[Paper Review] Eight Things to Know about Large Language Models

Samuel R. Bowman|arXiv (Cornell University)|Apr 2, 2023

Natural Language Processing TechniquesComputer Science97 citations

TL;DR

A synthesis of eight surprising, evidence-based points about LLMs, highlighting scaling laws, emergent behaviours, representation learning, steering limitations, interpretability challenges, performance vs. human benchmarks, values and biases, and the misleading nature of brief interactions.

ABSTRACT

The widespread public deployment of large language models (LLMs) in recent months has prompted a wave of new attention and engagement from advocates, policymakers, and scholars from many fields. This attention is a timely response to the many urgent questions that this technology raises, but it can sometimes miss important considerations. This paper surveys the evidence for eight potentially surprising such points: 1. LLMs predictably get more capable with increasing investment, even without targeted innovation. 2. Many important LLM behaviors emerge unpredictably as a byproduct of increasing investment. 3. LLMs often appear to learn and use representations of the outside world. 4. There are no reliable techniques for steering the behavior of LLMs. 5. Experts are not yet able to interpret the inner workings of LLMs. 6. Human performance on a task isn't an upper bound on LLM performance. 7. LLMs need not express the values of their creators nor the values encoded in web text. 8. Brief interactions with LLMs are often misleading.

Motivation & Objective

• Motivate informed discussion among researchers, advocates, and policymakers about the implications of LLMs.
• Summarize evidence for how LLMs scale, what behaviours emerge, and the limits of steering and interpretability.
• Highlight ethical, governance, and safety considerations tied to deployment and oversight.

Proposed method

• Survey and synthesize evidence from prior work on LLM scaling, emergent behaviours, and representational capabilities.
• Cite empirical results from scaling laws, BIG-Bench, and related studies to support claims.
• Discuss limitations of current steering, interpretation, and evaluation methods.

Experimental results

Research questions

• RQ1 What are the surprising claims supported by evidence about large language models?
• RQ2 How do scaling and investment affect LLM capabilities and behaviours?
• RQ3 What are the limits of steering, interpretability, and value alignment in LLMs?
• RQ4 What risks and governance considerations arise from current LLM development and deployment?

Key findings

• LLMs become more capable with investment and scale, even without targeted innovations.
• Some important LLM behaviours emerge unpredictably as models scale.
• LLMs develop internal representations of the outside world to some extent.
• There are no reliable techniques that guarantee steering LLM behavior across all settings.
• Experts lack complete ability to interpret LLM internals.
• Human performance is not a universal upper bound for LLM tasks.
• LLMs need not express the creators’ or training data values.
• Brief interactions with LLMs can be misleading about their capabilities.