Large Langunage Models - Strengths, Weaknesses, Opportunities & Threats

Understand the difference between Language Tasks and Knowledge Tasks. Language tasks involve understanding, generating language such as writing essays, formatting an output to JSON etc. Knowledge Tasks require accessing and providing factual information or real-world knowledge. LLMs are good at Language Tasks (reformatting outputs, reformulating etc) but struggle with Knowledge tasks:

• They produce incorrect and contradictory statements
• They produce dangerous and socially unacceptable statements (That include bias and other socially unacceptable output)
• Limited Training, Retraining and Inference is expensive
• The knowledge cannot easilty be updated: Updating just one fact is near-to-impossible
• Lack of attribution: no easy way to determine which document in the training data is responsible for which part of the knowledge
• Poor performance on non-language tasks (reasoning tasks etc.)

Retrieval Augemented Generation (RAG) seem to be the holy grail, but they are not the panacea:

• Implicit world knowledge (in LLM) can interfere with knowlege from retrieved documents (hallucination)
• Only as good as the vector embeddings generated for each chunk of data

Some academic research hints at the fact that “hacky” get-arounds allow for more robust solutions when dealing with LLMs:

• Improve consistency of answers by asking the same question multiple times and finding the most consistent answer (through majority voting)
• Leverage feedback based mechanisms

Threats:

• Multimodal models
• Smaller models (Alpaca, Llama, Mistral)
• Security & Safety Issues: Jailbreak attacks, Prompt Injection attacks, Data Exfiltration attacks, Data poisoning attacks etc.
• Evaluation: Latency, Tokens, Human Evaluation

Opportunities :

• LLMOps: Including Data Drift, Model Quality Drift
• Design for potential model retraining / Fine-tuning: Capture the API input/output to allow for proprietary model training

Interesting reads / Sources:

• Stanford Natural Language understanding
• Thomas Dietterich, whats wrong with LLM
• Adversarial attacks on LLM