Attraction
AI Research Laboratory
A holographic assistant for architecture and code
Console
Ask the lab assistant
A fully client-side demo assistant with curated, genuinely useful answers about code, architecture, and learning paths.
Research radar
What the field is chewing on
Six themes currently driving real AI research, in plain language, no jargon left unexplained.
Retrieval-augmented generation
Grounds a model's answers in your own documents at query time, fetched from a vector database, instead of relying only on what it memorised during training. Cuts down hallucination and keeps responses current without retraining.
Agentic workflows
Models that plan multi-step tasks, call tools, check their own output, and loop until a goal is met, rather than answering in one shot. The hard part in practice is reliably knowing when to stop looping.
Mixture-of-experts
Splits a large model into many smaller specialised sub-networks and only routes each token through a few of them. Gets you a bigger effective model without paying its full compute cost on every request.
Multimodal models
A single model reasoning across text, images, audio, and video instead of separate models stitched together. The interesting research problem is a shared representation space where a photo and its caption land near each other.
Quantisation
Shrinks a model's weights from 16 or 32-bit precision down to 8 or 4-bit, trading a small accuracy hit for a large drop in memory and inference cost. Often the difference between needing a data-centre GPU and running locally.
Alignment and safety
Techniques like reinforcement learning from human feedback that steer a model toward being helpful and honest rather than just fluent. An active, unsettled research area, not a solved checkbox.