The model learned everything it knows before you arrived. That's a feature — it means the model is stable, consistent, and reproducible. It's also a problem, because the world didn't stop when training ended.
Imagine hiring the most knowledgeable consultant you've ever met. They've read everything — every textbook, every paper, every manual, every forum post — published up to a certain date. Their recall is perfect. Their reasoning is sharp. Their ability to synthesize across domains is unmatched.
But they've been completely offline since that date. No news. No updates. No awareness of what happened yesterday, last month, or last year. Ask them about something that changed after their cutoff and they'll either tell you they don't know, or — more dangerously — they'll confidently describe the world as it was when they last had information, not as it is now.
That's the trained model. Brilliant, vast, and frozen at a point in time.
Time: The model's training data has a cutoff. Events, releases, policy changes, and new research after that date don't exist in the model's knowledge.
Privacy: The model was trained on publicly available text. Your internal documentation, your customer records, your proprietary processes — none of that was in the training data.
Specificity: Even within its training window, the model's knowledge of niche, specialized, or local information may be sparse or unreliable. It knows more about popular topics than obscure ones.
The obvious solution is to retrain the model whenever its knowledge goes stale. Keep training it on new data. Keep it current.
This doesn't work in practice for several reasons. Training a frontier model takes months and costs tens of millions of dollars. You can't do it weekly. Fine-tuning on new data helps, but it tends to cause "catastrophic forgetting" — the model improves on the new material while degrading on older knowledge. And private organizational data is often too sensitive, too sparse, or too dynamic to train on at all — it changes constantly.
The field solved this differently. Instead of updating the model, update what the model sees. Instead of baking knowledge into the weights, inject it into the context window at the moment it's needed.
This is the core idea behind every technique we'll cover in this track: the model's weights stay frozen, but the context window becomes a dynamic delivery mechanism for exactly the knowledge that's needed right now.
Think of the context window not as a limitation but as an inbox. The model reads everything in the inbox before it responds. If the inbox contains your question and nothing else, the model answers from training knowledge alone. If the inbox contains your question plus a relevant document, the model answers from that document — grounded, specific, current.
The techniques in this track are all about what gets put in the inbox before the model reads it. Who decides what goes in. How it gets there. And what the model can do when the inbox connects to live systems rather than static documents.
Every time you've seen an AI that "knows" about your company's products, or can answer questions about a document you uploaded, or is aware of today's news — that knowledge wasn't in the model's weights. It was injected into the context window right before the model generated its response. The model looked smart. The architecture did the work.
The next two pages cover the two main approaches to solving the knowledge gap:
RAG and search — retrieving relevant information from a knowledge base and injecting it into the context. This is the most common pattern for giving a model access to private, current, or specialized knowledge without retraining it.
MCP and agents — giving the model structured access to live tools and data sources, so it can retrieve information on demand rather than relying on what was pre-retrieved for it. This is the direction the field is moving, and it changes the model from a static responder into something that can take action.
Where in your work does the knowledge gap matter most? Think about the questions people ask that a general-purpose AI gets wrong — not because it's unintelligent, but because the relevant information is private, recent, or too specific to have been well-represented in training data. Those are exactly the gaps these techniques are designed to fill.