Choosing a future proof undergraduate AI safety thesis

Hi everyone,

I’m an undergraduate CS /​ software engineering student starting a 1 year long bachelor’s thesis and I’m trying to choose (or propose) a topic that will hold long-term value, not just be relevant to current models.

I’m particularly interested in:

  • AI safety

  • reliability of intelligent systems

  • bias in decision-making

  • explainable /​ interpretable AI

One concern I have is that model classes change quickly (e.g., today it’s LLMs, tomorrow it may be something else), and I don’t want my thesis or skill set to become obsolete because it’s too tightly focused on a specific architecture.

At my university, I can either:

  1. Choose an existing topic, such as:

    • Testing and verification for neural networks and LLMs (fuzzing and ACT)

    • Mechanistic interpretability for LLMs

    • LLM safety via RL /​ adversarial training

    • Explainability in large language models

    • Bias in AI assisted decision making

  2. Or propose my own topic, as long as it’s feasible at bachelor level.

I’m currently leaning toward proposing something more model agnostic, for example:

  • reliability and failure mode analysis of learning based decision systems

  • specification based safety evaluation

  • runtime monitoring and incident detection for AI systems

  • bias and automation risk in human AI decision pipelines

My main question is:

Which research directions do you think will continue to be valued and in demand as AI systems evolve beyond current architectures and which ones are likely to age poorly?

I’m trying to make a thoughtful, long term choice rather than just chase what’s popular right now and i would greatly appreciate any insights and opinions.

No comments.