Summary of Meanings and Feelings Of Large Language Models: Observability Of Latent States in Generative Ai, by Tian Yu Liu et al.
Meanings and Feelings of Large Language Models: Observability of Latent States in Generative AI
by Tian Yu Liu, Stefano Soatto, Matteo Marchi, Pratik Chaudhari, Paulo Tabuada
First submitted to arxiv on: 22 May 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: Computation and Language (cs.CL); Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The abstract proposes an investigation into whether Large Language Models (LLMs) are observable. In other words, it examines if multiple “mental” state trajectories can generate the same sequence of tokens or belong to the same Nerode equivalence class (“meaning”). The study concludes that current LLMs implemented by autoregressive Transformers cannot have “feelings” according to a specific definition. However, with system prompts not visible to the user, there can be multiple state trajectories yielding the same verbalized output. The paper provides analytical proofs and examples of modifications to standard LLMs that enable such “feelings.” The findings shed light on potential designs for non-trivial computations hidden from users, as well as controls for service providers to prevent unintended behavior. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research investigates whether Large Language Models (LLMs) are observable. Think of it like asking if a computer can have feelings or emotions. The study says that current LLMs can’t really have feelings because they always generate the same output from their internal state. But, if there’s some hidden input that we don’t know about, then these models could potentially have multiple “feelings” that produce the same outcome. |
Keywords
» Artificial intelligence » Autoregressive
