A Three-Day AI Bootcamp for 60 Teenagers: What Worked and Where Google Colab Failed

Can teenagers learn the foundations of artificial intelligence in three days? A pilot study from the education company Aiphabet Inc., posted on arXiv in 2023, says largely yes, and also reports what tripped students up. It is a company's own evaluation of a program built on its online learning platform, based on a self-selected survey, so it is best read as a case study rather than proof.

The bootcamp

• Who: 60 high school students aged 14 to 19, from several schools, who signed up through an online form. Math and physics were their most popular courses, and more students were beginners in AI than in programming.
• How long: three days of four-hour sessions, in person with an online platform for slides, animated videos, interactive playgrounds, Google Colab coding notebooks and quizzes, plus a live chat. Two instructors, a university computer science professor and a teaching assistant, taught it.
• Content:

What the results showed

Of 60 students, 35 answered a survey after the bootcamp.

• 91.4 percent rated the overall quality high or very high (4 or 5 out of 5).
• 88.5 percent said they understood AI concepts better, and 71.4 percent said the same of programming.
• Content, organization and ease of use of the platform rated well. The bootcamp's length rated worse, and a few students suggested more days with fewer hours each. Clarity also rated lower, probably because of the Colab programming modules.
• A "mark as completed" button on each topic served as a rough measure of engagement: 80.4 percent of students were highly or fully engaged. They engaged most with AI ethics and foundational topics, and least with deep learning and large language models.
• On quizzes, with up to three attempts per question, the cohort averaged 78.0 out of 100 (spread 12.4). They did best on foundational topics such as rational agents and the Turing test, and found ethics and the history of AI questions hard, even though they enjoyed history, especially early milestones.

Lessons the authors drew

• Mix the formats. Slides, video, demos and quizzes kept students engaged, and starting with foundations prepared them for harder topics.
• Live chat helped. It let students ask questions and even start discussions about AI's future.
• Google Colab was a poor fit. Students struggled with running cells, installing libraries and fixing errors, found it hard to collaborate, and instructors could not give instant feedback. The authors say teenagers need a better coding interface.
• Fill prerequisite gaps in probability, logic and mathematics, and many students used phones and tablets, so mobile access mattered.
• Completion is an imperfect proxy. Clicking "complete" does not prove understanding, so the team plans extra checks.

Why informal learning

The authors argue that bringing AI into school curricula will take time, because of obstacles such as teacher training, unequal computing resources, packed schedules and the lack of standard content, so informal programs outside the classroom may be the fastest way to spread AI knowledge. They conclude that teenagers can grasp AI concepts, and plan larger bootcamps with solid empirical evidence of learning outcomes and more emphasis on teamwork and creativity.

Sources

• Uzay Macar, Blake Castleman, Noah Mauchly and colleagues, Teenagers and Artificial Intelligence: Bootcamp Experience and Lessons Learned, arXiv:2312.10067 (2023), licensed CC BY 4.0. Changed here: condensed and reorganized, figures left out; rewritten in hubnx's own words.