Detail

Large Language Models: From Next-Token Prediction to Reasoning Agents

In just a few years, large language models have undergone a transformation that few anticipated: from statistical text generators to systems capable of multi-step reasoning, tool use, and autonomous action. This talk traces that journey.
We begin with the fundamentals: how LLMs work and what the scaling laws tell us about their behaviour. We then examine three pivotal developments that redefined the field: inference-time computation (chain-of-thought and beyond), reinforcement learning with verifiable rewards, and the rise of agentic systems.
A central focus of the talk is on evaluation: how progress is measured, which benchmarks matter and why, and what the dramatic improvements in cost-efficiency over the past three years let us anticipate about the years ahead. We also look at the open-source landscape, where models from Alibaba and DeepSeek have altered assumptions about the resources required to reach frontier AI.
We close with coding agents: how the agentic loop works, what tools like Claude Code and similar systems are already capable of, and what this implies for the future of the software engineering profession and the broader society.