ML misfits club or what to do when nobody wants you

TL;DR This summer I did not get into any of a bunch of research programs and schools. After crying into my pillow, I remembered someone smart saying that knowledge isn't something you're given; you have to take it. That seemed like a bright idea, so I am starting my own summer-ish school / research-padawan community with blackjack and hookers. An educational-research collective for the misfits (upd: backup link). Expect deadlines, homework, and most importantly other equally obsessed people right beside you.

Background and other blah-blah-blah

Hi there! I'm Andrey. For almost ten years I shuffled JSONs, then had an existential crisis and rolled into AI research by starting a Master's at Skoltech. After an amazing year (pinky promise!), surrounded by smart, high-energy folks and tons of ML, I started pondering what comes next. Slowing down was not an option; I wanted faster, higher, stronger - for the glory of ROC-AUC and accuracy! Applied to multiple research intensives, summer schools, internships, and extra-curricular programs… and got into exactly none of them.

Specifically:

• Rejected for internships and research-engineer roles at Google, DeepMind (to-may-to - to-mah-to?), Nvidia, Apple, Microsoft, Amazon, and Meta. Even with referrals. Even after spending my last 3 years before Skoltech as a software engineer at Meta.

• Was not even considered for a quant hotshot at Jane Street, HRT, XTX, and the rest of the shops that sacrifice to the almighty Alpha.

• Didn't make the cut for MATS or LASR (AI-alignment bootcamps).

• Flunked interviews at Anthropic and Huawei. Technically Huawei never replied after the take-home assignment. Maybe they're still grading it. It's only been three months.

• Did not make the cut for Yandex School of Data Analysis (top-tier tech program in Russia with a cutthroat selection process). Again. Not the brightest fella, huh?

They did actually admit me to the one-week M2L school, which I'm super happy about, but that hardly covers my long-term craving for structured growth.

There I was, ready to drown my sorrows in my usual bucket of ice cream, chewing on a deep dissatisfaction with the current state of things, when a (maybe) bright idea hit me. What if this is the first time I'm truly outside my comfort zone professionally (in personal life I managed to get out of the comfort zone after being outside of the comfort zone already)? If so, I'm exactly where I'm supposed to be. Lack of success just marks the growth zone. And growth means every failure is an opportunity.

Ok, let's re-frame. It's not that they didn't take me anywhere; it's that I just won the chance to design the most flexible learning program imaginable, level-up my self-organization, grow my network, and maybe help a few equally lost souls.

The Cunning Plan

Can't get into top-tier programs? Fine, I'll make my own.

Learning rests on three pillars: materials, community, homework with deadlines.

Materials and homework are easy. Kind of. Lots of MOOCs are lacking rigorous homework assignments (and I believe one should do at least twice as much practice compared to the theory to master anything). However, one can always compile a decently looking Frankenstein out of multiple sources. Not to mention that multiple courses from that Yandex School of Data Analysis are available for free. With complex homework assignments as required. Plus tons of lectures from every other major university.

Community is hard - and the most valuable. Sometimes I stall where a buddy flies; sometimes I'm just out of steam and keep working only thanks to a shared mission. Like-minded people rock. That's exactly why you are reading this post.

Action items:

• Grab a bunch of courses, add to them solid homework, start grinding.

• From 30 June we meet weekly: online via Zoom and in-person in London (extra points if someone spins up offline hubs elsewhere).

• Before each meeting we cover a chunk of theory, do the homework, then discuss what was unclear.

• Cross-review each other's homework - who else is going to grade it?

• Leave behind notes with exhaustive walkthroughs of tricky spots and share annotated solutions. No more "proof left as an exercise".

What's on the menu:

• A math-for-ML-research course covering “everything-everywhere-all-at-once”. I also want to blend classic linear algebra with Oseledets' godlike Numerical Linear Algebra. Audited it last winter, comprehended maybe half of what it can offer, crave more.

• Yandex's “ML Handbook” - back to basics.

• The ARENA curriculum - an awesome buffet of LLM alignment & interpretability.

• Karpathy's legendary “Neural Networks: Zero to Hero” series.

No need to take every course with me. Jump in only where you care. You can also propose your own course and run it. Vox Populi, Vox Dei?

What about research?

Simpler and messier. Simpler because I already have a lab at Skoltech with a fantastic advisor. Messier because I want to see how research ticks in other teams and help others poke their own research itch. We're building a community here, right?!

Let's self-organize that too. Among first-year master's classmates alone I saw no less creativity and drive than among professors.

The problem: it's hard to rally a team around one idea when there are a gazillion possible rabbit holes. My best idea so far: shortlist interesting profs/labs in town and start sending cold emails. Sadly, it does not extrapolate to multiple people, hence an attempt to self-organize.

Sold. How do I join?

Ta-da! (upd: backup link)

That GitHub repo already has (or soon will):

• Links to lectures/books and other study materials

• Post-discussion notes

• Homework assignments + spoiler-tagged solutions

• Links to chat(s) for coordinating in-person meetups

• One GitHub Issue per course we're dissecting

How to use it:

1. Introduce yourself in the thread whois.

2. Pick an interesting course, subscribe to notifications, join calls or IRL sessions.

3. If your dream course isn't listed, spin up your own and recruit people.

4. Find a co-working location or set one up.

5. Post or browse research ideas here.

Quirks:

• We'll prefer English-language content so any reject from anywhere feels at home. Quality still rules though. So, you will see some courses offered in Russian there. For instance, Yandex's ML handbook. You are more than welcome to join even if you are struggling with Russian, but the translation of the content is up to you. Sorry :(

Hooray! We've built a home for the misfits. Let's ship it to prod. See you on the first calls!

Got questions, proposals, or just want to talk? Hop onto the kick-off meeting—or come in person if you're in London.