Artificial intelligence & machine learning
Cross-disciplinary ecosystem spanning foundation models & generative AI, machine learning & deep learning, and reinforcement learning.
What's happening now
Leading organisations
The frontier labs, companies and research institutes building the models — and what each one ships.
Understanding the ecosystem
Model families
The model families that define the field — and the labs and companies that build them.
Anthropic's frontier assistant family, tuned for safety and reasoning.
Google DeepMind's natively-multimodal frontier family.
Google DeepMind's open-weight models, derived from Gemini research.
Meta's open-weight family; the backbone of much of the open ecosystem.
Mistral AI's efficient dense and mixture-of-experts models.
DeepSeek's efficient open-weight models, incl. strong reasoning variants.
Microsoft's small language models trained on textbook-quality data.
Benchmarks & evaluations
The evaluation suites the field is measured against, by what they test. Identities only — not a live leaderboard.
Massive Multitask Language Understanding — 57-subject knowledge test.
Harder, reasoning-focused successor to MMLU with more distractors.
Stanford's Holistic Evaluation of Language Models — multi-metric framework.
A framework / living leaderboard rather than a single test.
Graduate-level Google-proof Q&A in science; resists web lookup.
Contamination-resistant benchmark with continuously-refreshed questions.
Abstraction & reasoning corpus; probes fluid, few-shot generalisation.
Competition mathematics problems requiring multi-step solutions.
American Invitational Mathematics Examination — a frontier reasoning yardstick.
Function-completion coding benchmark from OpenAI.
Resolve real GitHub issues end-to-end; the agentic-coding standard.
SWE-bench Verified is the human-validated subset.
Reference
Papers shaping this ecosystem
The foundational work this field is built on, in chronological order — the architectures, methods and milestones that made modern AI.
Goodfellow et al. · 2014 · NeurIPS
Adversarial training launched modern generative modelling.
He et al. · 2015 · CVPR
Residual connections made very deep networks trainable — the backbone of modern vision.
Mnih et al. · 2015 · Nature
Deep Q-Networks fused deep learning with RL, reaching human level on Atari.
Silver et al. · 2016 · Nature
Deep RL plus tree search beat a Go world champion — a watershed moment for AI.
Vaswani et al. · 2017 · NeurIPS
Introduced the Transformer — the architecture behind virtually every modern LLM.
Shazeer et al. · 2017 · ICLR
Sparsely-gated mixture-of-experts — conditional compute that scales models to trillions of parameters.