Snipers vs Sharpshooters

EDIT: This whole article should’ve been one sentence: “specialists aren’t generalists, big company MLEs are specialists, startups need generalists, act accordingly“

all of my advice is under-qualified rambling

‍

ChatGPT says sharpshooting is the Olympic sport in which participants demonstrate exceptional accuracy in firing firearms or other projectile weapons. A sharpshooter is someone who takes part in such sports.

Sharpshooters are highly specialized, and their goal is to increase accuracy.

‍

On the other hand, a sniper is a military or paramilitary marksman who engages targets from concealed positions or at distances that exceed the target's detection capabilities.

While in popular media people think that a sniper’s goal is to be accurate - most of the time the real goal of a sniper is to “make sure your team gets home safe” (or at least, in my totally unqualified mind and for the purpose of the analogy, that’s what i believe.

‍

‍

‍

BigCo Data vs Startup Data

Academic researchers, MLEs at big companies, are like sharpshooters. 8 digit budgets a year across an enterprise to spend on compute to train massive models on petabytes of data. You’re measured by agreed upon criteria - in relatively controlled environments, by a judgement criteria. You never need to think about profits or costs - just solving the next highly specialized and technically difficult engineering problem. A 1% improvement on targeting accuracy is another $50M in revenue due to the 5% increase in conversions. You’re like a quant at jane street figuring out the right arbitrage to pour ad-money into to get revenue and profits out.

Your job is so valuable - bc your single digit percentage improvements are so valuable… that any time someone distracts you from the big problem, you SHOULD be saying “that’s not my job”.

Being a data scientist or MLE or whatever at a startup is a categorically different game. Your company is default dead. 10% increase in accuracy is a rounding error. Problems aren’t hard - there are just 10 of them so you need to find the 80/20 in 10% of the time and you might still not be doing enough.

Your success isn’t measured in accuracy or precision - it’s measured in whether or not your company goes bankrupt and everyone loses their job. At a startup, life is not a given - and you don’t have the luxury of optimizing. Even thinking about optimizing before something is already on fire is by definition over-optimization. You might be called on to close a deal. You might be called on to make some changes to the front end. “That’s not my job“ isn’t part of your vocabulary at all.

‍

If you’re at BigCo

Don’t join a startup if you want to stay at the cutting edge of whatever your domain of optimization is. I don’t know the stats - but if I had to guess, military snipers don’t do very well at the olympics. It’s nice to tell ourselves that if you can do all the general tasks - that toughens you up to do the specialist tasks even better.

This is not true. If you want to be a published researcher - stay in the environment where you’re so highly leveraged, companies will buy software to keep you in the zone.

‍

‍

If you’re hiring for a data scientist at a startup

Don’t hire from big tech. They’re not going to want what you offer - and if you manage to convince them to, they’ll be sorely disappointed once they start. “What do you mean I can’t query a petabyte of data whenever I want? - wait we don’t even have a terabyte of data to train on?”

‍

Unless…

I was going to write a whole inspirational thing about going into the arrangement with eyes wide open - aware of the tradeoffs on both sides… but I realized that this is already getting a little too preachy for my taste.

Just be honest with yourselves. A startup isn’t going to give you the power to do what you get to do at a company that spends 7 figures and up on compute for ML training. A startup is a place where every day is a fight to stay alive.

And as long as you’re honest with yourself everything will be fine