Management technique we learned early on: short-term machine learning deadlines can be set based on inputs (e.g. high-quality execution on a set of experiments) but not outputs (e.g. reaching some level of performance). Science does not bend easily to the wishes of managers.
hardmaru
 @hardmaru
11 May 2022
“Are there any software engineers that switched into a machine learning role and found it a lot more stressful due to deadlines combined with the uncertainty of research?” Discussion: old.reddit.com/comments/ulsuzn

May 11, 2022 · 1:53 AM UTC

21
136
21
908
Load newest
Replying to @gdb
In my previous job. We used to publish Model metrics to management. After giving same accuracy for couple of weeks management stoped asking about status.
4
Replying to @gdb
ML and specially research has inherent uncertainty, while software engineering is risk free.
Replying to @gdb
Agree 100%. This is one of the things managers need to be wise about when working with ML/DS teams ( mitsloan.mit.edu/ideas-made-…)
1
Replying to @gdb
More from a managers perspective: another interesting thing you learn as you calibrate performance across MLEs and SWEs is how much you weigh process in the former and outcomes in the latter.
9
Replying to @gdb
That's a really important point - machine learning is a complex process and it's important to be realistic about what can be accomplished in a certain amount of time. Trying to force results will only lead to frustration and ultimately lower quality results.
Replying to @gdb
Mostly that we tend to approach outputs.
Replying to @gdb
Hello, is your lesson documented anywhere other than in this tweet? I agree with what you say, but I'd like some reference to quote at people! If not I'll just use this tweet...
Replying to @gdb @hardmaru
How do you measure high-quality execution?
Replying to @gdb
AI engineering is not Markovian completeness
Replying to @gdb
💯💯 ML is more of an experimental science- we collect evidence and run tests - than engineering which is building and execution. When we combine both, that's when the magic happens and we get useful products.
1
Load more