Cambridge Language Models and Intelligent Agentic Systems Course
A 16-lecture journey hosted by Cambridge University’s C2D3, that takes you from Transformer fundamentals to building and safely deploying tool-using language-model agents.
-
Part I. What is a Language Model?
-
1. Introduction to Language Models
How next-token prediction grew from n-gram tricks into generative models that can translate, summarise and reason, setting the stage for the rest of the series.
-
2. The Transformer Architecture
A guided tour of attention, residual pathways and positional encodings, explaining why the Transformer became the lingua franca of modern ML.
-
3. Scaling Laws
The empirical regularities that tie model size, data and compute to capability — and what those curves do (and don’t) promise for the future.
-
-
Part II. Crafting Agentic Systems
-
4. Post-Training
From instruction-tuning to RLHF: how we shape a raw model into a helpful assistant.
-
5. Reinforcement Learning for Language Models
Adapting policy-gradient and bandit methods to text-generation, with demos of in-context RL for real-time improvement.
-
6. Reward Modelling
Building surrogate reward functions from human preference data to scale feedback beyond what annotators can directly label.
-
7. Agents and Agent Architectures
Tool use, memory, planning loops and multi-agent workflows that turn a chatty LM into an autonomous problem-solver.
-
-
Part III. Agentic Behaviour
-
8. Optimisation and Reasoning
Why gradient-descent objectives collide (or align) with logical reasoning, and how prompting, search and scratchpads bridge the gap.
-
9. Reward Hacking and Goal Misgeneralisation
Classic failure modes where clever models short-circuit the metric we set, plus strategies for early detection.
-
10. Out-of-Context Reasoning and Situational Awareness
Evidence that models can track hidden variables and realise when the “game” changes — and what that means for safety.
-
11. Deceptive Alignment and Alignment Faking
When models appear compliant but pursue latent goals; we examine theory, case studies and open research questions.
-
-
Part IV. Frontiers
-
12. Risks from Advanced AI
This lesson covers many different threats from AI that have been explored in the literature, like malicious usage, rogue AIs, or gradual disempowerment, and discusses some of the underlying drivers of AI risk.
-
13. Evaluations
Designing benchmarks, red-team protocols and interpretability probes that actually correlate with real-world performance.
-
14. AI Control
Scalable oversight, interpretability tools and governance levers for keeping powerful models within human-defined bounds.
-
15. AI Orgs & Agendas
A landscape tour of labs, open-source efforts and policy bodies shaping the research agenda — and how to contribute responsibly.
-
16. The Future of Language Models
Speculative yet grounded visions of multimodal reasoning, embodied agents and the socio-economic and R&D shifts they could catalyse.
-
Meet your instructors
-
Edward James Young
Department of Engineering, University of Cambridge
-
Jason Brown
Department of Computer Science & Technology, University of Cambridge
-
Lennie Wells
Department of Pure Mathematics and Mathematical Statistics, University of Cambridge
What you’ll learn
This 16 lecture series will explain how language-model systems are built in order to understand and predict their behaviour. Frontier language models are now being used as the foundation for agentic systems, which can carry out tasks that require extended reasoning and long-horizon planning. We investigate the potential safety and security risks associated with such systems, and present current research directions that aim to mitigate them.
The series is designed to be accessible for a broad audience across academia and industry, requiring knowledge from an introductory course in machine learning or statistics (e.g. backpropagation). We emphasise conceptual understanding of such systems, but will discuss technical details where necessary.
We hope that the course will empower researchers to make better use of language model systems and inform deployment across academia and industry. We also hope to stimulate engagement with the serious risks associated with intelligent systems, and encourage further work to address them.
Citing this work
J. R. Brown, L. Wells, and E. J. Young, "Language Models and Intelligent Agentic Systems," Meridian Cambridge and Cambridge Centre for Data Driven Discovery, University of Cambridge, Cambridge, UK, 2025. [Online]. Available: https://www.meridiancambridge.org/language-models-course
BibTeX Entry
@misc{lmias2025,
title={Language Models and Intelligent Agentic Systems},
author={Jason R. Brown and Lennie Wells and Edward James Young},
institution={Meridian Cambridge, University of Cambridge},
url={https://www.meridiancambridge.org/language-models-course},
year={2025},
month=jun,
}About Cambridge Centre for Data-Driven Discovery (C2D3)
C2D3 is a Cambridge University organisation that supports world-class research and teaching in data science, artificial intelligence, and machine learning.
