On the Biology of a Large Language Model

I know we have talked many times about how LLMs work, their implications, and their use cases. But despite so many articles on LLMs, they remain pretty much a black box.

But the real question we must ask ourselves is why it is important even to understand the internals of these systems. We need this understanding so that we don’t confuse ourselves about the capabilities of these systems. LLMs can easily fool us into thinking that they are solving a problem by reasoning instead of memorization.

And in today’s blog we are going to dive deeper into the internals and look at the biology of LLMs; to understand what makes them tick.

Table Of Content

• Background Of AI Interpretability
• What Is Monosemanticity?
• Sparse Autoencoders (SAE)
• Multi-Step Reasoning
• Planning In Poem
• Addition In LLMs
• Chain-of-Thought Unfaithfulness and Hidden Goals In Language Models
• Final Note

Background Of AI Interpretability

DL models can learn human-comprehensible algorithms. These models can be understood, but by default, they have no incentive to make themselves legible to us.

Mechanistic interpretability in the context of DL involves delving into the inner workings of these models to understand how individual…