How to start doing real AI research before anybody gives you permission. A talk for students who want to work on AI safety and think they arrived late. It argues from the field’s own numbers that no one has a head start: the safety literature is still small enough to read in a semester, the subfield someone spent a PhD on can stop mattering in eighteen months, and the field cannot reliably label its own papers. The speaker’s own timeline, rejections included, is the case study. The bottleneck turned out to be sending the application, not the college, the CGPA, or the GPU.

An introduction to technical AI safety for a technical audience that is new to it. The talk starts with the vocabulary, because much of the confusion in the field comes from the same word meaning different things to different people: what alignment means and what it does not, what an evaluation actually measures, what a model organism is, and why interpretability and control are different bets on the same problem. It then walks through the main subfields one at a time: alignment and post-training, evaluations and dangerous-capability testing, AI control and monitoring, mechanistic interpretability, and biosecurity screening for biological models. For each one: the open problems, which of them can be worked on today with open models and a laptop, and where the people doing that work publish and meet.