Beyond filter lists: using LLMs to rethink ad blocking
summary
AdGuard prototypes replace keyword filters with three LLM-powered selectors: contains-embedding (fast, local embeddings cached; a small local model hits 100% on the author’s tests), prompt-based semantic checks (cloud prompts reach ~100% accuracy in ~1s but cost more), and contains-meaning-vision (vision-capable LLMs detect image-first ads with very high accuracy but introduce 10–15s latency and UI flicker). The prototypes work end-to-end, but production requires hybrid use of existing filter lists to limit what gets analyzed.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:51How ad blocking works today
- 3:40Limitations of filter lists
- 4:24Previous ML attempts
- 8:05Why LLMs could change ad blocking
- 8:50Blocking by meaning demo
- 10:19Teaching blockers to see images
- 13:06Connecting LLMs with filter lists
- 15:55Practicality of LLM prototypes
- 18:05Future of ad blocking with LLMs
- 18:52Key takeaways and Q&A
related talks
15mDouble-Edged Shield: On the Fingerprintability of Customized Ad Blockers
21mPERCIVAL: Making In-Browser Perceptual Ad Blocking Practical with Deep Learning
19mBlock, unblock, block! How ad blockers are being circumvented…
29mAutomating Privacy Reviews with LLMs: Challenges & Opportunities
39mBeyond the Prompt: Evaluating, Testing, and Securing LLM Applications
50mBeyond the Prompt: Evaluating, Testing, and Securing LLM Applications
01h 19mEffective Transparency and User Controls: Are We Close to a Breakthrough?
46mLLM—Search Engine Killer?