Bani Chaudhuri, who leads content and video at Tars, explains the thing people usually read as the AI ignoring them. The context window is the entire conversation, everything you said and everything the model said back, counted in tokens against a hard limit that every model has.

The useful part is what happens before you reach that limit. Models hold on to the start of a conversation and the most recent messages, and the middle gets fuzzy, which is how you end up thirty messages deep with a confident answer that ignores what you specified halfway through. A big context window is not an instruction to fill it.

What this video covers

• What the context window holds, and why it is measured in tokens rather than messages
• Why the middle of a long conversation gets lost before the limit is reached
• Two things that fix it, summarizing the key points every ten to fifteen messages, or starting a fresh thread

Chapters

• 0:00 What the context window is
• 0:46 Losing the middle of a long conversation
• 1:30 Keeping conversations short and focused