What's the justification for an attrition target? It seems such a weird thing to want your staff to leave.
Of course teams of employees with emotions and institutional knowledge are a bit more complex than a hand of cards, but that has never stopped bad management strategies.
What it is good for in my view is making managers look and feel tough and establishing a culture of fear that is so necessary for hierarchical power structures. Power/status hierarchies are a common trait for most primates. So my shorthand for this sort of it-doesn't-make-sense situation is, "Primates gonna prime."
Not defending, just explaining.
But this makes sense _in a game with a fixed hand size_. But a company's staff should grow over time. So shouldn't one have an ROI-based criteria where you want each team member to help grow generate meaningfully more revenue than they cost to employ. If my "lowest" performing team members still make meaningfully more money than they cost to employ, they still increase my ability to hire more.
This acknowledges that the "optimizing a hand of cards" analogy leaves out factors -- which is true.
My point is, even if there were no emotions or institutional knowledge, even if I had a "staff" of AI agents or robots whose "knowledge" could be fluidly copy-pasted, the "optimizing a hand of cards" analogy is poor, because a robot that makes money is worth keeping even if there are better robots out there. An army of low-but-positive ROI robots can help me buy a high ROI robot.
This is where your cards analogy falls through, in a deck of cards, you know that there are better cards you don't have in your hand, you know exactly what your Jack is worth in absolute terms.
At Amazon, you can have a Royal Flush, and because of stack ranking decide to discard your 10, and if you know your poker, that makes you go from the best hand to one of the worse.
The issue is you wouldn't know your 10 at Amazon was a great engineer and a key asset to your team, and that all replacement candidates are going to be worse, except now you've lost precious time, ramp up, and moral, as well as a great employee.
For a smaller company, it might make sense as you're building a core team and might need to go through a few people before you find a set of good ones, but where Amazon is now, at its size, over time, it has lost engineers that in industry wide averages were above it, and it has created a reputation that hurts it from attracting them back or other engineer's above the industry's average.
Due to this, Amazon is becoming a second pick. Good engineers will pick Netflix, Google, Microsoft, and even sometimes Facebook (due to their higher comp) over Amazon. And since all these companies hiring process are the same, it's common that someone who can pass the Amazon interview can pass one of those other ones as well. In fact, almost 70% of everyone I've seen be released from a PIP went to work at one of the other big tech companies.
The other big problem is that the stack ranking works like American reality TV competitions. It doesn't matter if you've excelled year over year, if you have one bad year, where for whatever reason you get identified as a low performer, you'll get PIPed. In my experience that's often contextual, a new manager or you look like a bad performer because the business and leadership couldn't figure out good projects for you to work on and own that year, or they themselves couldn't get their act together and you're a scape goat. Other people on your team might have performed better simply out of luck of having been assigned better projects.
Amazon used to get away with simply finding the best talent, but in my experience there, that's now backfiring, and instead of growing the best talent, their quest to find it is turning into a system that recruits more and more mediocre talent as time goes on.
Amazon doesn't pay top dollar for engineers, and the delta between Amazon and other companies is growing every year. (The compensation ends up being competitive when you take into account stock growth, but the new hire offers are not attractive.) And it's a very results-driven, stressful work environment. The effect of that is that people who get better offers go elsewhere, both new hires and existing employees. Those who are left either don't feel like interviewing or interviewed but didn't like their other options.
There are elements of the culture I miss. The focus on writing was great. Shipping things quickly that impact a lot of people is nice. It made up for a lot of the deficiencies, and it isn't universal (cough Google). Amazon tends to make the right tradeoffs with tech choices to enable this, not always, but most of the time.
I enjoyed working with principal and senior engineers at Amazon. I didn't meet a single PE that I didn't thoroughly respect and enjoy talking with. I miss the internal PE talks, and I wish they'd make them public.
Compensation is deeply personal, but once I broke 400k, I really didn't care too much one way or another, and Amazon's total compensation was never significantly less than competitors. It was about the work, impact, and people.
Yes comp could help to attract better talent, that's what Facebook is doing. But of the about 150 interviews I've conducted for Amazon, 90% of almost all candidates always ask me: "So is it really as bad as they say working here? With how they treat you?"
I think that's a pretty good indicator that the reputation is just tarnished, and I wouldn't be surprised that that's having a sizable effect on the decrease of talent.
The other thing is, yes maybe some real bad apples leave from a PIP, but the whole culture around it, the stress, the feeling everyone has that they constantly have to fight for trust and respect, that is also a cause for a lot of the really good engineers and high performer to leave as well, of their own, no PIP involved, but it's the same root cause for why they leave.
I see so many good ones, ranking high every year, and after 3 to 5 years say: Well I had enough of this BS, too much hassle always playing the game. If anything, that's the biggest issue.
Every time an AMZ recruiter tries to poach me from my current job, I consider replying back with links to the NYT article, the URA article, and various anecdotes from friends that worked there. Ending with a question of 'Why in god's name would I work for your company?!'
source: recently left Amazon as SDM
Maybe you hired a 7 of clubs, but given time they can gain institutional knowledge and improve their skillset to become a Jack of Spades.
I'm happy there are companies that set attrition metrics as a KPI, it makes it easier for me to hire and build out a technical team for my company.
Statistically, some of your employees probably aren’t good enough to be worth employing. But managers are human, hard conversations are hard, there are lots of incentives to not manage those people out.
By having an attrition target and forcing people to cut the bottom 10%, you basically skip over the soft fuzzy human factors keeping people around, and instead fall back to the statistical truth that the worst 10% probably are negative value employees.
Not saying I agree with this, but I think that’s the strongest case you can make for it.
(Somehow stack ranking logic never gets applied that way.)
In particular quotas flow down into teams with zero percent of people who should be fired. The effect on the remaining 100-X% employees morale is negative.
For a company that's vaunted for it's long-view philosophy, it's a terribly short-sighted policy.
1. It doesn't work, obviously because they are not in a closed system. Their policies affect their reputation, which will affect their ability to hire ICs who have to contend with an increasingly higher bar to clear.
I wish I could have said that's better than the alternative, but it's not. I suspect the ground is always shifting beneath you at Amazon. Even if your performance is fine, a colleague or your manager could get PIP'd
Why is this a statistical truth? In a large, Gaussian distributed population MAYBE this is true (but hiring isn’t random, is it?). But the thresholds are not applied to large populations anywhere I’ve worked. It’s pressure applied to management at all levels.
That means people will devote SUBSTANTIAL amounts of their day-to-day labor in jockeying and positioning.
People will be hired (wasted money and time) explicitly to be fired/PIP'd.
The jockeying will result in disruption of team development.
People will sabotage their coworkers.
Trust disappears.
Paranoia grows.
Blame deflected.
Difficult problems won't be tackled because why risk it.
Bandaids rather than solutions will result.
YOUR BEST PEOPLE WILL LEAVE. Let me emphasize this. You are selecting for people that aren't good, because good people will have OPTIONS and THEY WILL LEAVE.
What are you left with? A management hierarchy of sociopaths. A subpar set of employees that furthermore will not cooperate in good faith (in the best case) or actively engage in institutional sociopathic behavior. You will have a cesspool, and cesspools repulse good people.
And the signal is right in front of people that apply: the hiring process is abusive hazing with the "raise the bar" bullshit.
Staff turnover means abandoned systems. Abandoned systems will be avoided as bad career risk, or reimplemented before the org has gotten good ROI from them.
THAT IS NOT GOOD MANAGEMENT.
Remember: AWS had a CRAP Christmas. Crap. This is starting to boil up.
Amazon management practices will breed sociopathy at the root worker level.
I agree that is kind of the average-stage evolution of an enterprisey company. Amazon is worse.
The kind of mean-spirited pettiness that drives actual post-2000 capitalism, and it gets worse the higher you go, is so depressing and repugnant that it can't be put on TV.
Also, characters with no redeeming qualities are considered bad writing. But most coprorate executives are people with no redeeming qualities. You can't put them in fiction; unlike a well-written villain, they'll suck everything into a hole. Even Bill Lumbergh on Office Space had his charms.
Where do we get one of those?
The central limit theorem doesn't give us one of those. Not even the Lindeberg-Feller central limit theorem. Neither does the strong law of large numbers. Nor the weak law of large numbers. Nor the martingale convergence theorem.
There is some old material, from about 100 years ago, in schools of education that given a performance measure and a large population the measure will be Gaussian distributed on that population, and the larger the population the closer to a Gaussian distribution. An example is supposed to be the size of the largest eigenvalue in a principle components decomposition. Uh, that largest eigenvalue was long called IQ (intelligence quotient). Nothing intelligent about it. Nonsense. 100% total crackpot nonsense. Brain-dead, cult nonsense.
"Low performers" in need of a "performance improvement program" (PIP)? Start with the managers who believe in a Gaussian distribution. Then move to the managers who believe that the "bottom 20%" are always low performers.
A lot is known about how to be a good manager, and the Gaussian probability density distribution is not part of that.
I've seen a lot of bad managers. A pattern is that they are tyrants, have lots of rules and measures, are big on formality over reality, dot i's and cross t's with great reliability, and have everyone with their nose to the grindstone, ear to the ground, shoulder to the wheel and trying to work in that position. But the organization is stagnant, is not changing or keeping up with the market or technology. So, after a few years like that, it becomes obvious to everyone, customers, BoD, stockholders, employees, even nearly all of the managers that the organization is about ready to die, and often that's what happens, e.g., by firing 50% of the people and going downhill from there.
Once I saw a failing organization where nearly all the managers except at the lowest level were part of a tight clique, cult enforcing failure. Finally the BoD installed a new CEO and fairly soon 80+% of the clique, cult were given "PIPs" and demoted out of management or retired.
But there is also good evidence that even such a sick organization does not have to die and, instead, just needs some good management, starting at the CEO level. In particular, with a good CEO, suddenly, wonder of wonders, 90+% of the employees can be seen as great performers.
So, net, as bad as the situation can be, it is fair to say that Darwin is on the case with improvement on the way!
At one time Amazon was a small mail-order book and record shop run by Bezos from a small office. At least looking from 10,000 feet up, Bezos was a good manager, and, thus, I doubt that he was playing with the Gaussian distribution, firing his bottom 20%, stack ranking, etc. Instead he knew all the employees and the work of each. Sooooo, it sounds like now Bezos should leave outer space, return to Amazon, and fix the destructive nonsense that is ruining the company.
HAHAHA well said.
Height cannot be Gaussian if only because height is never negative but every Gaussian density is positive for all real numbers, including negative real numbers.
"interactions of many genes that are measureable"
The interactions of many is commonly used to justify a Gaussian assumption, but there is no such theorem. The main theorem to get a Gaussian density is the central limit theorem which usually assumes an infinite sequence of independent identically distributed random variables (plus some), i.e., i.i.d. The i.i.d. is asking a LOT!!! Asking so much that in practice it is essentially unrealistic.
What people often mean by Gaussian is just a central peak, some symmetry, and long tails, but an actual Gaussian distribution has a lot more, e.g., sufficient statistics (as in a classic paper by Halmos and Savage, with a derivation in M. Loeve's two volumes, from Springer, Probability) which, as I recall, E. Dynkin showed are quite sensitive to deviations from actual Gaussian.
Gaussian is important in a lot of derivations -- a LOT is known. But in practice Gaussian has some utility but only as an approximation where don't need to be very careful about accuracy.
Sounds fine.
If you treat heights as some i.i.d.s, then can use the central limit theorem to argue that in the limit for large samples (we get convergence in distribution) the probability density distribution of the sum of heights expressed as a z-score (mean 0, standard deviation 1) has Gaussian probability density distribution.
But that does not mean that the probability density distribution of heights is Gaussian. Indeed, if include both males and females, then for the density likely get two peaks instead of just one. If also include children, then get a left tail longer than the right one.
So, net, we cannot expect that heights are Gaussian. And we will have a tough time finding a large population that is Gaussian. z-scores in the limit for large samples Gaussian -- sure, can get that. A large population Gaussian, not much chance.
How does one qualify the new candidate being better than 50% of current team. Better than 50% at what?
Every role has many dimensions, and most members of the team have a mixed set of strengths. You form a cohesive package if you hire complementary people and get 2+2=5 type of output.
The idea that everyone can be treated as an interchangeable cog and objectively, quantitatively ranked is pure silliness.
All the counterpoints about whether 10% is accurate, whether you get the right 10%, whether they’re evenly distributed, whether there are other costs, why not apply the same logic to managers and execs, etc etc etc, yes, I agree!
Further, in any big organization the weak performers are not evenly distributed. However the PIP / attrition % rates are uniformly forced. So the over performing team has to start cutting into good performers within a year or two, whereas the low performance team can cut and still have many lower performers..
All this being said, in the context of an organization that actually deserves to exist (a rarity, given that we live under capitalism), it is impossible to say what percentage of people will be ill-fit to their roles, let alone what percentage will be so ill-fit to any role that separation is the only option.
https://www.washingtonpost.com/archive/politics/1994/07/17/h...
I also want to point out that a lot ppl who suffered or even lost their lives during such period were actually sabotaged or framed by their peers (neighbors, coworkers), rather than the authority.
His answer: "because, of the remaining, a new bottom 20% will be formed"
The other bit is it's a form of resiliency. High churn improves the bus factor. I respect Amazon for building a system that's resilient like this. That's not saying it's a good system to work in, but I admire that they built a system that can operate like this.
High churn is a high bus factor, like by definition of a bus factor. Bus factor => people had kids, retried, got a better offer, got fired, or were incapacitated (the proverbial bus) on a moments notice, and thus are abruptly unavailable to maintain your systems.
Churn is not a mechanism in itself to reduce the bus factor. It doesn't encourage people to mitigate the bus factor, because if you're going to be pushed out anyway, why bother building something sustainable?
I could see the argument for a probabilistic approach in that the company continues to function with critical systems being black boxes that teams have so far successfully scrambled to replace (with varying degrees of success), kind of like a more permanent Chaos Monkey for an entire system instead of cells within an system.
However, that does seem like it would waste a bunch of time, money, and opportunity on churn to deliver what you already have instead of using experience to improve/replace the existing systems.
This is just one (very depersonalized) approach to not keeping 'dead weight' around, or optimizing staff to always be on an upwards trend. It's like an evolutionary algorithm; trim off the worst performers and the remainder will evolve to higher levels. In theory.
In practice however, attrition goals are terrible. It destroys team cohesion, encourages people to sabotage eachother and incentivizes sociopathic tendencies. If you know the lowest performers will get kicked out, then you have every incentive not to help those around you.
Typically the people who get pushed out are not the lowest performers, but rather those who have pushovers as managers. A vicious manager can fight and defend his reports. The weak ones end up sending someone to the "least effective" bucket, about half of which get fired in the next few months.