I’ve been in a bit of a limbo-land work-wise for the past year. My wife and I knew we were moving to the US at some point (the planned date was January 2023) but my visa application progress was slow and ended up being delayed longer than expected. The uncertainty around our plans coupled with some other logistical issues left me reluctant to dive into full-time employment, so I resigned myself to a few more months of messing about and making more free educational material like the Hugging Face diffusion models class and the latest fast.ai course I’d been involved in. But then Suhail got in touch with an offer I couldn’t resist.

Suhail tweeting out invitations to be a Builder in Residence (the Jan 7 tweet is about yours truly if I remember right)

This was the pitch: come work for Playground AI for the next few months, part-time, as a “builder in residence”. I could earn some cash and get my hands dirty building some cool stuff for real users, and I’d still have 2 days a week to work on my other projects. Sounds great, right?! Anyway, six months later I’m happy to report that it was indeed an excellent experience :) In this post I’ll share a little about what I liked, what I learned, and why you should consider something like this in your own company/career.

What I Liked

Playground is nice! I enjoyed getting to know the team, and they’ve cultivated a great culture that felt welcoming even when working remotely from the other side of the world. One thing I’d been missing in my recent solo endeavors was having other people to bounce ideas off of and learn from.
They ship - and getting things in front of actual users is thrilling. It’s one thing to write a little tutorial and hope some people read it. It’s another to train a model and then see it being used by real people to make millions of images! I was very impressed with how fast the team went from a paper or proof of concept to something being integrated into the product.
Part-time is amazing. Obviously :) I had plenty of time for non-work things like errands and hobbies and was able to do lots of things I couldn't have managed if I was working somewhere full-time.
Suhail is great. I was initially wary - when I started there were hundreds of founders diving into Stable Diffusion, and I was worried this was just another ‘startup bro’ leaping onto what they hoped was the Next Big Thing (TM). Needless to say, I’ve revised my opinion and now consider myself very fortunate to have worked with him. I learned a lot about startups, and he learned a lot about diffusion models, so I think it was beneficial both ways 🙂 It was fun watching him earnestly dive into the technical side of things - I don’t expect many CEOs to be so familiar with their companies’ codebase or research work!

What I Learned

Time ≠ capacity, and context switching is hard. Working Monday-Wednesday meant in theory I could do my research work with Jeremy on Thursday and work on my book on Friday… except that in practice it rarely worked like that. When I was hyper-focused on something for work it was very tempting to keep checking in and running experiments all week long, and even if this didn’t add up to many hours it often used up pretty much all my mental energy. I found having a call in the morning with team members/coauthors helped switch me into a different gear on the off days sometimes, but it required a lot of discipline to keep things separated.
We’re all learning together. I’ve had impostor syndrome for a lot of my career. It’s always refreshing to join a team and realise that everyone is still learning, and we all have things to teach each other. I liked how often everyone said things like “help me understand X” - I really hope I never have to work somewhere where admitting ignorance isn’t cool.
There is a lot of room to improve even in a saturated space. I would have assumed the market for AI image generators would quickly become saturated, and so even though I knew a lot about diffusion models the idea of basing a company on them seemed unappealing to me. BUT, it turns out that there are a lot of ways to improve things, and watching Playground continuously improving and moving from the “type into the magic box and see an SD1.5 image” towards “a magical new type of image editor” has shown me the value of this consistent improvement
Tracking hours does add a mental burden. This is one downside of any time-based contract work - any missed hours feel like lost money. This hit me especially hard - my sense of what “good pay” is is skewed from living in a country where people earn a few hundred dollars a month, so ‘giving up’ on a few hours of AI researcher pay seems insane! If you’re considering this path, be very careful to set the right mental expectations around salary, and maybe go for a fixed fee rather than trying to track and bill hourly as I did.

Why You Should hire/be a BIR!

Flexibility is great. Having options besides “devote all your waking hours to this company” seems like such an obvious win to me. It’s also a nice way for smaller companies to benefit from some deep technical expertise in a subject where maybe there isn’t enough capacity for a full-time role.
Diminishing returns / low-hanging fruit: When I joined I had lots to offer, and could give high value fairly quickly. After six months I can still contribute, but the effort per unit value has gone up substantially. This won’t apply to all roles, but in this case, I feel like it was a great deal for both parties that I could come in and be super useful, and then move on once we’d got most of the low-hanging fruit.
“Internships” shouldn’t just be for fresh grads, consider it a trial period for both company and employee. If I was looking for something full-time, I’d be much more likely to join after getting a real taste of what it’s like here vs. if all I had was a few short interviews to suss out the team.

Final thoughts

This post was mostly written as a journalling exercise, but if you’ve made it this far I hope you’ve found it thought-provoking. If nothing else, let this be a reminder that the ‘rules’ of work are more like guidelines, and it is possible to find a better option than the default in many cases. If you’ve done something like this yourself, let me know - I’d love to hear how others’ experiences of things like this have gone.