I work at a HFT (not Cit) and have interviewed a bunch of Meta employees. They have impressive claims on their resume (boosted engagement by X%, saved Y% in capacity, removed Z% of bad content, etc). However, when you actually ask them questions about their work they know nothing. I had one interviewee tell me they made a new recommendation model but couldn’t tell me what type of model it was other than “some ML model”, or how they chose their features, or how they got their training data. Some would explain how their systems work using Facebook tooling terminology and couldn’t even explain to me what these tools actually do. Others wouldn’t even know the traffic or storage volume they deal with on a daily basis.
Write code, why? What do Citadel/HFT engineers do?
Get grilled over single ms variance.
I know one thing they do very well, which is bringing value to shareholders
maximizing shareholder value **
Meta is a stock market darling right now, but I wouldn’t bet on it being same long term really.
सुंदर नाम।
Why would an engineer know the traffic or storage volume if they never deal with those numbers? There are tens of thousands of software engineers at faang companies working on services that are less than 1qps (like maybe 1 request per hour or day or week) and plenty never interact with storage past using an AWS API or equivalent.
You really do need numbers as they can influence your design choices. Although most companies can get by without knowing/estimating them, but at scale you need to understand the numbers so that you will be confident that the underlying systems (database etc) can handle the scale, latency requirements or storage. The more senior you are you will need to care about these numbers and that’s a primary difference between a junior and senior. Recently, We had choose a file based storage(we don’t need to rebuild a database around file system, there’s existing infrastructure around it) instead of spanner db as we cannot not achieve the latency at the scale we have using spanner.
This happens when you build tiny pieces of big systems, and you don't have much visibility re: the big picture. That's not unusual in big tech, particularly for lower level ICs. I don't think you should judge them too harshly - the skill of linking your work to value isn't as important there as it is in many other places. But you should know that those candidates may not have that skill.
There is a bespoke culture where Meta engineers “sell” their work - spending hours on wordsmithing and manipulate fake metrics for PSC. Another aspect is the culture encourages high-impact hacking, so there’s generally not a lot of thinking behind things.
The biggest skill one learns at Meta these days is massaging words to sell impacccc
How much does your HFT pay for ML engineers? Might be trying to GTFO
This. I've found that finance doesn't pay nearly as much as I used to think it did. Big tech easily eclipses it. Since the WLB as a quant is also awful, I'm not sure why anyone works there
Some people (actually many people, look at the explosion in retail trading and stock picking) like finance way more than tech.
Meta has some incredible tooling in place that allows for impressive dev velocity. MLE can train models by creating a training data table, pointing towards it, and choosing some basic parameters. A lot of hyperparameter tuning and automl happens in the background. There’s still room for creativity though as you can apply active learning but you might spend more of your time on model analysis, evaluation, calibration, etc. The downside of this tooling as a swe is that it places many layers of abstraction between you and the basic ml framework like PyTorch (which is still high level) so you might not learn a lot about what’s happening under the hood. Same thing applies with their data systems etc. Product ml seems to be moving towards more high level work however, as ML toolkits and frameworks are adopted by more companies, so this experience may extend to midsize companies as opposed to just big tech as it is today.
I mean you can train a model very quickly with the high level configs but to actually do the analysis, evaluation, and debugging you should actually know wtf model you're using and why at the veryyyy least. Not even knowing that is quite shocking
We sit in meetings and fight for scope.
> fight for scope What does it mean? Trying to make as big part of the project to be assign to yourself as possible?
Trying to do something new that multiple teams are interested in doing. Whoever wins gets to claim the impact.
We interview candidates form HFT’s who don’t know wha they’re doing
Extremely predictable response
Touché