What people say after going through the cohorts

I came in knowing how to run training scripts but not what was happening when the loss stopped improving. The mathematics course changed that. By week four I was reading the gradient flow sections of papers that had previously been opaque. The problem sets were hard in the right way — they forced me to re-derive things I thought I had understood from the session, and the corrections were specific about where my reasoning had gone wrong.

Mathematics for ML · July 2025

The transformer architectures programme is the most demanding course I have taken. Twelve to fourteen hours a week is accurate, probably underselling weeks where a code review came back with substantial comments. But finishing it with a full reproduction of a published paper — one that I had to actually read, implement, and defend — felt different from any certification I have collected. The architecture defence in the final session was not comfortable, but it was fair and useful.

Transformer Architectures · June 2025

The research programme takes a while to find its pace — the first eight weeks of literature surveying and question scoping felt slow, and I was impatient for the experiments. Looking back, the time spent on that phase made the rest of it far less wasteful than it would have been otherwise. The writing clinics were genuinely useful. My supervisor's comments on the paper draft were direct, occasionally difficult to read, and almost always correct.

Research Practice · May 2025

I watched the full sample session before enrolling, which is the right way to make this decision. The teaching is methodical and expects you to follow along with a notebook open. If you want something faster and more survey-like, this is not it. I wanted to actually understand what eigendecomposition was doing in PCA and SVD, and after week three of the mathematics course I did. The derivation clinic was particularly good for catching the mistakes I had made and corrected badly.

Mathematics for ML · July 2025

I had been fine-tuning models through a platform at work for about two years before taking the architectures programme. The course made it clear how much I had been operating without understanding. The section on training instabilities alone was worth the cost. Being able to read the papers that are usually cited as justification for these techniques, and now understand them, is something I use every week.

Transformer Architectures · April 2025

The research programme is genuinely a commitment. Sixteen hours some weeks is not an exaggeration. My supervisor's fortnightly reading of my draft sections was thorough and honest in the way that useful feedback needs to be. The ethics review module was more rigorous than I expected — most resources in this area are either dismissive or theoretical, and this treatment was actually useful for the scraping decisions I was making in my own project.

Research Practice · June 2025