AI interpretability and alignment are tools that look inside a trained AI model to see what it has learned, adjust it so that it behaves as intended, and often compress it so that it runs more cheaply. Interpretability means seeing inside the model. Alignment means making it behave the way its makers intend. Compression means making it smaller.
A model is a very large set of numbers, called weights, arranged in layers. When the model reads an input, each layer produces its own set of numbers, called activations. Nobody writes these numbers by hand; they are learned in training. So even the model's makers cannot easily say why it gave a particular answer. LLM evaluation looks only at the answers, from outside. Interpretability tools look at the activations themselves.
One common method trains a small helper model, called a probe, to detect a particular idea inside the activations, such as 'this input mentions a person's address'. Another method traces which words of the input, and which internal parts of the model, led to an answer. Once a pattern is found, alignment tools can change the model's behaviour. They may weaken the pattern directly, or train the model further with examples that correct it.
Consider a lender in Mysuru that uses a model to read small-business loan applications and write a short summary for its credit team. In tests, the summaries are more often negative for applicants from certain areas of the city. The lender's data team uses an interpretability tool. It finds a pattern inside the model that reacts strongly to the applicant's pincode, and that pushes the summary towards words such as 'risky'. The lender does not want its summaries to depend on where an applicant lives, which can stand in for community or income. So the team trains the model further to ignore that pattern, and tests again. The gap in negative summaries between areas falls from 18 percentage points to 2.
Compression uses the same knowledge of the model's insides. If a tool can see which parts of a model matter least for a task, it can remove them, or store the numbers with fewer bits. A model reduced from 16-bit to 4-bit numbers needs about a quarter of the memory. That is one way a model is made small enough for edge AI on a phone or a camera.
Alignment findings can also become new rules in LLM guardrails. Interpretability is still mainly a research field. It matters most where a large language model (LLM) helps make decisions that affect people.