OMG, four, or possibly half a dozen departures. LeCun’s comment comes off as glib the way it is presented in the article, but after reading to the end of this very short article, the comment sounds reasonable.
That is significantly incorrect. Facebook is safely the number two AI company in the world. In addition to all of the basic AI research Facebook does, they also have numerous ML components built into their products such as translation, ads, photo recognition, moderation, feeds and recommendations. Plus, with their strong pivot to VR/AR/Metaverse, they use a lot of computer vision and object detection capability, much of which is built into their current VR offerings and running on commodity cameras.
It's not a vanity division, but I would definitely put Facebook in fourth, below Google Research, DeepMind and Microsoft, so definely not safely on second place
Number 2, based on what? Among OpenAI, Google, Waymo, DeepMind, Tesla.
FB is mostly deploying models developed elsewhere, instead of advancing the frontier of SOTA.
I'm not entirely sure what the OP meant, though if you asked me "list the top AI companies" as a reflex I'd probably mention Google and Facebook first without any prior evidence. I think it's kind of hard to really compare. Could Tesla AI engineers do the AI work of Facebook? Probably not, and vice-versa. Tesla's AI goals seem to be even more specialized than most other companies (which is ok and not a negative thing). Are we comparing "solve any general AI problem" or "best at solving their specific problem"? No idea.
And Does it make sense to separate Waymo and DeepMind out from Google since that's the owner? Idk. Just a thought.
But even if they weren't exactly #2, they're probably inarguably top 5 right? That's still pretty amazing. And that's without really considering international companies. I have no idea what the AI capabilities of Chinese tech companies are, for example.
Yes, that was my thought as well. It's a pretty subjective statement, but Alphabet and its subsidiaries are pretty clearly on top and there are not too many obvious competitors for the silver. OpenAI is quite small and narrow in focus. I have not heard anything from Baidu in a long time and Tesla also has a very narrow focus and does not contribute much to the research or tooling environments. Facebook is extremely active across a range of AI applications, publishes papers, creates tooling and releases products that millions pay for.
Funny I would also not count Facebook and found it really weird when they announced the DC for meta.
With Google and ml I connect so much positive and life-changing like: health care, weather, language model, deepmind ... There research blog and papers are great.
And then Facebook.
Yeah what is Facebook doing with ml? Optimizing for the next bullshit to make our society worse?
They are pretty top notch though- and also in a different field rather than ads - that alone makes them innovators. Also, doesn't Karpathy work at Tesla?
They aren't top notch, though. And while they are a different field than ads, even in a list of top self-driving AI companies they'd only make it on if the list were quite long. Karpathy is fine and writes (or at least used to write) some interesting things, but he's far away from single-handedly turning them into a top-AI contender (I mean, if one person could single-handedly do that, he'd be far down that list of people).
This comment is ridiculous. Facebook has some of the best researchers in machine learning and deep learning. And if AI at Facebook were a vanity division, Facebook would never have a viable ad business, and the entire Facebook would have shut down years ago.
I suggest finding more polite ways to critique a post. Otherwise discussions tend to get boring quickly.
Also, the GP was explicitly describing their perception, not trying to promote it as truth. One interesting response would be "I really think that's false. I wonder how we reached such different conclusions."
The original commenter was talking out of their ass and not remotely qualified to comment, given they gave such a patently false opinion - they are just providing an ignorant hot take. The parent comment to yours is completely appropriate IMO, it's not attacking the poster, it calls out the comment. Your condescending follow-up only takes away from the overall level of discourse, to say nothing of the kind of responses (like mine) is ends up invoking. Nothing wrong with calling out BS, it would be much worse, both in terms of honesty and condescension, to give the fake "I wonder..." that you suggest
Facebook has some of the best researchers in machine learning and deep learning.
Which is exactly why some consider it to be a vanity division.
All that brain power devoted to a company nobody (genuinely) respects, and whose sense of vision -- in the current year, 2022 -- teeters between stale and outright delusional.
But hey, if it gets you $300k a year for easy work and resume cred for the kind of job you really want -- go for it.
Maybe the resume credibility is not what it was once perceived to be. I also find it hard to take Facebook AI seriously and have no reason as of yet to believe it won’t be used for low brow reasons like increasing user addiction to their apps, censorship of user opinions, or increasing advertising engagement. In fact, I am skeptical of AI and ML in general for its effectiveness, purpose, and applications.
In fact, I am skeptical of AI and ML in general for its effectiveness, purpose, and applications.
Agreed. It's all about the application and purpose (and the hubris of some of its practitioners), not the theory (or even the demonstrated effectiveness in certain areas).
When we see these disconnects at "work", on the ground -- in terms of lack of relevance / applicability to what the business actually needs -- they are sometimes quite startling, in fact.
Just for curiosity , can someone please point out what are the AI things that came out of Febu that is now in the public domain and used everywhere? Like for example Tensorflow from google.
To be fair, these ad tech platforms are built on top of algorithms that predate the latest AI revolution significantly. Logistic regression and matrix factorization get you very far in adtech. A lot of the performance then comes from heavy engineering.
The claim was that AI wasn't used in Ad Tech at all. Not the latest AI. I agree with your assessment, but to say that AI/ML isn't used in AdTech is just nonsense. AdTech, especially the last 20 years has been built on core AI/ML concepts. I also highly doubt that the latest AI tech isn't being applied somewhere in the AdTech space, but I left AdTech a while ago and am never looking back.
I don’t what are you talking about. Machine Learning is one the fundamental technical pilars of modern Facebook. It’s used in almost every process imaginable in some way or another.
I doubt it would have to be used that much, and results are questionable and even malicious at times. So, not sure if that makes people working at AI at Facebook that great...
A significant portion of the current tooling and research methods used in AI come from facebook. Not just in adtech either. I work in computer vision and literally have never done a project that didn't rely heavily on something from facebook (even ignoring pytorch).
I have a lot of bad things to say about the company and their products, but they and google are almost single (double?) handedly responsible for the advances in ML/AI that we've seen over the last years. (Not to say they are at all the only ones who have contributed, but if you subtracted their contributions we would have a very different world, certainly an average schmoe like me wouldn't be able to fire up world class models with millions or billions of dollars of R&D in them in a free hosted notebook connected to GPU, and actually solve difficult ML problems)
> Are the people who depend on it seriously incapable of taking up the load?
Probably. First of all most are domain experts not at all systems programmers. That’s the main value of such libraries.
And even for those who work close to the iron: plenty depend on, say, the Linux kernel without being able to write patches or even diagnose problems. And few, if any kernel developers have submitted patches to, say, GCC.
There seems to be serious talk about reimplementing the Torchscript core in Julia (i.e., reimplement the C++ part in Julia), something that makes sense to me given the excellent support there already is in Julia for autodiff. Cf. e.g., https://dev-discuss.pytorch.org/t/where-we-are-headed-and-wh...
If these suggestions are credible and don't come from the Facebook team, that might sidestep your concerns.
While I would love for that to happen. Quite recently (e.g. [1], also discussed on HN, I believe) the Julia community seemed content with more niche applications, quite deliberately choosing not to challenge the biggies (TensorFlow, Torch, etc.) on deep learning, because the amount of mainainers needed to be competitive was just not there.
Is there reason to expect this to change? Maintaining something like Torch and keeping it competitive in terms of speed is a HUGE amount of work for systems development, writing insane numbers of GPU kernels, etc. After having read it, I wouldn't quite yet call the tone of the discussion you linked "serious talk"...
Being reputable and being a vanity division are not mutually exclusive.
You could argue that at it's peak Bell labs was a vanity division. That research may have changed the world, but very little of it likely ended up benefiting AT&T in any major way financially. It's telling that once AT&T was broken up Bell labs, while existing in some form for years after, was never reestablished.
Bell Labs still exists but is owned by Nokia, who now makes cellular infrastructure. AT&T Labs also exists.
But the days of telecom research making rapid and monumental advances peaked decades ago. Nokia or AT&T or Huawei or Ericsson could quadruple their R&D spending and it wouldn’t reestablish the impact of Bell Labs of last century, because it’s simply a much more mature field.
Whether it was a vanity division or not, it did serve one practical purpose for AT&T: it presented the company as a benevolent monopoly, spending its profits on developing technologies that benefitted the nation. This helped stave off anti-trust action for a long time.
I work for a big AI consultancy, but we haven't seen relevant work from Meta for a while. Their non-bayesian stuff is "meh" and often seriously misguided in their business approach. (which makes sense, they're not accustomed to work as consultants on setups with limited data)
IMO their area of expertise is in software and tooling. The fact that they maintain PyTorch is a big deal and I've used a few of their other libraries like Detectron and torch-points3d and some have become an integral part of my work.
This isn't even remotely true. wav2vec alone is the most impactful model in speech recognition in a while and their subsequent work in low-resource languages makes it pretty obvious they're working with limited data...
No clue what a "big AI consultancy" is, but as someone who works in ML on a big feed-driven social networking site, what you are saying is just not true.
That caught my attention too. I work on AI consulting and have a reasonable sense of the ecosystem. There are certainly big players that claim "AI" as part of their consulting business, e.g. Deloitte, but they are well removed from any of the research and really just selling dashboards or something with a new name. The GP comments is the kind of thing I would expect a consultant working at a place like that to come out with
So if I understand correctly: they are working in a domain different from yours, thus their work isn’t relevant to you, and even though they are not claiming otherwise their reputation is overrated?
Maybe you meant something else, but that’s how your statement reads.
Self driving at Uber was a core part of why investors pumped so much money into them and pushed them public. People expected Uber to roll out self driving cars and completely replace their human driven fleet of cars, thus increasing their profits immensely. It definitely was not a vanity division.
No, they are responsible for optimising the ad targeting. It's a really mature operation, with consideration like faster decisions means more time to run campaigns and less greed algos means more ressources for marginal improvements.
I always wondered how much the pytorch team could make if they quit and forked the project. I'm sure plenty of companies would be willing to sponsor them.
As I see it there’s a difference between sponsor and maintain. Sponsorships help but they don’t help in the same way someone with active knowledge of the domain does.
I see it as this, you can throw money at a project to build a new OS but you still need the right people to have an actually good OS.
I'm saying the team should quit and spin out an open source company to work on pytorch (same for react). They'd get much more buy in from the other big tech companies (like amazon) by being a neutral party and not part of Facebook.
There's a ton of companies that would be happy to pay them to keep working on it.
counter-example: CPython, NumPy, SciPy and many critical well-used open-source projects that have tiny volunteer teams and operational budgets that are minuscule compared to something like React / PyTorch
When these sort of highly sought after researcher and ML/AI talent begin to leave a company in large groups even if it's just a small number of people at first it's a sign that something is profoundly wrong at that company. These people work on the long term big roadmap items and can typically see the the future initiatives clearly. Essentially means that innovation at Meta has slowed to a trickle or has completely dried up.
These are AI/ML researchers, and so you can probably talk about that area at Meta, but it doesn't apply to other departments necessarily.
But I think it's kind of obvious that AI/ML is a dicey proposition there. See for example[0] how they decided to stop doing facial recognition. I can easily imagine a lot of AI projects being canned or back-burnered like that now. What's ML used for at Meta? Feed recommendations, social graph inferences, face identification, etc? All those are under heavy scrutiny, and a lot of the work of doing it "right" is not even an ML question at all, but one of policy, regulation, and product. It's not like self-driving cars where people can generally agree that "getting from A to B without crashing" is a good thing, and where the obvious ML and engineering problems line up with product problems. But if you "increase engagement" with your feed recommendations, that could be good or bad in ways that ML can't really tell you. If you get better at identifying people in pictures, people are going to hate you and there will news articles about stalking. If you identify who a potential friend might be, that's a privacy consideration, etc, etc.
But what's going on there is pretty independent of the "metaverse" stuff, I think.
> These are AI/ML researchers, and so you can probably talk about that area at Meta, but it doesn't apply to other departments necessarily.
When you look at their products as a whole at their root are essentially AI/ML driven or rely on the capability. This is very relevant to and applies to those that are most directly revenue generating like advertising.
> I can easily imagine a lot of AI projects being canned or back-burnered like that now. What's ML used for at Meta? Feed recommendations, social graph inferences, face identification, etc? All those are under heavy scrutiny, and a lot of the work of doing it "right" is not even an ML question at all, but one of policy, regulation, and product
Correct due to multiple factors there's little growth or innovation possible in the main products that Meta generates revenue from. Right now they only sell a single product which is advertising.
Metaverse is a highly speculative venture that may not work and there are companies like Sony, Nintendo or Microsoft Gaming that are better positioned to rely on existing technologies, platforms and synergies that have a major edge here.
That's why I would interpret AI/ML researchers leaving as a sign that Meta's best days are behind it and Metaverse is an act of desperation.
Facebook/Meta stock lost 30% of it's value a few months back and may not recover any time soon.
Many of these people are high level, so I'm guessing (based on a quick skim of levels.fyi) about 50% or more of their TC comes from stock.
This amounts to an effective 15% pay cut in a year with record inflation.
I think most of us would leave our job over a 15% pay cut, especially if we were well established in the field. On top of this it loosens those golden handcuffs quite a bit for anyone who was on the fence about being employed by facebook, but couldn't say no to the comp.
A lot of folks get into a zone where the money no longer matters, as long as it’s in a range where you don’t have to think about it. Still they might leave simply because they’ve been there a while, or want a change, or want to work with some particular people who are at a different organization.
Not to say that some might be motivated by money, but at that level I would be surprised if it motivated many.
I find this is true for me, but I ultimately don't work for a FAANG because the money isn't that important to me and there are many things about FAANG companies I don't like. However I don't think this generalizes well for people who do aggressively pursue careers at FAANG companies.
My experience has been that most of the people I know that work for FAANG are extremely TC conscious, that's a large part of why they work at a FAANG in the first place and the source of a non-trivial amount of their self worth. If you look around the comments on Blind (a biased source for sure, but certainly a non-zero part of the community) you'll find plenty of people where TC is all that matters.
I interviewed at FB a few years ago and one thing that surprised me when I asked "what's your favorite thing about working here?" I got the same answer from everyone I talked to: the compensation/benefits. I thought for sure the answers would involve working on hard problems at scale, having incredible amounts of data etc. However without fail every interviewer immediately pointed out their comp and other perks as their primary motivator for working there.
In general I would say that people that make a lot of money, make a lot of money because making a lot of money is important to them.
This idea that "eventually you make enough where money doesn't matter anymore" is a meme that couldn't be more wrong. Its the kind of thing said by someone who has never made that much money, but who is extrapolating their current relationship with money to a compensation multiples higher; then saying "well if I can live on $X right now, then the Y in $X+Y wouldn't matter, so money must not matter (past some level) (and, coincidentally, that level is always somewhere around what I make) (weird how it always works out like that, right?)".
You can say that it shouldn't matter; that a $200k salary decrease to someone making seven figures shouldn't matter, because they're rich and they can take it. Maybe that's a correct assertion; that it shouldn't matter. But: it does.
Well, personally, I'm in a situation where I struggle to spend more than half of what I earn, including my mortgage repayment.
I'm not a cheap person, if I need to spend money on something, I will. It's just that my life style and affinities mean I'm not spending much.
I'm not much of a fashion guy, so no huge collection of expensive clothes.
Nor am I a car guy (I don't even currently own one, and if I were to have one again in the future, it would simply be a tool).
When I travel, the destination is often decided at the last moment, and I'm more of a backpack kind of guy.
When I buying something somewhat expensive (price ~$1000 or more), I'm always evaluating the usefulness of the thing, for example, I kind of want to replace my folk guitar but in fairness, I rarely play my current one, and it's not like this will change with a nicer one (I also have two very nice electric guitars I've not touched in years).
Simply put, partly out of how I was brought-up, partly out of some ecological-consciousness, I do not buy something simply because I can.
Also, regarding this part:
> Its the kind of thing said by someone who has never made that much money
When I make this statement, opposition comes far more often from people "who have never made that much money". My friends with similar or greater earnings get my point of view, even if they don't have the same views, but my friends with lower incomes generally strongly disagree with this kind of view.
A Principal Engineer at my company, could easily increase his salary 20-50% by moving to a FAANG, and can do Leetcode problems in his head, doesn't move. Because he has more independence here and can work on more interesting problems here.
He grew up in relative poverty. He already gets paid a lot. It's enough for him.
Well that's not true... at least not for me. I'm not even at the point of earning excessive amounts of money (research scientist at a german university E13, so about 60kEur). And already now I don't care enough about money to have it influence where I'm thinking about applying...
I went for years not thinking about money, not negotiating or worrying about comp, and making enough that I didn't need to worry or think about it, I just wrote code and did research. It is definitely a thing
I think this is to some extent personality-dependent and not as universal as you posit.
I somewhat fall into the demographic you describe but find myself not really caring about maximizing TC because I’m working on interesting and meaningful problems. (I would never be interested in working on adtech or infosec even for double my pay — I’m just not interested in those areas.)
I have met many in tech whose game is maximizing TC. They’re very vocal but I don’t know if they represent everyone.
Until the dot com era money really wasn’t much of a “thing” here (SV), and you’re right: even now it’s more of an SF thing, but has infected the Valley too.
I'm not sure, I can't imagine feeling that way. The jump from $500k to $650k, or 1mil to 1.2mil, would still feel pretty significant, in terms of cash hitting your bank account.
Honestly, many many people feel that way. I have been there most of my working life.
When you don’t make enough to make ends meet, or to do so comfortably, sure, fixing that situation is going to be very important to you.
But once you have what you need to live the life you want, why not simply live the life you want? One of my closest friends has been at google for almost 20 years. He doesn’t want to be promoted because he likes his job and his coworkers.
I don't think this is correct. I'm lucky enough to be on a level where turning down a 100k increase in salary is something that I have to routinely do because my current job ticks so many boxes and I genuinely love it and don't have to risk it for that amount.
I'm not saying these people wouldn't take a six figure pay cut for a job they like better. Just saying a six figure increase can still be pretty significant and material.
Once your basic needs are taken care of, inflation at current levels is irrelevant. I'm by no means rich. By HN standards I have an average income. If not for the daily news stories about it, I wouldn't even know there was any inflation bubble right now. The top AI people at Meta make a substantial multiple of what I make. They might be leaving for more money, but it has absolutely nothing to do with inflation.
People at that salary level usually leave for other reasons, like not wanting to work for such a morally bankrupt company whose sole purpose is selling ads.
Genuine question, when you shop for groceries do you not look at prices at all and just buy what you want and tap your phone at the checkout?
Because to me, it seems impossible to not notice the inflation in food prices over the past year. It hasn't really changed my life, but it's very noticable that prices are going up.
No, I almost never look at prices for food. If there were no prices printed on food, it wouldn't change my shopping habits one bit. I do sometimes notice deals that are prominently displayed, like "50% off!" or "2-for-1!" and take advantage of that. If I'm ordering online instead of in-person, the grocery store's UI helpfully points out when there are coupons, and I'll use that. If it's something where I don't care at all what brand I get, I'll buy the store brand knowing that it's cheaper. But I never intentionally look at prices for the purpose of saving money. I don't know what a gallon of milk or a loaf of bread costs.
I know this is "privilege", but it's emphasizing my point, which is that these top AI engineers making 3-10x as much as me certainly aren't worrying about inflation.
> Because to me, it seems impossible to not notice the inflation in food prices over the past year. It hasn't really changed my life, but it's very noticable that prices are going up.
I earn significantly less than a FAANG employee, and I agree with GP. From groceries alone, I'd never have known we had inflation. The only increase in prices that I did notice on my own is that eating out has gotten more expensive. And that's because many meals I would have for under $10 have magically hit $10+ - a notable boundary point, and the price point where I start thinking twice about "convenience" lunches.
Depends on the city you live in. Rent here has been going up for a long time - even when inflation was near zero. So for locals, the rent increase alone is not an indicator of inflation.
When people refer to "inflation" without any additional modifiers, they're generally referring to the Consumer Price Index numbers released by the government. When the news reports that "inflation is at a 40 year high" or "inflation is at 6%", that is based on the CPI. The CPI does not include housing costs. The most visible single component of the CPI for most people is the cost of gasoline. That's the only thing that I've personally noticed costing more.
That was stated surprisingly confidently. It's entirely wrong though, the CPI does include housing costs. I don't know why you would single out gasoline as "most visible", but I guess that gives you flexibility to claim that your real meaning is whatever you want. Gasoline prices actually (currently) impact CPI about half as much as food prices and one tenth as much as shelter costs.
I am fairly senior and make tons of money. The ratio of income to my annual grocery bill is stupid. Also, I eat mostly salads and meat is not my most common protein source. Groceries could go up 10x and fluctuation in my investments would still grossly dwarf .
But one of my earliest memories is my mother crying at the ATM because the checking account was over-drawn and there was no food at home (by the time I was in high school we were much better off than most, but the treatment effect stuck -- my younger sibling and I think about money in very different ways. Crazy what a $2K savings buffer can do). I then spent a good part of my early adult life stressing about affording food & choosing between rent and desired groceries.
I still stress over every single thing that goes in the grocery basket, for absolutely not rational reason. In particular, I don't stress about e.g. restaurant bills. Why? Because the times in my life when food was scarce do not overlap with times I was in restaurants as a patron.
The grocery store in particular is a source of psychological terror for many Americans at one point or another in their lives, and those memories are visceral.
It's not on the level of food insecurity, but I see a ton of comments in this thread by people who obviously are too young to remember the dot-com bust of 2001.
I was just out of school in 1999 and can vividly recall one of the senior engineers warning me, with mortal seriousness in his eyes, "it isn't always like this". I also remember having absolutely no fucking clue what he was talking about, except on the most superficial level. He was then just about the age that I am now.
>Facebook/Meta stock lost 30% of it's value a few months back and may not recover any time soon.
I think a significant driver of that was the tightening screws re: criticism and pending federal investigation.
The brand reinvention was in the context of changing perceptions that could culminate in existential threats to the company. There is no good way out of it, but changing the conversation from "Break up facebook" to "gee that Meta sure is weird", is a mixed success, and exchanging catastrophe for mixed success made quite a bit of sense as a strategy.
When companies cap their pay increases for existing hires to a 3-4% pay increase a year and inflation on goods is close to 10% yoy and housing is jumping 20% yoy in most markets it makes sense to look for higher pay increases than what your company is locking you in at.
I'd be surprised if that's the case for such specialized workers. I know for a fact that at least some places has been giving out double digit salary increases at six month reviews.
