Capstone Project: An AI Research Brief

You have completed the AI Foundations track. You have traced the full machine-learning pipeline from raw data to a trained model. You have stepped inside a neural network and watched backpropagation adjust weights layer by layer. You have studied how large language models tokenize text, predict continuations, and can be steered through fine-tuning and prompting. You have grappled with the alignment problem, with fairness and accountability, and with the ways AI is reshaping industries and institutions. Now the question is: can you think like a researcher? This capstone is not a quiz or a fill-in-the-blank exercise. It is a structured investigation — a miniature version of the process that produces real AI research. You will pose a focused question, design a method for investigating it, run that investigation using the Institute labs and other available resources, and communicate what you found in a disciplined written brief. The habits you build here — precise questioning, honest reporting, acknowledgment of uncertainty — are the foundation of every serious AI research program.

Your Mission

Pose one focused, answerable research question about AI. Investigate it using a structured method. Write a short research brief — approximately 600 to 900 words — that communicates your question, your method, your findings, your limitations, and the open questions your investigation raised. The brief is the deliverable. The lab work, the experiments, the notes — those are the evidence behind it. A reader who has not done your investigation should be able to read your brief and understand exactly what you did, what you found, and how confident you are in your conclusions.

Phase 1: Pose a Question

A good research question for this capstone has three properties. First, it is focused. "How does AI work?" is not a research question. "How does increasing the number of hidden layers affect a neural network's ability to learn a non-linear function?" is a research question. Second, it is answerable with means you have. You cannot run a study on 10,000 users or retrain GPT-4. You can run controlled experiments in the Neural Net Internals Explorer, analyze publicly available model outputs, or compare documented system behaviors using structured observation. Third, it genuinely interests you. The best research questions come from noticing something you do not fully understand and wanting to understand it. Example questions across Tier 3 topics: ML Pipeline: How does the choice of training-to-test split ratio affect a model's generalization error on the same dataset? Neural Networks: How does varying the number of neurons in a hidden layer affect how quickly a network converges on the Neural Net Internals Explorer's binary classification tasks? Neural Networks: What happens to a trained network's decision boundary when you freeze the first layer's weights and continue training only the output layer? Large Language Models: How does prompt length and specificity affect the factual accuracy of an LLM's response on a topic you know well? LLMs and Attention: Does adding explicit chain-of-thought instructions to a prompt change how an LLM handles multi-step arithmetic — and if so, in what types of problems? Ethics and Alignment: In three documented AI deployment cases, what reward specification choices led to misaligned behavior, and what would a better specification have looked like? AI in Practice: How do three different AI coding assistants handle the same ambiguous requirement, and what does the variation reveal about their design priorities? Choose one question — or construct your own. Write it down precisely before moving to Phase 2.

Phase 2: Investigate

Your investigation must have a method — a clear, reproducible description of what you did and why. Design your method before you begin collecting observations. A method designed after the fact to explain results you already found is not a method; it is a rationalization. For questions about neural networks, the primary resource is the Neural Net Internals Explorer at /institute/lab/neural-net-internals-explorer. This lab lets you configure network architecture, set hyperparameters including learning rate, batch size, and activation functions, watch training unfold in real time, inspect weight changes layer by layer, and visualize decision boundaries. It is a genuine experimental environment. Experiments you can run there: hold architecture constant and vary learning rate across at least four values, recording loss at convergence; hold hyperparameters constant and vary architecture depth or width; observe how the decision boundary shifts as training progresses on a linearly inseparable problem; compare training curves between sigmoid and ReLU activations on the same task. For questions about LLMs or AI systems: design a structured prompt comparison, define your evaluation criteria before you begin, and apply the same criteria consistently across all cases. If your question involves fairness or alignment, select documented cases from published sources and analyze them against an explicit framework you define in advance. In all cases: record your raw observations separately from your interpretations. Note what you expected. Note what you actually found. If a result surprises you, run the condition again to check whether it repeats.

Phase 3: Write the Brief

A research brief has five sections. Each section has a job. Question and Background (approximately 100 to 150 words): State your research question precisely. Provide the background a reader needs to understand why the question is worth asking. This draws on the AI Foundations content you have learned — cite the relevant concept (backpropagation, tokenization, reward specification, etc.) and explain why your question probes something genuinely open or interesting about it. Method (approximately 100 to 150 words): Describe what you did in enough detail that another student could replicate it. Name the tool or environment. State the conditions you tested, the variables you controlled, and how you recorded observations. Be specific about quantities: not 'several learning rates' but 'learning rates of 0.001, 0.01, 0.1, and 1.0.' Findings (approximately 150 to 200 words): Report what you observed. Use specific numbers and descriptions where possible. Present results that support your hypothesis and results that complicate or contradict it. Use precise language: 'the network failed to converge' is more informative than 'it did not work well.' Limitations (approximately 100 to 150 words): State clearly what your investigation could not establish. Was your sample of conditions small? Did you test only one architecture? Is your evaluation of LLM outputs subjective? Limitations are not failures — they define the boundaries of your claim. A brief with honest limitations is more trustworthy than one that overclaims. Open Questions and Next Steps (approximately 75 to 100 words): What would you investigate next if you had more time or resources? What questions did your findings raise that you cannot yet answer? This section transforms your capstone from a closed report into the beginning of an ongoing inquiry.

The AI Foundations track began with a question: what actually happens when a machine learns? You have followed that question through data pipelines and gradient descent, through hidden layers and attention mechanisms, through alignment failures and real-world deployments. You have seen that AI is not magic — it is a set of techniques, built by people, shaped by choices, and subject to scrutiny. You have also seen that the field is genuinely open. The questions that matter most — how to build systems that reliably do what we intend, how to make AI accountable and fair, how to align powerful optimization processes with human values — do not yet have settled answers. The researchers working on those questions need people who can think carefully, investigate honestly, and communicate precisely. That is who you are now. Congratulations on completing AI Foundations.