A recent LessWrong post challenges the binary choice between low-bandwidth natural language and high-bandwidth uninterpretable vectors in transformer architectures.
\nIn a recent analysis published on LessWrong, the author explores a technical nuance in transformer architecture that has significant implications for AI interpretability and efficiency: the mechanism of information transfer between model components.
The Context
Current Large Language Models (LLMs) process information using high-dimensional vectors (embeddings) but typically communicate with the outside world-and often with other internal components-via discrete tokens (words or sub-words). This conversion from a rich vector state to a single token represents a severe bottleneck. The author describes this as embed(sample_token(output)), a process that strips away the subtle probability distributions and "thoughts" of the model, leaving only a single data point.
To solve this, some researchers propose passing the raw output vector directly to the next input stage. While this maximizes data transfer (bandwidth), it creates "neuralese"-a form of communication that is mathematically efficient but completely opaque to human observers. This renders the model's reasoning process a black box, complicating safety auditing and debugging. The industry faces a difficult trade-off: lose information to stay readable, or keep information and lose transparency.
The Core Argument
The LessWrong post argues that this is a false dichotomy. We do not need to choose between the low bandwidth of a single token and the uninterpretability of raw vectors. The proposed solution is to "use more text." By allowing the model to generate multiple tokens, phrases, or even paragraphs to represent its internal state, we can increase the bandwidth of the communication channel without abandoning natural language.
The author challenges the assumption that natural language is inherently low-bandwidth. It is only low-bandwidth if restricted to a single token at a time. By expanding the output to include "patches of bytes" or longer textual descriptions, the model can convey complex internal states in a format that remains readable by humans. This approach aims to maintain the "glass box" nature of natural language processing while mitigating the information loss associated with standard token sampling.
This concept is particularly relevant for the development of Chain-of-Thought (CoT) reasoning and multi-agent systems, where one model's output becomes another's input. If the communication channel is too narrow (one token), reasoning degrades. If it is too abstract (vectors), we lose oversight. The "more text" approach suggests a path where AI systems can be both highly capable and transparent.
For engineers and researchers working on model architecture and interpretability, this post offers a compelling perspective on how to structure internal data flows.
Read the full post on LessWrong
\n\n