Technical AI Safety Seminar
AI capabilities are advancing faster than safety. Six weeks on closing that gap: training safer models, evaluating them, looking inside them and limiting the damage when they fail. For people with a technical background.
12 October to 22 November 2026. Places are limited: apply by 9 October.
Six weeks
12 October to 22 November 2026, one two-hour session a week plus an hour of material on your own
Mixed: in person and online
Sessions at Universidad de los Andes, with free food, and sessions online
Free, with a certificate
Materials by BlueDot Impact; discussion in Spanish, readings in English
Capabilities are advancing faster than safety
AI capabilities are advancing faster than safety, and the field needs more technical people working on it. The problems are still open: we don't yet know how to reliably measure what a model can and wants to do, how to understand what it computes inside, or how to guarantee that a capable system won't cause harm if something fails.
The Technical AI Safety Seminar is a tour of those problems and of the methods used today to try to solve them: RLHF, evaluations, mechanistic interpretability and AI control. The readings were designed by BlueDot Impact; AI Safety Colombia runs this cohort independently, with facilitators from the community.
Would you rather take it online and in English? Apply directly with BlueDot.
For people who already know how a model is trained and want to work on making it safe
You just need to understand the basics of how an LLM is trained; if you don't yet, BlueDot's two-hour AI Foundations course covers it. The seminar is designed for:
ML research
You do machine learning research and want to point it at alignment, evaluations or interpretability.
Software engineering
You build systems with models and want to move into working on the safety of the models you use.
Mathematics and computer science
You have the formal foundation and want to see which problems in the field need it.
Students who are already serious
You're thinking about fellowships such as MATS, SPAR or ARENA, or about graduate school, and want to arrive with judgement.
This isn't an introduction to AI. It assumes you already know how a model works and want the safety underneath.
If you'd rather start with the strategic picture, see the Seminar on the Future and Risks of AI.
A small cohort that discusses every week
Each week starts with an hour of light material (videos and podcasts) and continues with one two-hour session with your cohort: close reading and discussion. That is about three hours a week, at a time set with the selected participants. All readings and videos are in English, as BlueDot publishes them; the discussion is in Spanish. The format is mixed: some sessions are in person at Universidad de los Andes and some are online, and it fits alongside a job or university. The in-person sessions come with free food.
The facilitators are Helen Stefany Penagos (Mastercard) and Andrés Mosquera (Universidad de los Andes). The discussion is where assumptions get tested and where the readings connect to the work being done in labs and research organisations, and the cohort stays on as a community of peers to keep exploring the field with after the seminar.
Six weeks, from the underlying problem to your next step
Each week: an hour of light material and one two-hour session, close reading and discussion. Sessions alternate: the first week is in person, the second online, and so on to the end. Session times are set around the availability of the selected participants.
- Week 1In person
What it takes for AI to go well
Why today's models are not safe by default, and which technical approaches exist to change that.
- Week 2Online
Training safer models
How RLHF works, which other methods are used to fine-tune a model, and where they fall short.
- Week 3In person
Evaluations
How to measure what a model is capable of, and how to look for whether it is hiding something.
- Week 4Online
Mechanistic interpretability
The techniques for seeing what a neural network computes on the inside, and how far they reach today.
- Week 5In person
Minimising harm
How to limit what a model can do even if it is misaligned, with AI control and constitutional classifiers.
- Week 6Online
How to start contributing
Which open problems fit what you already know how to do, and how to take the first step.
A technical map, judgement, and a next step
A map of the technical approaches to safety and which problem each one attacks.
Judgement to read a paper in the field: what an evaluation measures, what an interpretability result shows, and what it doesn't.
Clarity on where your skills fit and what your next step is: a hackathon, a fellowship, or a project of your own.
A cohort of technical people in Colombia, and an AI Safety Colombia certificate.
Applications have closed
Applications for this cohort closed on Friday 9 October.
Frequently asked questions
How much technical background do I need?
You should understand the basics of how LLMs are trained and fine-tuned, that AI progress is driven by data, algorithms and compute, and that a neural network is optimised with gradient descent. BlueDot's two-hour, self-paced AI Foundations course covers it.
Do I need to take the Seminar on the Future and Risks of AI first?
It isn't required, but it helps: the Seminar on the Future and Risks of AI gives the context this seminar builds on. If you apply to this one and aren't accepted, the form lets you ask to be considered for the Seminar on the Future and Risks of AI.
Is there a selection process, or does everyone who applies get in?
There is a selection process: places are limited. We read every application and choose based on what you write and how well you fit the profile this page describes. We let everyone know, accepted or not, before the cohort starts.
When are the sessions?
The time is set with the selected participants: before the start we ask which slots work for them and fix the one that suits the cohort best.
What language is it in?
The discussions are in Spanish. All readings, videos and podcasts are in English, as BlueDot Impact publishes them, so you need to read English comfortably.
What do I need to receive the certificate?
Attend the discussion sessions: you can miss at most one. It's the same rule BlueDot Impact uses in its cohorts. The certificate is issued by AI Safety Colombia for completing the seminar, based on the BlueDot Impact curriculum.
How is this related to BlueDot Impact?
BlueDot Impact designed the curriculum and publishes its materials. AI Safety Colombia runs this cohort independently, with those materials and with facilitators from the community.