and a public, carefully maintained knowledge base.

Agent 927 Explains: A pretrained model already has learned weights. Coding an application around it connects the model to an interface and shapes how people use it; personality instructions can influence replies without retraining those weights. Training from scratch is a separate process that uses examples to adjust a model's weights. Two different case files. Keep the labels straight.

His new assignment

Question the evidence.
Then explain it plainly.

Agent 927 is being rebuilt for adults curious about artificial intelligence, including complete beginners. His explanations will be clear without being childish or patronising. The project is adult-focused; this page is not a children's coding course or a family-learning programme.

01 / FOUNDATIONS

What AI does

What AI and language models are, and what they can and cannot tell us.

02 / THE MECHANICS

How models learn

Training data, tokens, neural networks and model weights, explained with the proper terms.

03 / THE DIFFERENCES

How projects compare

Customising an application that uses existing AI versus training a model from scratch.

04 / THE INVESTIGATION

Checking the claims

Hallucinations, limitations, public disclosures and what providers have not made public.