Most free online education is free because it is cheap to produce, which usually means video plus a quiz and nobody on the other end. BlueDot's courses are free for a different reason, which is that a non profit decided the bottleneck in AI safety was people who understood the field rather than money, and built something that costs real money to run and gave it away. The difference shows in the format. Every session you read and write for two to three hours on your own, then you sit in a two hour Zoom discussion with roughly eight peers led by someone who actually works in AI safety.
That structure is expensive and it is also the single most effective way to learn a contested subject, because alignment is not a field with settled answers you can memorise. It is a field where reasonable people disagree about what the problem even is, and arguing about it with a facilitator who can tell you why your intuition is wrong is worth ten times the equivalent hours of watching lectures. There are now several tracks. The technical course covers alignment and RLHF, mechanistic interpretability, evaluations and red teaming, AI control and scalable oversight, and it expects you to already understand roughly how a language model is trained and what gradient descent is doing.
There is a short free self paced foundations course if you do not. The strategy course is the one I would point most people towards, and it is less about maths and more about how you reason under deep uncertainty: the incentives operating on AI companies, threat modelling in the kill chain sense, and evaluating interventions on defence in depth grounds. There are also project sprints for people who have done the reading and want to produce something, and a biosecurity strand that sits outside the scope of this site. Both main courses come in a compressed few day intensive or a five to six week part time version with identical content, which is a sensible accommodation for people using annual leave versus people fitting it around a job.
What makes it work, beyond the facilitation, is the curation. The AI safety literature is a sprawl of blog posts, arXiv preprints, forum comments and lab publications of wildly varying quality, and a newcomer trying to self study it will spend a month reading things that turned out not to matter. Someone has done that filtering for you, and kept it current, which for a field moving this fast is a substantial gift. I have seen self study lists for AI safety and none of them are as good as this, partly because a list cannot tell you which disagreements are live.
Now the reservations, and I want to be direct about the biggest one because a lot of writing about these courses is not. AI safety as a career has far more enthusiastic applicants than positions. The frontier labs, the government safety institutes and the handful of research non profits hire small numbers of people, often with strong research track records, and a five week course does not make you competitive for those jobs on its own. BlueDot is honest about being a starting point rather than a pipeline, and the alumni outcomes are genuinely good relative to the field, but you should go in with clear eyes.
If your plan is course then job, adjust it. The realistic value is that you find out quickly whether you actually want to do this work, and if you do, you leave with a map, a reading habit and eight people who are in the same position as you. The practical frictions are worth knowing too. Applications are competitive and places are limited, so this is not something you decide to do on a Sunday evening.
The calls are scheduled, live and timezone bound, which rules out a fair number of people through no fault of their own. And the reading is genuinely load bearing. If you turn up to a facilitated discussion having skimmed the papers, you will contribute nothing and you will make the session worse for everyone else, which is a stronger obligation than most courses put on you. That is a feature rather than a bug, but it does mean the honest prerequisite is time rather than knowledge.
One last thing I appreciate: nobody here is selling you a certificate. There is no badge being positioned as the outcome, no LinkedIn credential push, no tiered pricing where the useful version costs money. The product is understanding, and the organisation appears to believe that is enough. Given how much of the AI education market is the exact opposite of this, I am giving it 4.4 and the only reason it is not higher is the structural mismatch between how many people want to work in this field and how many roles exist.