Illustrative Mechanisms You Could Study

Below are illustrative examples of mechanisms you might investigate.

These are suggestions, not restrictions.
You may choose other systems if the mechanism is clearly defined and structurally identifiable.

Application-Based Mechanisms

■ Social Media Feeds (e.g. TikTok / Instagram)

▪ Ranking systems

    • Content in your feed is ranked before being shown.
    • Platforms prioritise posts using signals such as engagement, recency, or predicted interest.
    • Study how ranking determines what appears first and what becomes visible.

▪ Engagement feedback loops

    • User actions (likes, watch time, shares) reshape what the system shows next.
    • Study how interaction data updates and reinforces future recommendations.

■ Streaming Platforms (e.g. Spotify / Netflix / YouTube)

▪ Similarity-based recommendation

    • The system suggests content similar to what users previously consumed.
    • Study how similarity is computed using behaviour patterns or learned embeddings.

▪ Exploration vs reinforcement

    • Platforms must balance recommending familiar content and introducing new items.
    • Study how systems prevent feeds from becoming too repetitive or predictable.

■ Conversational AI (e.g. ChatGPT)

▪ Response selection

    • Language models generate many possible continuations and must select one to output.
    • Study how probability ranking or sampling influences the final response.

▪ Reasoning vs standard generation

    • Some systems introduce reasoning steps or structured prompting instead of direct completion.
    • Study how reasoning pipelines differ from simple next-token generation.

▪ Context window limits

    • Conversational systems remember only a limited amount of prior text.
    • Study how limited context affects long conversations and response quality.

Model Architecture Mechanisms

■ Attention-based models (Transformers)

▪ Learning relationships across information

    • Some models analyse input by allowing different parts of the data to attend to one another.
    • Study how this structure helps models capture long-range relationships in language or data.

■ Adversarial models (GANs)

▪ Learning through competition

    • One model generates outputs, while another evaluates them.
    • Study how this competitive interaction improves the realism of generated images or media.

■ Encoder–decoder models

▪ Compressing and reconstructing information

    • One component converts input into a compact internal representation, while another reconstructs the output.
    • Study how this structure enables tasks such as translation, summarisation, or reconstruction.

■ Diffusion-style models

▪ Gradual refinement

    • Some generative systems create outputs through many small refinement steps rather than a single prediction.
    • Study how iterative refinement produces detailed images or media.

Technique-Based Mechanisms

■ How systems represent information

▪ Encoding meaning into numbers

    • AI systems convert text, images, or behaviour into numerical representations.
    • Study how vector representations enable similarity search, clustering, or retrieval.

■ How systems learn what to prioritise

▪ Training signals and objectives

    • AI models improve by adjusting themselves to reduce error or maximise a reward signal.
    • Study how different objectives influence what the system becomes good at.

■ How systems update themselves during learning

▪ Iterative improvement

    • Models repeatedly adjust internal parameters based on feedback from previous predictions.
    • Study how repeated updates gradually improve system performance.

■ How systems generate outputs step-by-step

▪ Sequential generation

    • Many AI systems produce results one step at a time rather than all at once.
    • Study how sequential generation influences creativity, randomness, or consistency.

Key Advice

Do not attempt to analyse everything.

A strong submission focuses on one clearly identifiable mechanism and explains how it shapes system behaviour.