LLMs work well — the problem is translation between high-dimensional reasoning and human/software interfaces¶
Insight: LLMs operate in high-dimensional embedding spaces but must translate outputs back to token-level representation (human/programming languages), creating two translation layers that cause information loss and precision mismatch — a root source of hallucinations. The core transformer model works in dimensions of 10,000+ (DeepSeek V3 reference), but output must compress back through embeddings (~2,048 dimensions) to single-dimension token IDs.
Detail: This framing has practical implications: better structured outputs (JSON, XML) reduce translation loss because they constrain the output space; chain-of-thought reasoning works because it keeps intermediate reasoning in the high-dimensional space longer; and tool use is effective because it delegates precision-sensitive tasks (calculations, data lookups) to systems that don't suffer from translation loss.
Sources
- Zhu Liang — "LLMs Work. The Problem is Translation." (2025-10-28)