About this course
A serious, interdisciplinary seminar for students who care about the stakes of AI and want to engage with the questions leading researchers are actually working on. Framed around a “for humanity” mission and anchored in the work of Cambridge's Leverhulme Centre for the Future of Intelligence and the Centre for the Study of Existential Risk, the course moves from foundations — what the alignment problem is, how today's models are trained and aligned, scalable oversight, interpretability, governance, and ethics — toward each student making something and producing an original, argued, properly cited research paper by the end. It draws on the field's foundational papers and on widely used curricula from BlueDot Impact and 80,000 Hours. No programming is required, though technically inclined students can take their project in a technical direction.
Is this the right fit?
- Who it’s for
- Thoughtful students aged 14–17 who care about the societal stakes of AI and want to engage seriously with alignment, governance, and ethics.
- Prerequisites
- Strong reading and writing, and a willingness to engage with challenging academic and technical texts. No coding or advanced math required; Introduction to Machine Learning is a helpful but optional companion.
- Tools students use
- Readings from CFI/CSER, BlueDot Impact, 80,000 Hours, and foundational AI-safety papers
Class schedule
Two weekly time-slot options, so students in Asia, the Gulf, and the Americas can all join live. Pick the one that fits; a coach confirms it when you enroll.
Times convert to your zone automatically and account for daylight saving. A coach confirms your slot when you enroll.
What students walk away with
- A working grasp of the alignment problem — reward misspecification, specification gaming, Goodhart's law
- How today's models are trained and aligned: RLHF, scalable oversight, and Constitutional AI
- The landscape of interpretability, AI governance, and the ethics of automated decision-making
- The ability to read foundational papers and build a rigorous argument with honest treatment of uncertainty
- An original, argued, properly cited research paper, defended in a final seminar
What a session looks like
- 0:00
Reading in focus
A short framing of the week's assigned reading and the question it raises.
- 0:20
Seminar & debate
Structured discussion and debate — steelmanning positions, not just stating them.
- 1:20
Research workshop
In the second half of the term, supervised work and peer review on each student's paper.
Curriculum
Week 1 — Why AI safety, and “for humanity”
The trajectory of capability; near-term harms vs. long-term risk. Introducing the Leverhulme CFI and CSER framing.
Week 2 — What is the alignment problem?
What we want vs. what we get: reward misspecification, specification gaming, and Goodhart's law. Reading: Concrete Problems in AI Safety.
Week 3 — How models are trained & aligned
Supervised learning to language models to RLHF; what fine-tuning and instruction-tuning actually do.
Week 4 — Scalable oversight
Supervising systems too capable to easily check: debate, recursive reward modeling, and Constitutional AI.
Week 5 — Robustness, control & misuse
Adversarial examples, jailbreaks, distribution shift, and control as a complement to alignment.
Week 6 — Interpretability
Why we can't yet read a model's “mind,” and why it matters: mechanistic interpretability, features, and circuits.
Week 7 — AI governance & policy
Standards, evaluations, compute governance, and international coordination, drawing on CSER and CFI's work.
Week 8 — The ethics of AI
Fairness, bias, transparency, accountability, and value pluralism — whose values should AI align to?
Week 9 — Existential risk & the long view
Superintelligence, instrumental convergence, and steelmanning the skeptics. Readings from Bostrom and Christian.
Week 10 — From reader to researcher
Choosing a tractable question and what makes an argument in AI safety rigorous. Students lock in their paper topic.
Week 11 — Build your argument
A drafting workshop — thesis, evidence, counter-arguments, uncertainty — with a peer-review round. Technical students may add a small demonstration.
Week 12 — Publish & defend
Students present their papers and take questions in a seminar defense, and discuss sharing work responsibly.
What’s included
- 12 live small-group seminars, capped at 6 students
- A curated reading list drawn from CFI/CSER, BlueDot Impact, and foundational papers
- Structured discussion, debate, and peer review
- Supervised research and writing in the second half of the term
- An original research paper, defended in a final seminar
How to join
- Purchase your seat. Choose your time-zone slot at checkout and pay securely. Classes are capped at six, so seats go in the order they’re booked.
- We confirm the details. After checkout a coach confirms your slot and asks a few quick questions so the first class is pitched right for your child.
- Sit the first class risk-free. If it’s not the right fit, tell us after that class and we’ll refund you in full.
Not sure yet? Book a free consultation or email hello@redwoodacademy.ai and a coach will help you choose.
Frequently asked questions
Does my student need to code?
No. The course is a reading-and-writing seminar and requires no programming or advanced math. Technically inclined students can take their final project in a technical direction — for example reproducing a specification-gaming example or a simple interpretability probe — but that is optional.
What is the reading like?
Serious but accessible: foundational papers like Concrete Problems in AI Safety, work from Cambridge's Leverhulme CFI and CSER, and units from widely used curricula by BlueDot Impact and 80,000 Hours. Most readings are light on math.
Why is this class two hours?
It runs as a seminar: framing the reading, structured discussion and debate, and — in the second half of the term — supervised research and peer review on each student's paper. Two hours gives that room to breathe.
What is the deliverable?
An original, argued, properly cited research paper on an AI-safety, alignment, or ethics question of the student's choosing, presented and defended in a final seminar — a genuine intellectual credential for applications and research programs.