Why does AGI matter?

What benefits could AGI bring?

What risks does AGI pose?

Terrorism

Authoritarianism

Loss of Control

Summary

Three ways you can contribute

1. Use your career to reduce AI risk 

2. Advocate for sensible AI regulation 

3. Donate to impactful AI safety organizations 

Your AI Safety Action Plan

What can the most powerful AI systems do today?

The speed of progress is unprecedented

What does agency look like?

So... how do AI systems work?

What is AGI?

The Vending-Bench evaluation

How could AGI get built?

AI Agents at Work

Teaching AI the next action

Five years ago, AI systems struggled to form coherent sentences. Today, more than 5% of the world use AI products like ChatGPT every week for help with work, studies, and creative projects. These systems extend far beyond a simple chat. They can produce images and videos, write complex code, and control robots to do real-world tasks. 

This unit explores how AI is evolving from simple "tools" into autonomous "agents",  capable of setting goals, making complex plans, and acting in the real world.
AI systems can generate images, videos, songs, and websites from a single text prompt.

🎵 **Music**

Platforms like [Suno](https://suno.com/) and [Udio](https://www.udio.com/) generate songs from short descriptions like “a pop jazz song about sunshine”. Leave this playing while you explore the rest of this page! 
<Embed url="https://suno.com/embed/e02aebab-1fe8-4b64-b849-7393e01eeafd" className="!h-40"/>

🖼 **Images**

The first text-to-image AI system was built in 2015, producing tiny 32x32 pixel images. DALL-E was released by OpenAI in 2021, producing very basic photos. Today's AI-generated images are indistinguishable from those created by experts.

Explore more from [Midjourney](https://www.midjourney.com/explore), [DALL-E](https://openai.com/index/dall-e-3/), and [Stable Diffusion](https://stability.ai/stable-image).
<Embed url="/images/content/foai/midjourney.webp"/>

🎬 **Videos**

Four years ago, AI-generated videos didn't exist. Two years ago, they were terrible. Now, they're becoming hard to distinguish from real videos. 

<Embed url="https://www.youtube.com/embed/56jmufYqMjM?playlist=56jmufYqMjM&loop=1"/>
Explore more from [Veo](https://labs.google/flow/tv/) or [Sora](https://sora.chatgpt.com/explore/videos).

🔬 **Science and math**

AI systems outperform expert humans in answering science and math questions. One measure of this is [GPQA](https://arxiv.org/abs/2311.12022) - a test with extremely difficult science questions created by PhD experts. 

These "Google-proof" questions are so challenging that even skilled people with internet access only get 34% correct after spending 30 minutes on each question. The best AI systems get 87.7%.

**🛠 Building web applications**

AI can generate working code. Just tell it what kind of website or app you want, and it will build it. 

Try it yourself and create your own web app below! Type in your simple idea, or click on one of the examples. It'll take ~30 seconds to build.

<Embed url="https://course-demos.k8s.bluedot.org/generate-react-component" className="h-450"/>


Just five years ago, these capabilities didn’t exist. Now, breakthroughs are happening almost weekly.

Explore a time-lapse of model improvements since 2016:

[Open full visualization →](https://asteriskmag.com/issues/03/growing-up-overnight)

### 🤖 From tools to agents: AI’s next step
So far, most AI systems have been tools—responding only when prompted. Now we’re seeing early versions of “agents”, for example:
- **Tool:** A calculator that solves problems when you input them
- **Agent:** A research assistant who investigates topics & writes full reports

Agents can:
- Search the web
- Make decisions
- Act independently
- Complete tasks without step-by-step instructions

As their capabilities improve, they'll be given more license to roam the internet, control computers, and interact with the physical world via robot bodies.


**🌐 On the internet**

Imagine an AI assistant that manages your inbox. It doesn't just suggest replies; it reads your emails, understands what people are asking for, and responds on your behalf. 

It can even shop online: searching across different stores, comparing prices, and making purchases based on your preferences. Unlike a simple search tool, these agents take real actions online just like you would.

This now exists! [Try this agent-in-a-browser demo →](https://theaidigest.org/agent?state=N4IgxghgTgLiBcBtAugGnAewLYAcMGcBTAUSwgEsAbBUAIwwBMBPBEEdfAV1oCtCw48NuhgZWIAL7owGAHYA3QlHwQY5OfgSJQhMlU1JQ5BqwCMESaiMmhp2petmwktCGNn2b%2FAEEoYABbkijYAZhCUROjk%2BABKhDiUTOSyAOYIYRGE6FDxiQAiUBAhgnSMLEKeXLz8gsIgouISUjp6EVqOQgAsFs1uNiCd9hKu7l2e0b4BQYSh4ZFecQlJqelzWSA5SwVFJSD0zOIc3HwCh%2FViFU1WILoUbYZ9rABsPdejIE9DI%2F1P4z5%2BgWCq0yUViuWWaXgGXmm3yhWKND2ZTOVROtU8DUuvVu%2BnajyEAHYet9WAS%2FpNATNgfNootEslIdD1rCmNsEfBSgcKkdqqduedGtjWgZtPiQAAOYlRfri8kA6azEELcEM6nM8Fs3b7cp1VE1M6YthXFp3EUdEAATilYotcqmQKha1BdIhauyGvhWuR%2FL1fLqhskQtNePepgADNbQ2G7ZTFTSwUtVY6lSzNYjtSjjvr%2BQHjTdhSH%2BqYAEyRovFmMKt3KxMrZMwj07dPe3VZv0Yi5GoO4h6hgDMZbMfcrDqZzpVdbHG0b7M5OsqbfRIk7gdcYE4UByshgAAUIClCKx8KIcp515vCNud1BGJwBHuDwhTNIN1uYABlE%2BEB%2BHoQ4G8MHeMCaOgOKUO%2BhDQAEACKnBKPOoIAMLYHg%2BBJlO0Tvv4GAAO4Mu%2BsgQGAADWtDQNWGBQAwSgiq4%2F63gIEFQf43g4Ak5AzAAKkoWBnPRgGMZBALcVAvH8vxQH4DuSg%2FggfYcIQlA1DMiGqIQKSUeUsicJQlAKUpAgqXIijKKo6iyAAkjY2m6fpykMMhCg0WZchWbRdmGQwsEQNu5BqIQBjAFIIBEAZMAzO%2B5AAF4BTQwX4NhOEAPJUUoACyjDhNW%2BCESRZFQGlAUqI%2B%2FLyBxOGsMk9AAB6eDh0CyAygVNEAA%3D%3D%3D)

**🖥 In a virtual computer**

The above demo is in a constrained environment: just one email inbox and an online store. But AI agents are starting to control entire computers:
<Embed url="https://www.youtube.com/embed/jqx18KgIzAE"/>

**🦾 In the real world**

Robots powered by AI can now manipulate objects and perform tasks in the real world. Today, they're slower and make more mistakes than humans, but they're improving fast. 
<Embed url="https://www.youtube.com/embed/4MvGnmmP3c0"/>


Models like ChatGPT operate by predicting the next word in a sequence:
> **Input:** "The cat sat on the"
> 
> **Prediction:** "mat"

Doing this repeatedly allows the model to generate long, coherent paragraphs:
> **Input:** "The cat sat on the mat"
> 
> **Prediction:** ", watching"
> 
> […]
> 
> **Result:** “…watching the world outside with lazy indifference…”

It can also act as a helpful assistant:
> **Input:** “How much water does a basil plant need?”
> 
> **Prediction:** "Basil…"
> 
> **Full Answer:** “Basil loves a good drink but doesn’t like soggy feet…”

These models, while based on something as simple as predicting the next word, have shown surprising capabilities—writing code, solving problems, and creating stories.

<Callout title="Bonus: Want to learn more about how AI's trained?">
To learn more about basic machine learning concepts, check out our [beginners guide on how AI learns](https://aisafetyfundamentals.com/blog/how-does-ai-learn/).

Techniques to make AI systems even more powerful include [RLHF](https://openai.com/index/instruction-following/), [reasoning](https://openai.com/index/learning-to-reason-with-llms/) and [self-play](https://www.microsoft.com/en-us/research/blog/ai-self-play-for-algorithm-design/).
</Callout>


Think about where AI capabilities were 5 years ago, and where they are today. What do you think AI will and won’t be able to do in 5 years’ time?


5 years ago, 5 years ahead...

In your own words, how would you describe the difference between a ‘tool’ and an ‘agent’? Give one example of each


Tool vs. Agent

Future of AI

Learn what the most advanced AI systems can do today, and how they work.

What can AI do today?

Learn what AGI means, and why it might arrive within the next decade.

Grapple with how AGI could cure diseases and automate boring work, while also posing severe risks.

How could AGI change our lives?

Explore the challenge of building safe AGI amid heated global competition.