10 LLM Challenges

Challenges of building LLM applications for production by Chip Huyen

  1. Inconsistency
    • how to ensure user experience consistency?
      • same input -> different outputs
      • small input changes -> big output changes (temperature = 0 won't fix it)
    • how to ensure downstream apps can run without breaking?
      • no output schema guarantee
  2. Hallucination
    • poor performance on task that require factuality
      • bad text2sql performance
    • why?
      • Deepmind says lack understanding of cause and effect.
      • OpenAI says it's how we're annotating data in RLHF helps
  3. Compliance + privacy
    • buy: are APIs compliant?
    • build: what if your in-house chatbot leaks sensitive info?
  4. Context length
    • use cases: document processing, summarization, narrative, tasks involving genes and proteins
    • 16.9% of NQ-Open tasks have context-dependent answers (SituatedQA)
  5. Data drift
    • even when provided new data, existing models trained on past data fails to generalize to answer questions asked in the present (SituatedQA)
  6. Forward & Backward compatibility
    • same model, new data
    • how to make sure prompts still work with newer models?
  7. LLM on the edge
    • healthcare devices, autonomous vehicles, drive-thru voice bots, personal AI assistant trained on personal data
    • on device inference
    • training
      • on-device training: bottlenecked by compute + memory + tech available
      • if trained on server:
        • how to incorporate device's data?
        • how to send model's updates to device?
    • choose a model size
      • 7B param model (depending on sparsity)
        • $100 to fine tune
        • $25,000 to train from scratch
  8. LLM for non-english languages
    • performance:
    • tokenization (latency & cost differs with language)
  9. Efficiency of Chat as an interface
    • chat is not efficient but very robust
  10. Data bottleneck
    • rate of training dataset size growth is much faster rate of new data being generated
    • internet is being rapidly populated with AI-generated text

More on Chip Huyen's article: Building LLM applications for production and in this video series.

7/3/2023