| Dear Author, Software Safety and especially AI Safety are hard technical problems. You're not going to figure out how to add a few 9s to the reliability of a computer vision system by studying org theory. Making progress on The Hardest Problems requires hiring the excellent ML engineers and scientists, and then having even half-decent management. Management does matter, but the strong IC talent is a precondition to progress. If Defense had armies of competent ICs but was still failing, I guess focusing on non-compensation-related organization issues might be reasonable. That's not even remotely the case. Just shy of 70% of CS PhDs from US institutions are foreign nationals, which means Defense is already talent-constrained. The 50% or so who are qualified to work on AI safety and qualify for relevant clearances would probably want a close to a 0 added to what Defense offers (and not even for moral reasons... security clearances and "must work at the office" are real drags on quality of life that require significant additional compensation). I've seen first-hand that defense systems are insecure and unsafe because defense chooses not to purchase excellent IC talent. The biggest organizational challenge in Defense is the lack of adequate compensation for engineers and scientists. The largest management problem in the Defense industry is figuring out how to get the org to pay for excellent engineering and scientific talent. Your competitors in the labor market are paying high six or low seven figures, don't require security clearances, don't require drug tests, and have much more hybrid/WFH flexibility. For me to take a job in defense to work on ML Safety, you'd probably have to pay north of $1M. |