Resources: Projected AI progress

Exercises: Projected AI progress

Resources: Today's AI systems

Exercises: Today's AI systems

Resources: Distribution

Exercises: Distribution

Next steps: Programs

Resources: Automation

Exercises: Automation

Resources: Explosive growth

Exercises: Explosive growth

This unit examines how current AI systems might evolve into transformative artificial intelligence – technology that could radically accelerate economic and scientific progress. We'll analyze the key drivers: scaling laws that predict capability improvements, feedback loops in AI development (especially in R&D automation), and the interplay between compute, data, and algorithms.
The evidence for transformative potential comes from demonstrated scaling laws and early signs of AI accelerating its own development. We'll examine technical evidence suggesting rapid progress alongside fundamental bottlenecks that could prevent transformative outcomes.
For clarity, we'll focus on transformative AI rather than artificial general intelligence (AGI), as transformation tracks concrete outcomes rather than harder-to-define capabilities. When we do reference AGI, we'll use OpenAI's definition: "highly autonomous systems that outperform humans at most economically valuable work."

By the end of the unit, you should be able to:
\- Understand the concepts and potential of transformative AI 
\- Assess the relationship between scaling and capabilities
\- Model potential feedback loops in AI development (especially in R&D automation), and explain paths toward transformative AI
\- Understand the bottlenecks that might prevent transformative outcomes from AI


You’ve seen already that predicting future AI progress is hard. Experts and forecasters have messed up many times—for example, [drastically underestimating performance improvement on the MATH benchmark](https://bounded-regret.ghost.io/ai-forecasting-one-year-in/), or GMU Professor Bryan Caplan losing his bet that AI would not reliably get “A”s on his Labour Economics exam by 2029 in _[three months](https://www.betonit.ai/p/gpt-retakes-my-midterm-and-gets-an)_. On the other hand, over its long history the AI field has also seen previous waves of over-excitement followed by “[AI winters](https://en.wikipedia.org/wiki/AI_winter)”.

**Now it is your turn to have a try at AI forecasting!**

For each of the following dates, give a _concrete task_ that you think AI is 1% likely, 10% likely, 50% likely, and 90% likely to be able to do in the given years.

Dates:
- 2026-01-01
- 2028-01-01
- 2032-01-01

Examples of tasks that current AI (as of January 2025) can do might be:

- “Answer multiple-choice obscure PhD-level science questions about as well as human experts with internet access”
- “Beat 99% of human competitive programmers on competitive programming websites” (competitive programming problems are usually solved by expert humans in 20min-2h per problem, and similar to problems that software companies often give software engineers they are interviewing)
- “Make a humanoid robot walk, but not as smoothly as a human”
- “Having a voice-to-voice conversation with a human in real time”

Examples of tasks that current AI would fail:

- “Reaching human-level accuracy at <5min open-ended computer interaction tasks” (this is roughly what’s measured by the [OSWorld benchmark](https://os-world.github.io/), where before October 2024 top score was 8%, and as of January 2024 the top score is 38%, compared to a humans score of about 70-75%)
- “Solving a [Millennium Prize](https://en.wikipedia.org/wiki/Millennium_Prize_Problems) problem” (these are 7 extremely hard maths problems considered important by mathematicians, selected in 2000, of which only 1 has been successfully solved since then)
- “Writing and publishing theory papers in economics end-to-end, without human intervention.”
- “Adapting to perform any manual assembly task in a car factory that a human can do, based on seeing the same number of instructions and demonstrations that the human workers see.”


Economics of TAI

Exercises: Projected AI progress

Exercises