Where there's a will, there's a way: The mad genius of LLMs as a classifiers
summary
Argues that LLMs can replace or augment intent classifiers in production if used selectively, with Swisscom using 1,300 labeled utterances to prototype. Presents concrete architectures: a fast few‑shot classic vs hybrid/filter/embedding architectures, an embedding retrieval step, and a tiny custom head mapping token probabilities to intent labels. Reports gains in prototyping speed and multilingual robustness but real costs: latency, unpredictability, transcription noise, multi‑turn context loss, and the need for careful prompt and business‑impact testing.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:10LLMs as classifiers
- 5:39Intent detection use case
- 14:03Techniques and architectures
- 25:57Generalizing classification
- 29:12Summary & Q&A
related talks
20marchive – Ethan Ferdosi
49mA Practical Approach to Using LLMs: Peculiarities and Pitfalls
24mConversational Search with LLMs
30mLLM-Based Recommendation Systems
39mHallucination-Free LLMs: Strategies for Monitoring and Mitigation
30mLLM Revolution: Necessity or Excess? A Study of the Most Efficient Solutions
35mReading the Mind of an LLM
36mLLMs and AI agents demystified