Some teams shifted 20-40 percent of developers to annotation work.
Minh Nhat Nguyen#1716
human data all the way down
Meta is pulling engineers from unrelated orgs (eg ads) to become data labelers for MSL.
Many users criticized Meta for reassigning infra engineers to AI data labeling as burning talent and destroying engineering culture, while some defended it as essential hands-on work.
No Digg Deeper questions have been answered for this story yet.
Never heard so many standout infra engineers + AI infra eng actively wanting to leave Meta than now.
A month ago they were building cutting-edge infra and then got assigned to AI data labelling
Most of them went “WTH” and now I’m the middle of interviewing
Madness from Meta
Do you know what all of the top AI researchers do that their subpar counterparts don't? They actually look at the data
The top google search engineers spend their mornings doing a/b blind tests of search results on Google sheets with their own eyeballs
Has been for weeks, now. Covered it two weeks ago in @Pragmatic_Eng
fwiw those forcefully reassigned to data labelers were not laid off and probably will not be let go this year (no more mass layoffs). One positive at least
if you want to hire from Meta, now is the time btw
an ironic thing about the meta layoffs is that they also function as a pretty good retention mechanism. it’s hard to quit your job when you know there’s a chance that the next round of layoffs could pay you 4+ months of severance to leave instead.
Never heard so many standout infra engineers + AI infra eng actively wanting to leave Meta than now.
A month ago they were building cutting-edge infra and then got assigned to AI data labelling
Most of them went “WTH” and now I’m the middle of interviewing
Madness from Meta
Has been for weeks, now. Covered it two weeks ago in @Pragmatic_Eng
fwiw those forcefully reassigned to data labelers were not laid off and probably will not be let go this year (no more mass layoffs). One positive at least
if you want to hire from Meta, now is the time btw
Meta is pulling engineers from unrelated orgs (eg ads) to become data labelers for MSL.
Also I am talking with long-tenured and experienced engineers NOT assigned to data labelling, and not laid off.
Many of them started to interview as
1. 40–60% if their team is gone
2. Some of them got equity grants as they are deemed critical now
3. They stopped trusting direction & assume they could be next… best to leave now, when they have so many more options
4. It’s less the layoffs but more the senseless forced reassignment of standout devs to data labelling.
Never heard so many standout infra engineers + AI infra eng actively wanting to leave Meta than now.
A month ago they were building cutting-edge infra and then got assigned to AI data labelling
Most of them went “WTH” and now I’m the middle of interviewing
Madness from Meta
To clarify “assuming they could be next” is not about laid off in most cases: but being assigned to data labelling
Meta made probably 3-5x more devs manual data labellers than how many they laid off
Just insane, seemingly deliberate attrition they are forcing to happen
Also I am talking with long-tenured and experienced engineers NOT assigned to data labelling, and not laid off.
Many of them started to interview as
1. 40–60% if their team is gone
2. Some of them got equity grants as they are deemed critical now
3. They stopped trusting direction & assume they could be next… best to leave now, when they have so many more options
4. It’s less the layoffs but more the senseless forced reassignment of standout devs to data labelling.
If you think you are above labeling data you are not going to make it
Do you know what all of the top AI researchers do that their subpar counterparts don't? They actually look at the data
The top google search engineers spend their mornings doing a/b blind tests of search results on Google sheets with their own eyeballs
i find interesting that data labelling is essential but frequently disliked. i suspect it's bc in most cases, people aren't really rewarded for putting in very high effort and attention to detail versus phoning it in w mid quality, so people phone it in.
Never heard so many standout infra engineers + AI infra eng actively wanting to leave Meta than now.
A month ago they were building cutting-edge infra and then got assigned to AI data labelling
Most of them went “WTH” and now I’m the middle of interviewing
Madness from Meta
@paularambles it just effectively creates zombie & fearful employees… continuous layoffs are a major morale killer. i can’t believe they are doing it so regularly.
an ironic thing about the meta layoffs is that they also function as a pretty good retention mechanism. it’s hard to quit your job when you know there’s a chance that the next round of layoffs could pay you 4+ months of severance to leave instead.
Original article: https://newsletter.pragmaticengineer.com/p/the-pulse-did-capacity-shortages
Has been for weeks, now. Covered it two weeks ago in @Pragmatic_Eng
fwiw those forcefully reassigned to data labelers were not laid off and probably will not be let go this year (no more mass layoffs). One positive at least
if you want to hire from Meta, now is the time btw
everyone i know who survived this round of layoffs wants to quit, but then immediately goes “but there will probably be another round in august”
an ironic thing about the meta layoffs is that they also function as a pretty good retention mechanism. it’s hard to quit your job when you know there’s a chance that the next round of layoffs could pay you 4+ months of severance to leave instead.
Imo this is true for data collection too, at least a little bit
If you think you are above labeling data you are not going to make it
@chris_j_paxton It's also true for pure RL! You need to look at rollout samples the entire way through
Imo this is true for data collection too, at least a little bit
the other thing is that if you spend 3 weeks creating "code datasets" it feels like an accomplishment than improving a live codebases . it's spiritually more like a "code factory".
i find interesting that data labelling is essential but frequently disliked. i suspect it's bc in most cases, people aren't really rewarded for putting in very high effort and attention to detail versus phoning it in w mid quality, so people phone it in.
@GergelyOrosz On the data labeling effort, META does not believe that third-party contractor data is of sufficiently high quality and believe this is an important point of differentiation relative to competing labs that largely rely on Mercor, Scale AI, Surge AI, etc. See more here:
@menhguin This is one of the reasons that working in Biglaw was usually considered so soul crushing. It required extreme attention to detail for very boring tasks over a sustained period of time. It is extremely unpleasant
@AndrewRHarvey they are labelling code, sometimes adding tests. no infra being built lol
@GergelyOrosz sad for my former colleagues but glad i left
@GergelyOrosz @Pragmatic_Eng The only new update to this is since you wrote: there have been more drafts leading to the most hilarious outcome where this org is larger in headcount than the whole headcount of Anthropic or OpenAI 🤣
@PawelJLisowski @Pragmatic_Eng yes they did
that's the one positive in all of this
one more reason morale is at a historic low at Meta. None of them signed up for this when joining to build cool software
But yeah, very highly paid data labellers