The alignment problem

A computer is a calculator. When you type 1 + 1, it gives you 2.

AI today is a very advanced calculator. It reads text and predicts what text should come next.

Say you are studying for an exam and you use a calculator to check your answers. The calculator does not know about your exam. It only calculates the numbers you give it.

AI works the same way. It follows the instructions and goals it is given. It does not know why you are using it or what you need the result for.

The alignment problem is the gap between the goal an AI is given and what the people using it actually need.

Why it matters

A startup builds an app that matches low-income high school students with scholarships. A student enters their grades, background, and interests, and the app lists the scholarships they qualify for.

The company measures the app's success by how many matches it shows each student, so the app learns to show as many as possible. The list fills with large national scholarships that thousands of students apply for. A small local scholarship that this student had a strong chance of winning ends up at the bottom of the list, where the student never sees it.

The app reached the goal it was given: more matches. The student did not reach theirs: winning a scholarship.

Solving problems starts with knowledge everyone understands.

The lab

Most people outside the tech industry do not want to learn the vocabulary: AI, LLM, alignment, agent, benchmark. They should not have to.

Knowledge about AI should not stay inside a small group of companies and researchers. Technology does not have to feel scary or out of reach.

Differend Labs studies the gap between what people need and what AI does, and explains what it finds in plain language that anyone can understand.

• Small. Each study asks one question.
• Honest. Each study is planned before it runs, and failures are published.
• Open. Code and data are shared.
• Plain. Everything is written for everyone.