Resources: Contributing to AI safety (ITAI)

Exercises: Contributing to AI safety (ITAI)

Next steps: Programs

Resources: What AI systems are people building and why?

Exercises: What AI systems are people building and why?

Resources: How AI systems work

Exercises: How AI systems work

Resources: The Promise of AI

Exercises: The Promise of AI

Resources: The Risks of AI

Exercises: The Risks of AI

This unit focuses on the foundational concepts in AI that you'll need to know to understand the rest of the course.
We'll learn the basics of AI: what it is, how it works, and what it can do right now. We’ll examine the crucial architecture behind deep learning (the leading approach to machine learning): neural networks. We'll also do a deep dive into large language models (LLMs), the technology behind most state-of-the-art foundation models like ChatGPT.

By the end of the unit, you should be able to:
\- Understand that ML systems are good at predicting outputs based on a set of inputs
\- Explain to a non-technical audience the structure of a trained large language model (LLM), and the base and fine-tuning LLM training stages
\- State that neural networks are trained with gradient descent
\- Explain why compute, data and algorithms are key to training ML systems


These questions should be answerable using the core resources above. Write your answers to the following questions in the box below. 
1. In a neural network, a \_\_\_\_\_ is a number that defines how much one neuron should influence another.
2. In a neural network, a \_\_\_\_\_ is a number that adds a constant amount to a neuron's activation.
3. What's the value of A, B and C in **[this neural network](https://docs.google.com/drawings/d/1raj4TGZG_F1mPOUKawHKiS-DaGULNFodjK-bTzxf5IE/edit)**?
4. What is "loss" in the context of training a neural network?
5. Why does minimising the loss improve the neural network's performance?
6. How does gradient descent work - i.e. minimise the loss? (at a high level)
7. Explain how this except from our website might be turned into data to train an LLM: “BlueDot Impact was founded in August 2022 in Cambridge, UK.”
8. What are the two main stages Karpathy splits up training an LLM into?
9. Describe the three factors which have contributed the most to AI improvements over the last few years.

You can check your answers against **[the answer key](https://docs.google.com/document/d/1vlObqXxvxPqcuIh2zFYFl0iAnHWFIzhOQ4Bmtv0kkXs/edit?usp=sharing)**.


Comprehension questions

Get a feel for how current AI models work! In this exercise, you'll try out several state-of-the-art models. We recommend playing around with 3-4 demos.

Before you start each demo, spend 1 minute trying to predict what you’re about to see or hear. How good do you expect it to be? What do you expect its shortcomings to be?

As you’re watching the demo, consider the following questions:
- Are the state-of-the-art AI models more advanced than you expected?
- How might these systems disrupt different industries?
- What capabilities might be coming soon?
- Based on today’s demos, what _can’t_ AI do?​

**Part 1: Large language model**
LLMs take text as input, and output more text. They're commonly used for tasks like chatbots. To complete this exercise, you'll need access to one. We recommend **[Bing Chat](https://www.bing.com/chat)** because it's free and doesn't require signup, but you can use any chat model you'd like such as **[OpenAI's ChatGPT](https://chatgpt.com/)**, **[Anthropic's Claude](https://claude.ai/new)**, **[Google Deepmind's Gemini](https://gemini.google.com/app)**, **[Inflection's Pi](https://pi.ai/talk)**, or **[Meta's Llama 3](https://ollama.com/blog/llama3)**.
Ask your chatbot for a recipe for your favourite food, but delivered as a rhyming poem.
1．What did it do well? What did it seem to struggle with?
Now ask it how gradient descent works, summarised in a few sentences. Ask it follow up questions if you want more information. (Tip: You can use AI models while on the course to help understand resources - but be aware that they sometimes make mistakes!)
2．Does this align with your understanding from the resources?
Now, ask your chatbot to calculate '13943642 × 1343' without using tools. Compare this result with a calculator.
3．Did it answer correctly? Why might it be difficult for an LLM without tools to answer these kinds of questions?

You can compare your results against **[our sample results](https://docs.google.com/document/d/11v2UxgUNOanub4DWiSIoOb5A_892zxX_K-HmDPuMXW8/edit#heading=h.qrtutjqgmx4v)**.
​
**Part 2: Text to image model**
Text to image models take text as input, and output an image. To complete this exercise, you'll need access to one. We recommend **[Bing Image Creator](https://www.bing.com/images/create)** because it's free, but you can use any image generation model you'd like such as **[OpenAI's Dalle-3](https://chatgpt.com/) **(integrated with ChatGPT), **[Midjourney](https://docs.midjourney.com/docs/quick-start)** or **[Stable Diffusion](https://clipdrop.co/stable-diffusion)**.
Ask your model for a picture of your favourite animal, playing your favourite sport on the moon.
4．What did it do well? What did it seem to struggle with?

**Part 3: Text to video**
Text to video models take text as an input and output an image. You can try [Open AI's Sora](https://sora.com/) (requires a free account, not available in some countries), [Google DeepMind's Veo 2](https://deepmind.google/technologies/veo/veo-2/) or watch [this video](https://www.youtube.com/watch?v=1BDqfPEhsCA) showing demos from several video generation models.

**Part 4: Text to Audio model**
Text-to-audio models takes text as an input and output as an audio file. We recommend [Udio](https://www.udio.com/home) (requires a free account to create songs, can browse examples without an account). You can also try the podcast generation features in [NotebookLM](https://blog.google/technology/ai/notebooklm-audio-overviews/).

Ask your model to create a song using the following prompt format: _A [genre #1] [genre #2] song about [topic]. _Uncheck “instrumental”, and set the version to v3. It should take 30-60 seconds to create.

**Part 5: Robotics**
There are several humanoid robots powered by AI. Here are a few examples:
- [Tesla](https://www.youtube.com/watch?v=DrNcXgoFv20)
- [Figure](https://www.youtube.com/watch?v=Sq1QZB5baNw)
​
**Part 6: Defence systems**
AI has been used in wars [for over a decade](https://disarmament.unoda.org/the-convention-on-certain-conventional-weapons/background-on-laws-in-the-ccw/). These demos show how far we've come with these systems:
- [Palantir Gotham](https://www.youtube.com/watch?v=rxKghrZU5w8) - an AI-powered defence system from 2021, used to help analyse data and make decisions
- [Anduril Bolt](https://www.youtube.com/watch?v=EEXI6r08908) - autonomous drones that can track and strike targets.
- [DARPA's RACER ](https://www.youtube.com/watch?v=5t3-eXHC5Zs)- autonomous tanks.


Intro to Transformative AI

Exercises: How AI systems work

Exercises