What might AGI look like, concretely?
The term AGI consistently inspires different visions for people. I've seen this while leading specialised AI courses for thousands of people.
Most definitions are simply too abstract. For example, OpenAI defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." This doesn't give most people a strong intuition for what it looks like.
In this short, jargon-free post, you'll develop the same understanding of AGI that many experts have: an extremely-competent, fast and copyable remote employee.
Expert knowledge and reasoning
Models like ChatGPT can already:
- 🧑‍⚖️ pass the bar exam to become a lawyer
- 🎓 answer PhD-level multiple-choice questions better than domain experts
- 🧑‍⚕️ beat generalist doctors at answering medical queries
This doesn't stop at memorising facts - ChatGPT solves programming and math puzzles better than almost all humans, and translates passages as well as junior professional translators.
However, these are just a few disciplines. AGI would continue this trend across a much broader range of tasks.
Have a question for an expert ? AGI will answer at least as well as they could, if not better.
Tool usage
AGI won't be just a better question-answering machine - that's not what most economically valuable work is about. Instead, AGI would be able to use tools to interact with the world.
Today's systems can use computers, although they make mistakes and move a bit slowly:
AGI would instead nail computer use, as you'd expect from a competent remote employee. AGI could also connect to systems directly to operate many times faster than operating with a computer would allow them to.
End-to-end execution
Combining its smarts and ability to use tools, AGI would carry out long tasks end-to-end.
This would look like being able to give a task to an AGI like "build Uber for dogs" - and it would:
- raise money from venture capitalists
- hire other AI agents with different skills
- coordinate work and make strategic decisions
- continue pursuing this effectively over months or years
- potentially realise Uber for dogs isn't the great idea you think it is, and pivot
(This is meant to give a sense of AGI's capability, rather than how society would likely use it.)
This contrasts sharply with current AI that can't manage a simplified simulated business for more than a few months. Although systems can achieve smaller goals e.g. setting up an Uber for dogs landing page:
Video: You can see the AI thinking and taking actions on the left. Coding sequences have been sped up.
Speed, scale and stamina
While not inherent properties of AGI, its digital nature means that AGI will likely have incredible speed, scale and stamina.
Speed
AGI could digest information, produce outputs and take actions at superhuman speeds.
On specialised hardware, today's models can reach speeds of up to 1,347 tokens per second. (A token is a fragment of text, roughly the size of a word or punctuation mark.)
AI vs Human Output Speed
Scale
Not only will AGI be fast doing one thing, it might be fast on many things at once. AI systems can be copied and run in parallel: so if you need another , just run another copy of the AGI!
Every second, ChatGPT responds to over 11,574 messages. That's 0 since you opened this page.
Stamina
AGI would have infinite stamina: no sleeping, holidays, or calling in sick. Organizations built from AGIs wouldn't stress about staff turnover, and training new hires would be a simple copy-paste job.
Image: Best of all, AGI would never leave leftovers to go bad in the office fridge.
Summary
To recap, AGI would look like the perfect remote employee:
- Expert knowledge and reasoning - Matching or exceeding human experts in virtually all knowledge domains
- Tool usage - Using computers as well as an experienced remote employee, plus integrating with systems directly
- End-to-end execution - Handling complex, multi-stage projects autonomously over extended timeframes
It's likely that it would operate with superhuman...
- Speed - being many times faster than the fastest humans
- Scale - doing thousands or millions of tasks in parallel
- Stamina - never taking a break, getting tired or calling in sick
Bonus: Remember what you've learnt
Want to remember this, and know how to apply it in the real world? Learning research shows deeply processing material makes a big difference.
Spend 1 minute answering each of:
- Pick a random friend or colleague. How would you now explain AGI to them?
- What are the key differences between today's ChatGPT, and the described vision of AGI?
- Which of the three digital advantages—speed, scale, or stamina—do you think will most affect how impactful AGI is, and why?