The good

Overall, I’m very glad I decided to take the Master’s. It gave me a first taste of research, and it turned out to be great fun! Trying new things, coming up with ideas, flexing the creative problem-solving part of the brain; after years of writing code where I more or less knew what I’d get in advance, it’s been refreshing to work on problems without predefined solutions. In a lot of professional work you’re expected to arrive at the answer people already know they want, but research feels honest. You don’t know the answer, but nor does anyone else, so I guess it’s up to you to figure it out.

A big selling point of the MSc turned out to be the other students. UCL’s course may not be perfect, but it does attract some sharp people. On the first day of the programme, I ended up chatting to coursemates who had just done things like graduating from maths at Cambridge or submitting their 5th patent. One guy had even done a PhD in maths. Being surrounded by so many intelligent, motivated people is really energising, and it’s one of the best aspects of the programme.

But perhaps the best aspect of a Master’s is time. Time to work on side projects, to read about new ideas or work on a thesis for three months uninterrupted. After working at startups for years, the breathing room has been precious. Reading and enjoying learning again is what ultimately convinced me to pursue a PhD—hopefully I won’t revisit this post and curse that choice in the future.

The bad

The teaching was, politely, variable. Some lecturers were excellent, and you got a clear sense they deeply understood what they were talking about: Marc Deisenroth teaching Gaussian Processes; John Shawe-Taylor teaching SVMs (which he wrote the book on); Jack Parker-Holder (now of 60 Minutes fame) teaching world models and robotics. But other lecturers were switched off, or sometimes even actively confusing. There’s a strong reliance on “flipped-classrooms.” That’s what they call using pre-recorded videos in place of normal lectures, with fairly dry in-person Q&A sessions. The professors who rejected the flipped classroom in favour of typical lectures usually did a better job. Their lectures were more engaging and elicited better questions.

The quality of the exams was also, politely, variable. I’d never really considered exam quality before, but you start to when they’re bad. Question-for-question repeats of previous papers, patchy coverage of the course content, or problems requiring concepts the lecturer hadn’t bothered to teach. I think this stuff is mostly symptomatic of disinterested lecturers, and I suppose you get those everywhere, but it was still a bit shabby. One exception was Brooks Paige’s excellent paper on Bayesian Deep Learning—a subject about which I originally had little interest—it had genuinely interesting questions, and you got to extrapolate from ideas you’d learned in fun ways. What if a VAE were also optimising a posterior distribution over the latent space? I have no idea. But it’s fun to work through the maths and figure it out.

The content at UCL definitely felt a bit outdated. Occasionally while doomscrolling in the wee hours I stumble across course videos from Stanford or MIT, like this excellent introduction to distributed training. Wow, so cool! Their homework is to implement Flash Attention from scratch. The lecture is given by a Meta researcher who actually works on this stuff! Well, there’s none of that sort of guff at UCL; no transformers, no diffusion, certainly no GPUs (not even in the “Applied” Deep Learning course). There’s a balance to be struck between teaching fundamentals vs. the newest flashness, but it does sting a little when you learn graphical model-based translation rather than the methods that replaced it 15 years ago…

The rest

Wow, grade inflation is actually real. Marks were noticeably easier to earn compared to my Bachelor’s. My highest mark in undergrad was around 80% for a course on General Relativity, which was a surprise for all involved. But at UCL loads of students rake in 90+% across multiple courses. It got to the point that after exams, before I’d written a word of my thesis, I was already guaranteed a distinction. I can’t say whether or not this really matters, but success did feel a bit cheaper.

Generative AI was absolutely ubiquitous. I reckon half the cohort’s coursework could have been submitted via ChatGPT share link—you’d even see jokes that people’s research topic was “whether ChatGPT can write a thesis.” To be clear, I also use generative AI. I don’t think there’s anything inherently evil about it, although I pared my usage back after finding myself too dependent. UCL’s system seemed totally unprepared for the scale of LLM usage.

Not to bow to stereotype, but it was a surprisingly antisocial crowd. Maybe it’s because it was only a one-year Master’s, or people already had friends in London, or the students were all avowed-workaholics (doubtful). Maybe I’m a poor conversationalist (plausible). Either way, it was a little isolating at times. Fortunately I made a few good friends to message and gossip and revise with. But I got the impression that making friends was a little uncommon.

A few words more

Regardless of the complaints—and I am a complainer—I believe my life has been much improved by this degree and the friends I’ve made during it. I’ve had a chance to pivot, and those don’t come by every day. It’s been a great year.