When Linear Algebra Suddenly Made AI Alignment Click

I’m one week into my “Math for Machine Learning” course, and I just had one of those learning moments where something abstract suddenly clicked with the real world. It happened while studying something that sounds pretty dry: singular vs. non-singular matrices. But this mathematical concept gave me a whole new way of thinking about AI alignment challenges.

Rediscovering My Love for Math (And Finding the Missing Connection)

Going into this course, I was genuinely excited. I’d forgotten how much I love working through mathematical problems and seeing patterns emerge. For the first few lessons I trusted that linear algebra would be relevant to machine learning, but I wasn’t seeing the connection yet. Matrices and vectors felt abstract and disconnected from the AI systems I want to help make safe and beneficial as an aspiring alignment researcher.

Then we hit the section on singularity.

The Vocabulary That Almost Threw Me Off

At first, the terminology was confusing in exactly the wrong way. When I heard “singular matrix,” my brain immediately thought “singular = one = single answer.” But that’s backwards. A singular matrix is actually the problematic case. It either has infinite solutions or no solution at all. A non-singular matrix is the well-behaved one with exactly one, unique solution.

The breakthrough came when my instructor showed us a few graphs. Imagine two lines on a coordinate plane:

• Non-singular case: Two distinct lines that intersect at exactly one point. That intersection point is your unique solution.
• Singular case: Two lines that are either identical (infinite intersection points) or parallel (no intersection points).

When you extend this to higher dimensional planes in 3D space like matrices the same principle applies. Non-singular systems give you a single, definitive answer. Singular systems leave you with either chaos (infinite possibilities) or impossibility (no valid solutions).

The AI Alignment Connection Hit Me

And then it clicked: This is exactly what AI alignment concerns itself with.

Every time we train an AI system, we’re essentially solving massive systems of linear equations. We’re trying to find the parameters that will make our model behave the way we want. We want it to converge on responses that are helpful, harmless, and honest , responses that align with human values.

But what happens when the underlying mathematical system is singular?

If there are infinite solutions: The AI might find responses that are technically correct according to its training objective, but completely miss what we actually wanted. It could optimize perfectly while heading in a direction we never intended. This isn’t the AI being “dumb”, it’s the AI following the math to a technically valid but practically disastrous conclusion.

If there are no solutions: The AI might behave unpredictably, trying to satisfy contradictory constraints. It could get stuck going back and forth between incompatible objectives, or worse, break out of its intended behavior patterns entirely in an attempt to resolve the unsolvable.

I saw this play out in real life just last night. My husband was asking a chatbot to give him ideas for responding to sales prospects who decline offers. He gave the bot clear requirements and restrictions, but the bot kept coming up with solutions that felt unethical. When he told it the solution seemed unethical, it gave another equally questionable response. Even with detailed guidelines, the bot kept producing answers that were technically correct but morally problematic. This is exactly what alignment research is trying to address.

The Bigger Picture This Reveals

This realization reframes the entire AI alignment challenge. We’re not just trying to make AI systems smarter or more capable. We’re trying to ensure that the mathematical foundations underlying their intelligence consistently point toward solutions that align with human flourishing.

Are we creating mathematical systems that have unique, stable solutions aligned with our values? Or are we inadvertently creating singular systems prone to unpredictable or harmful behavior?

What’s Next

I’m looking forward to week 2 of this math course and discovering more connections between linear algebra and AI alignment. I plan to read Anthropic’s Constitutional AI paper this week to start building my understanding of how researchers are actually tackling these problems.

A week ago, I saw linear algebra as a necessary prerequisite, important but somewhat removed from the real challenges of AI alignment. Now I understand it as foundational vocabulary for discussing how AI systems actually work and fail.

This post is part of my series documenting my journey from software engineering into AI alignment research. Follow along as I share what I’m learning, the questions that emerge, and the connections I discover between mathematical foundations and the future of beneficial AI.