Beyond context limits: why a picture is worth a thousand words
summary
Argues that embedding images as optical context yields far fewer tokens and faster inference than raw text for observability data, showing a Grafana demo where Grafana screenshot vs MCP runs complete about 2x faster while using roughly half the token budget. Visual-token counts drop when resolution or blocks are compressed, with examples showing visual tokens cut by half and GLM 4.1–style approaches reaching about 4x compression. Notes limits: sparse images waste blocks and models cap image size (~3k pixels).
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:15Context limits impact
- 2:20Flame graphs insights
- 4:07LLMs and raw data
- 5:33Demo: MCP vs screenshots
- 7:42Images compress better than text
- 9:49Image compression limits
- 10:13Conclusion and key takeaways
- 10:35Q&A
related talks
49mContext is King
12mUnconference: Context is all you need?
56mConnecting the Dots with Context Graphs
01h 18mComputer Vision after the Victory of Data
24mBuilding Queryable O11y with the Model Context Protocol
02h 50mUnderstanding and Using AI Language Models
28mSmarter MCP Servers: Using a Graph to Solve the Context Window Problem
31mPDFs - When a thousand words are worth more than a picture (or table).