---
title: 'Can LLMs Actually Do Things? (Not Just Talk)'
source: 'https://youtube.com/watch?v=6NjVnSxJCTY'
video_id: '6NjVnSxJCTY'
date: 2026-09-16
duration_sec: 70
channel: 'Stanford Online'
---

# Can LLMs Actually Do Things? (Not Just Talk)

> Source: [Can LLMs Actually Do Things? (Not Just Talk)](https://youtube.com/watch?v=6NjVnSxJCTY)

## Summary

This video introduces the concept of AI agents, which are systems that can take a goal, reason about steps, use tools, and adapt based on outcomes. It contrasts them with traditional conversational AI, highlighting their action-oriented nature. The video also outlines key challenges in the field, such as reliability, evaluation, and planning, and mentions a course (CME 295) that covers these topics.

### Key Points

- **From Language Models to Agents** [00:01] — Large language models excel at understanding and generating text, but the next step is enabling them to take action. Agents are systems that can take a goal, reason about needed steps, use tools, and adapt based on what happens.
- **Agent in Action** [00:26] — Instead of just answering from knowledge, an agent might search for information, write code, call an API, analyze results, and decide what to do next. This moves AI from conversational to action-oriented.
- **Key Challenges for Agents** [00:38] — The shift to agents raises important challenges: how to make them reliable, how to evaluate them, and how to ensure they can plan without going off track.
- **Course Overview** [00:51] — The video mentions that these ideas are covered in CME 295, 'Transformers and Large Language Models,' grounding them with the latest technical advances.

### Conclusion

The video presents AI agents as the next evolution of language models, moving from conversation to action. It highlights the key challenges of reliability, evaluation, and planning, and points to a course for deeper learning.

## Transcript

Over the past few years, large language models have become very good at understanding and generating text. But the next question is, can these models actually do things?
This is where agents come in. An agent is a system that can take in a goal, reason about needed steps, use tools, and adapt based on what happens along the way.
For example, instead of answering a question only based on knowledge, an agent might search for information, write code, call an API, analyze the result, and decide what to do
next. And this is exciting because it moves AI systems from being mostly conversational to being more action-oriented, but it also raises important challenges. First, how do we make agents reliable?
How do we evaluate them? and how do we make sure they can plan without going off track? In CME 295, Transformers and Large Language Models, we cover these ideas by grounding them with the latest technical advances.
Shervin and I hope to see you there!