Industry talk · ICVGIP 2025 (VISIONxFAIL workshop)
A talk at ICVGIP
“Challenges in AI: Beyond the Context Window”
· IIT Mandi, India (presented remotely)

On Wednesday, 17 December 2025, I gave a short industry talk at VISIONxFAIL, a workshop at ICVGIP 2025 about constraints, challenges and negative results in computer vision. The conference was at IIT Mandi and I joined remotely, with a recorded talk called “Challenges in AI: Beyond the Context Window”. It came out of my own experience of working with AI, and the problems I keep running into.
I talked about where the context window becomes a bottleneck, using software engineering as the example. You want a model to write code without running out of room, to understand a huge codebase, to share context when several agents are generating code in parallel, and to build a new feature without missing something that matters.
The idea I put forward is something I call Minimum Critical Context, or MCC. It is the least context you need to carry over from one task to do the next one perfectly, and a task can be anything from a file to a document to a set of logs. Once every file has its MCC, a big job can be broken into small ones. Agents work on them in parallel, and the results are checked and put back together. More compute always helps, but it costs money, and the smaller you cut the pieces, the faster they run and the worse the results get, so it is a balance.
Then there is the question of who writes the MCC. It is easy to assume that takes the very best domain knowledge. In my experience that is true only for specialised fields that models have seen little of, and software is not one of them. With some base instructions and a persona built around the downstream task, a model can work out most of the MCC on its own. It is sensitive to small nudges, so getting it right takes a fair bit of testing.
I finished with the open questions. It is still not clear how to do this efficiently in specialised fields, or whether models could write the MCC instructions themselves. I have also seen results get worse as more of the context window is used, long before it is full, and there is no good metric for that trade-off yet. Thank you to the organisers for having me.
The slide uses a presentation template by Slidesgo.