So a tiny handful of top performers - plausibly few enough that you could fit them all on a schoolbus - in an ultra-hot technical field are raking in huge sums, for the moment.
Unless you are a top professor at an R1 research university or have a PhD from an R1 university studying under the other top minds in the field or are otherwise in the most elite fraction of the profession, it doesn't sound like this is particularly useful information for the average practicing or aspiring data-tician. What's the typical salary for someone with, say, a two-year professional masters?
$300k. I work at Google, have a few good papers, and a couple years of professional experience, and dropped out of grad school. TBH not really clear that I'm doing much better than peers without ML expertise.
Google is still the top of the field. Many people go entire careers without publishing a paper. You are probably smarter than average and get paid accordingly.
I make 220k salary and 480k total comp as a staff software engineer (not in AI) - so this really could be either a higher performing staff engineer (salary) or a average performing senior (total comp)
And I frowned at the downvotes for the other comments. Friends, why do we turn into insecure conservative bullies when it comes to femininity? Why accept the implicit normalizing and othering of last century's favorite pronouns?
I've almost done my masters - I finished all the courses, but did not submit my graduate thesis in time - so I have what we call "absolutorium", i.e. finished the program without pursuing a degree. The other day, a friend from my alma mater wanted me to come and do a semester of programming classes as an "industry consultant". It was all arranged properly and I was getting ready with my lectures when, few weeks before I was supposed to give the first one, I became a political point in inter-faculty negotiations. People from one of the other faculties argued it's disrespectful to allow someone with mere BSc to teach students. Ultimately, my faculty had to let me go.
So yeah, that's the kind of bullshit that having an academic title may help you navigate. Whether or not it is worth the effort to get one is arguable, though.
To be honest, the academic field is about the only field in which it perfectly makes sense to finish a degree, I mean, what else can you boast with if you have none?
The other three reasons are visa (proof of "ability"), wanting to work in government (if you really, really want that) or early-stage startups that need Ivy-League credentials for their investors, because they don't have anything else to offer.
Other than that, as a software engineer in the private industry, nobody will care about any of your degrees.
Well, I care, to some extent. Like many other things, it can be used as a filter. At the least, lacking other demonstrable evidence (such as open source or industry experience which can be confirmed by referrals), a finished master or PhD shows that someone is able to finish a non-trivial project.
I have a friend who's been in grad school a decade (although to be fair he had to take several breaks for family issues - so let's say 7-8 years).
I don't think he's published a paper as a first author. He's about to now, but his advisor is not happy with it and has made it clear his work on that paper is not sufficient. His advisor is sending all sorts of signals that he has given up on him - he often ignores his emails for requests to meet, etc. And he's expressed displeasure about his progress several times. The University will require his advisor's signature every semester because he has gone beyond the university's deadlines for a PhD.
He's a very smart guy and is good at software (his PhD is not CS, though). He can get a good job with a decent (or even top) company. He is almost 40 years old. Oh, and on top of all that, given how his research is going, there's a good chance no one will hire him for his research work even if he gets a PhD.
I asked him "Why not quit and get a job?"
Him: "Man, I put so much time into it and I'll feel all that time would have gone to waste. And please don't talk to me about sunk cost!"
Me: "Did you every consider that you will finish the PhD, and then end up asking yourself why you wasted so much time pursuing it?"
I've been in his shoes, and I quit my PhD as well (although spent a few years less than he did in school). Then I went to industry and saw the types of jobs I would have gotten if I had completed. Although they paid better, I would have hated those jobs. As for the time I spent in grad school? How could it possibly be a waste? I went because I had a passion for the topic, and I learned a lot while there. You don't lose all that knowledge and experience when you quit.
In a hypothetical universe, I could've done a stats/cs double major in undergrad and dropped out of grad school after a year or two to go into data science, but in the real universe, data science wasn't a thing anyone in south louisiana had heard of in 2005, so I needed 10 years of phd+postdoc to find my way to it.
If you drop out of grad school is there a way to prove you were ever actually even in grad school? Or can anyone say they were in grad school but just didn’t finish?
There is something I see on LinkedIn from people I happen to know never graduated: they list the years they were at university and their subject. It’s not technically lying, and most people never notice.
that makes sense, just be weird if you dont finish your masters to have you bachelor, then a few years of nothing, and then your first job. I'd put the master course there too, but perhaps include a note that you didn't graduate?
I’m pretty close to just listing that I was a Stanford graduate and seeing how it affects my LinkedIn profile. Seems like no one would really bother to check anything anyway.
You have to actually be good at it, but I'm not famous, and I make over seven figures annually as an AI researcher at one of the big companies.
True talent is extremely rare as well as lucrative. That said, I'm glad I turned down openAI's offer of only $200K. Yikes. Ilya is better than I am, but not 10x better.
They say that OpenAI are forced to disclose the salaries. Does anyone have a link to the source, to the document with the disclosed salaries? I'd be curious to see the range.
Yeah I make 100k and live in Los Angeles sometimes I can't eat food I work remote so my coworkers outside cifornia don't have it too bad. But despite 30 years of experience I'm hardly worth enough to this society to let me eat food or have a roof over my head. 100k is about average for LA but at one point I had a job that paid me only 45k with no benefits demanding I generate a sentiment analysis system I literally had to stop eating to make my rent on that low pay. But in LA engineers are treated like worse than dogs, we have no social currency and if you work for an LA company they expect 140 hours for 40 hours pay far below market average.
You should be able to get more but it would probably require moving to Riot, Snap or Google. You would want multiple offers etc... I would go the personal network approach if possible with 30 years experience.
“AI Celebrities Are Making More Than $1M, Even at a Nonprofit”
There, fixed it. These are celebrities within the field and there’s nothing surprising about their pay. YouTube celebrity pays more but I suppose you take what you can get.
Key quote:
““When you hire a star, you are not just hiring a star,” Mr. Nicholson of the start-up Skymind said. “You are hiring everyone they attract. And you are paying for all the publicity they will attract.”
This is correct. There are many people who would pass up on many a hedge fund position (and the associated difference in compensation) to work with Geof, Yann, Schmidhuber, Goodfellow, Zoubin etc. People know this; the salaries honestly don't seem wholly disproportionate in light of it.
Wow... what a hyped up clickbait article! If you are a top performer in virtually any field in the world, you are likely making 10X than average norms. It doesn’t mean everyone is getting 10X. To get $1M in AI field you will need to have PhD from top institute, probably under well known adviser, might have even written a book, published award winning papers at top conferences and so on. It’s a big list and there are likely less than dozen people would qualify. And yes, they do deserve 10X income.
It’s not that author doesn’t know this. He even tries to protect himself by making sure facts are included:
At DeepMind, a London A.I. lab now owned by Google, costs for 400 employees totaled $138 million in 2016, according to the company’s annual financial filings in Britain. That translates to $345,000 per employee, including researchers and other staff.
is this just salary costs- rule of thumb is an employee costs 3x their salary. However working at the bleeding edge can increase this multiplier a lot as you need a lot lot more support (large clusters, large equipment costs etc)
I used to work at a world leading Rnd Organisation and they where trying to get the OH rate under 500%.
It's sad that the comments here mostly fall into a few categories:
- "This is only a few people, so not something that could affect the rest of us"
- "Those people had a specific skill, so the rest of us couldn't possibly make that much"
- "This is only happening in the Bay Area, so anybody living anywhere else shouldn't expect to make that much"
It's maddening. How about, instead, when we see another datapoint showing how ridiculously valuable we developers are to employers and our potential earning power, we talk about how to get one of those million dollar salaries. Because that is in fact doable, and this article highlights just one of the many ways a smart guy in his 20s could position himself so as to be making that million dollar salary within the next few years.
That would be much more productive than trying to prove to everybody that the market does in fact only ever pay $70k/year because that's what you see on that Java job ad on Monster.com.
This is a reality that developers need to take on board. And sadly, most people here are actively fighting against doing so. To their detriment.
I applaud the positive spin, but the responses are due to the headline which suggests these salaries are normal.
If an article said “supermarket checkout clerks are earning over $1m” and the article explained that a couple of retired property investors decided to work at a supermarket, would you respond the same?
The harsh reality is, very few people will have the tenacity, intelligence, and opportunities to become a top 20 AI researcher, especially in their 20s.
You're doing it again. If you truly believe you can't bring in that kind of money unless you're "top 20" then you're never going to try.
And it's to your detriment, because there are a lot more than 20 people making that kind of money in software today.
Note further that AI is in fact a subset of software. There is nothing magic about it that somebody like yourself or anybody else here can't pick up.
You seem to grudgingly accept that doing so can make you a lot more money. But still you go out of your way to insist that you, personally can't do that.
I don’t believe your statement that AI is just a subset of software. It’s almost entirely math. Furthermore, “somebody like yourself or anybody else can’t pick it up” is wildly untrue. You realize that the people in the articles are researchers, not practitioners, right? They’re the Albert Einstein’s of the industry dreaming up new deviations and architectures. I think you’re severely downplaying what they’ve achieved.
You might as well say that Michael Schumacher just drives a car pretty well, so most of us could spend a bit of time in training and win F1
The reason I go out of my way to insist I can’t do that is that I’m not in my 20s, and it would be delusional of me to think AI is “nothing magic” and I can watch some YouTube videos to become the worlds best.
Agreed. At the same time—believing one can do something is something that can help people continue to be persistent when it doesn’t come right away? It gives a goal even if never achieved. I never made it to the Olympics but I did become a scholarship-college athlete by dreaming of being an Olympian(hopefully this doesn’t seem boastful). I like to know who the people are at the top of their game. It’s a good way to measure how much I may be able to improve. I can create a gap-analysis of how to get there.
> How about, instead, when we see another datapoint showing how ridiculously valuable we developers are to employers and our potential earning power, we talk about how to get one of those million dollar salaries.
In the case of the original article, those million dollar salaries go to researchers, not to developers. I haven't heard about developers making a million a year.
Not unusual for people at the top of their game. I just turned down a 2M a year offer, and I have a friend who is making 2.9M USD. It's just a value proposition, he makes the company a lot more than 2.9 million a year, and when you negotiate you negotiate on value (and not on time) then you can justify these salaries.
What is considered to be "top of game" in deep learning in your experience... does that apply to just researchers like Ian Goodfellow who come up with completely novel methods for ML algorithms, or does it extend as far as people who are just using the methods that others developed effectively or in new ways? I know thats a weird question, but I am planning on looking into deep learning jobs after finishing my (MS) curious what the market value is for people who have experience implementing the systems , vs the people inventing new architectures. Because I won't have a PHD... It seems like somewhere along the way its a pretty extreme jump to ask for 1,2M or even 500K instead of just 100K~200K... wondering if you have any advice for how someone new might prove themselves... I guess beyond the standard stuff (have a nice github, try to replicate papers etc...)
I think a smart, hard working person who re-uses modern results from others well and in potentially new ways can create a vast amount of value. Short term, more than the top guys, as a lot of their work may be speculative, and yours would be getting-it-done. Long term, they'll invent some method that leaves you in the dust, but that's fine, just learn that too!
I respect fake-it-till-you-make to some degree, but I find too much "spinning" actually hurts the ML industry (though no doubt profitable for the spin artists). Communication style straight from the ICO world.
It's experience in the two examples I gave : I singlehandedly built my company : RAMM Science (https://ramm.science)
And he singlehandedly built one of Europes largest online betting system backends (and scaled it to 800 transactions a second using Deep Learning, Kafka and a Hadoop cluster)
Is 800 transactions per second good? It seems like a low number but what do I know. I am impressed that Europe's largest betting system only does 800 transactions per second.
That's 800 credit card transactions a second --- thats people making deposits into their accounts to bet with. (btw I think that's peak - I'm not 100% sure)
The frontends (where the casino / betting games run) do a lot more than this, but they work against the balance in the account, it's the payment gateway that's doing 800/s
It was a bit simplistic of me to summarize it as a payment gateway, it was actually 50-60% of their backend systems that was enhanced, it was called their "Big Data Project"
They do do fraud detection on the deposits, because of anti-money-laundering. The DL models also monitor various other types of abuse (bonus abuse and in-game abusers). They also are experimenting with Deep Reenforcement Learning to actually play some of the games
Well, I think random forests on a feature set probably plays a big role and is probably easier. But to use a deep neural network maybe, you might be able to get away with a convolutional network to extract higher order associations among some input set.
An online betting system has to do fraud detection on a higher level than random merchants, since they can expect to be a target for large scale non-amateur fraudsters, and just as in any other business, if your incoming payments are fraudulent, you lose that money.
Degrees don't matter. (Business focused) results and a track record of delivering do. Still impressive remuneration though -- he must be adding tens of millions in profits to the bottom line.
Degrees matter deeply among the untalented in my experience. I have one of those fancy pee HUDs, but I only whip it out when I'm dealing with the above sort.
Results don't really matter either, you just need to convince people that you are valuable. Many people in business don't have any sort of "provable" results but have excellent people skills.
A very specific skillset which will have that value until the next wave of techniques comes up.
I really don't see the point of these articles. If you're the expert corporate lawyer in mergers and acquisitions with an expertise in EU-US mergers in the agrochemical sector you'll make bank when you hit that one huge multinational client. And then that combo of skills goes back to being not particularly valuable.
I don't understand the surprises here. New grads out of undergrad these days are making 120k in base + about 250k RSUs vested in 4 years + 20k cash bonus every year. That's about 200k / year for a new grad from UNDERgrad. Ian Goodfellow invented GAN and he's paid 1M a year and people are shocked?
Genuine questions for those participating in this market (on either side):
-How is the performance of $1M AI researchers measured? Like, at their six month review, what do you look at to tell if your money is being well spent?
-How do you tell if it makes more sense to hire 1 famous AI researcher vs 10 additional junior data scientists? What sort of products are suited to different hiring strategies?
-How much of $1M AI researcher hiring is long-term/speculative (i.e., based on hopes of future products/revenue) vs short-term/measurable (immediately measurable as soon as a code commit hits the product)? Are there examples of products where profit went up by millions after a famous AI researcher was hired?
It sounds like quite a few companies are grappling with these questions today. I'm really curious to know how their thinking is evolving on this topic.
It seems more akin to building a labor war chest, i.e. filling your company with the best AI talent so that your competition can not hire them. You might not have an exact plan for what to do now, but there's no reason not to build up a talented team of engineers in a hot field when you've got effectively unlimited money and resources.
I don't think it works like this. It's a strategic investment for the company to attract talents. People want to work with smarter people than themselves.
Like they say in the article “When you hire a star, you are not just hiring a star, You are hiring everyone they attract. And you are paying for all the publicity they will attract.”
The top players in a field with a lot of commercial potential can claim high salaries, even for doing quasi-academic research mostly on things they want to research.
That said, I'm reminded the champion of the effective altruism movement made his name arguing that charities ought to rigorously audit their efficiency in turning donations into results before recommending $30m of foundation cash went to this particular speculative nonprofit research project his roommate and future brother in law happened to work at...
OpenAI is funded by a few VCs and big name tech leaders. It got to be a research incubator and pay well otherwise everyone will go off to a big tech company. To some extent they can focus on research as directed by the board as well. Otherwise you wait for academics to publish something OR for people to leave the big techs.
Oh I don't doubt OpenAI is more likely to produce useful research if can keep hold of staff. Mostly I was reminded of the people in the HN thread [1] at the time suggesting it didn't pay well, when actually it's picking up elite researchers on elite salaries and hoping something will come of it.
Some occupations such as football players or C-level executives make tens or hundreds of millions of dollars.
I am not surprised that an AI researcher can make that much.
Consider the AI researchers behind the optical recognition systems for handwritten text used in ATMs. Now you can massively deploy ATMs that read checks 24/7 without supervision... Paying $1m 2018 dollars to those guys is not a proportional compensation with respect to the amount of wealth they brought in.
He opened up two subfields of AI, GANs and adversarial attacks, and coauthored one of the most recommended books. His fame attracts more researchers and for lower salary, just to work with him.
GANs (which he invented) are a game changer, and have enabled things to be built which I thought were 20 years away.
For example, the life-like faces NVidia research created in Progressive Growing of GANs for Improved Quality, Stability, and Variation[1] are just an incredible achievement (it's worth looking at the pics in that paper if you haven't seen them).
Approaches like that have real-life impact in things like games, architecture, design, film, etc (and that is just the obvious implications for images). GANs of course can be used outside images and they have just as much potential there.
Anywhere an auto-encoder is used now can probably be done better with a GAN.
To the degree that this is true, my impression is that it's due to poor fundraising departments at those organizations, which appears common enough to be a stereotype. "Nonprofit" really only affects how an organization's taxes are reported, all in all.
People who are really good at what they do often thinks what they do isn't worth much because it's easy for them. It's part of why they are good at it.
>Mr. Zaremba said big tech companies were offering him two or three times what he believed his real market value was.
By definition, if he was offered some amount then his market value is at least that amount. That's literally what the term "market" is doing. Maybe he was offered above his value, but not above "market" value, by definition.
AI is driving cars and planes, beating experts at disease diagnostic, ruining popular competitive board games... But as the definition goes, anything that can be done today is already old hat.
I know someone starting a masters under Yoshua Bengio at UdeM. It's been a great school for statistical machine translation for quite a while. I wonder what he'll be making in ten years.
He might be trying to say that there are enough people like him in the market that it really shouldn't be a seller's market; but those others are mostly undervalued/ignored, while his value has been artificially inflated because of the (incorrectly!) perceived rarity of his skills.
This is the pattern that goes on all the time inside companies: constant attempts to hire talent from the outside, coupled with complete ignorance of the skills (perhaps developed post-hire) of their existing employees. I know more people than I can count who know everything required to be doing "data science" for a company, but—since they were hired to be maintenance dev-ops people—they will never be considered for "promotion into" a data-science role when there's a vacancy.
2) show me some code you have written and deployed
3) here's Learn Python the Hard Way, start reading and come back with questions.
In the modern era, I have not found anyone who can pass test 1 and doesn't know code (except for 1 mathematician who's like 70 and no one cares if he can write code, his ideas are so insanely good that other people write his code for him, despite their day jobs). These people execute and come back with challenging problems for me.
Anyone who passes test 2 seems to be able to pick up whatever task I give them. I'm not entirely sure they understand what they're doing all the way down, so there's more review, but they can generally execute. Questions go back and forth.
If I get to test 3, well, no one has come back with a question yet.
What this tells me is there is plenty of signal in the culture that math and programming ability are valued. The people don't, ain't never gunna.
If you know linear algebra, machine learning is not much further to go. You can understand pretty much all the models by looking under the hood at the math.
At that point it’s just experience to learn the rules of thumb that guide practical implementations.
Haha. I guess if I had to ask any question about python, it would be, "why on earth did they do that ridiculous unicode thing in version 3?" Most builtins should use bytes, and the few that can't should just use bytes annotated with an encoding.
Most builtins can't use bytes instead of string safely for all data, and the problem is that it propagates - in Python 2 you often have the case that library A uses library B that uses a builtin that treats a piece of text like a string of bytes, and so you can't use library A because it will give you broken results in certain conditions. We've spent time updating some third party open source libraries to support python 3, and that was well worth the time to avoid the waste of programmer time that'd come from working in python 2 due to lack of sane handling of unicode strings; and if you're working with user-facing software (as opposed to, say, scripts for system administration or physics calculations), pretty much every string you encounter nowadays is unicode string. Names of people, names of files, contents of files, results of http requests, results of database requests - all of those can be treated as streams of bytes only if you treat them as a single atomic token and don't look inside that stream in any way whatsoever. As soon as you make the first index operation, substring or split, you can't treat them as bytes anymore safely.
However, you can simply think of it as syntactic sugar that manages the encoding annotations in the default case. Where is it creating problems for you? Is it some performance hit or something else?
I think there are a lot of deep learning dilettentes, and a large number of them on HN, overestimating their abilities relative to the experts being hired by google. The guy who is stuck doing dev ops but can be doing data science is not the 1mm+ guy at google even though some of them find it hard to accept
True, but they probably are worth more to their own companies—especially mid-sized companies—than those companies realize. Whether or not the market would pay more for them, more money could certainly be being made off their backs (as with the “administrative assistant” roles at medium-sized companies that often end up substituting for much of the role of HR, PR copywriting, office management and often even team coordination. All for the cost of a secretary!+)
Not harnessing these untapped assets you already possess is a failure for a company, in much the same way that not shaving cost centers or negotiating purchases would be.
+ Yes, I’m trying to hint here that one would be crazy to allow oneself to be exploited in this manner. If you truly have the ability to do these additional jobs, then you should be applying for roles that explicitly, rather than implicitly, use those skills, and offer compensation for them. Sadly—for the same reason many people find it hard to negotiate salary—many people won’t try for jobs that no one has told them they’re “allowed” to apply for.
That’s correct, but the pragmatic reality of most companies and data science with respect to internal politics and other bullshit is that they’re not going to find that mobility even if they network internally
Funny -- I actually see that as the single most important factor distinguishing folks making $1MM and $100k. There are plenty of overpaid geniuses, but really even more underpaid geniuses.
I am not in agreement- while I’m sure there are a few, there’s quite likely more people who believe themselves to be underpaid geniuses.
The overpaid genius, at the very least, has a preponderance of evidence that he understood all the math he’s using and has evidence he can innovate with it rather than regurgitating code (maybe that’s slightly too lenient)
> The overpaid genius, at the very least, has a preponderance of evidence that he understood all the math he’s using and has evidence he can innovate with it rather than regurgitating code (maybe that’s slightly too lenient)
This assumes a separating equilibrium. There is one, of course, but it's biased against genius. Businesses don't want to overpay. Not should they. People don't negotiate their salaries. They should.
Not sure what you mean by separating equilibrium... bimodal? I think we’re talking about different things. I’m comparing (in response to the parent) the devops guy with a few dl/ml side projects who is possibly skilled enough to join a data scientist team and contribute vs. a high ranking stem phd with a few years of experience who the devops guy may be supporting. Both can certainly undernegotiate, but they’re in different situations/roles usually.
The difference between the two you mention could be qualitative, as you imply, or it could be that some of the DS folks send the wrong signal. Not choosing a top-tier AI university would be a poor signal, and fail to differentiate quality candidates of equal ability.
It comes down to whether the mental model of meritocracy is actually practiced by the business world, which also includes whether screening performed by employers is accurate. It's not, ergo it stands to reason there are poor AI geniuses too.
>Otherwise they might as well not know how to do it.
This is some real crazy talk here. The idea that you should limit yourself to the particular set of knowledge you wish to make money off of is insane.
There's almost no such thing as wasted learning, even if you're not interested in pursuing it for a career. Maybe it's a hobby, maybe you touch it in a tangential way for your normal work where a basic understanding brings value but is not necessary.
This is the same attitude that undervalues previous experience and builds in favor of specific lingual or stack competence.
>> Otherwise they might as well not know how to do it.
> This is some real crazy talk here. The idea that you should limit yourself to the particular set of knowledge you wish to make money off of is insane.
Crazy talk? That's what you get when you blatantly ignore half of what I said. In the previous sentence I quite literally asked if these want to be doing data science to begin with. If they don't have any interest in it, then it doesn't matter whether or not they know how to do it as far as the company is concerned; they're not going to be doing it either way. Nowhere did I ever suggest that it's somehow a good idea for people to be limiting themselves to a particular knowledge set.
Sure, his statement contradicts the classical definition of market, but we all know markets can become divorced from reality during bubbles. I'm assuming that was his real point.
Perhaps a more charitable explanation: Mr. Zaremba has expert knowledge of statistical analysis and hence what his true value is (in terms of return on his salary, and in comparison to other equivalent candidates). The current market, though, is essentially in a "bubble" for AI candidates, and so is mispricing the actual value of this "asset".
If Bill Gates takes a fancy to a $100,000 mountain cabin and the owner doesn’t want to sell, he could offer $10 million for it, but that doesn’t increase the market value of the cabin.
Real-estate runs on comps, and the sellers of other houses in the neighborhood love to use the overvalued ones to boost their asking price. In my experience, homeowners always nosily follow their neighborhood sales hoping for overvalued prices because they bring up the going rate in the neighborhood.
Plus, if Bill Gates buys a hours in your neighborhood, it's more or less guaranteed that the prices go up on name recognition alone.
Depends on the fungibility of cabins. If the $100,000 cabin next to yours sells for $10 million, you'd better believe it's going to change your cabin's appraisal.
In the context of real estate, what the remaining potential buyers are willing to pay after Bill Gates spends ten million on a hundred thousand dollar property.
If someone offered me three times what I think I’m worth I would assume something extremely illegal is going on. Like if you talk a ot this they will never find your body.
Unless you are a top professor at an R1 research university or have a PhD from an R1 university studying under the other top minds in the field or are otherwise in the most elite fraction of the profession, it doesn't sound like this is particularly useful information for the average practicing or aspiring data-tician. What's the typical salary for someone with, say, a two-year professional masters?