Working in the ML/data space and have been fortunate to have largely steered clear of this problem so far. My heuristic when evaluating potential jobs is only to consider positions where the output of statistical analysis and machine learning models has a clear and immediate impact within the organization, while avoiding adtech for ethical reasons and because of concerns toxic work could enable a toxic environment.
Jobs like this exist, though you may have to take a pay cut compared to the bullshit. My first job after finishing my PhD was as a staff scientist in an academic lab using ML engineering + data science to support scientific research. There seems to be a fair number of grant supported jobs like this and pay isn’t terrible. Just under or just at 6 figures. You can make more in industry, but scientific work feels very meaningful.
Now I’m working in credit risk modeling, something I never expected to be doing, but so far it’s been a good fit. The models are applied directly in decision making for the business, and there’s a real incentive to get everything right because mistakes could harm real people’s lives. The team I’m on is strong and ethically sound and I feel good about what I’m doing.
For anyone in the data space who’s despairing about the state of the industry, non-bullshit jobs do exist, you just have to look for them, and use your judgment when scoping out new roles.
I really like your heuristic, and I'll add to your list of examples where one could work
- Any tech company where the Stats/ML model is one of THE products and differentiators. You'll need to cut through a lot of buzz word and sales-speak to find the good ones.
- Banks and other financial institutions where making uneducated guess is a big no-no when it comes risk, pricing and anything related to financial products. Your example of Credit Risk Modeling is a classic example and very interesting problem.
- As much as consultancy gets a bad rep due to some shady practice from big players, there exists a solid demand for professionals that know Statistics in the Large Construction Projects space. Let's say modeling demand and return financial for a project, proving environmental impact, preparing/implementing/analyzing unbiased surveys in the area, and so on.
- Government agencies where data is one of the Key Outputs. Such as the Census, Bureau of Labor Statistics, CDC, and so many others.
As you said, sometimes it will mean a pay cut, especially if you want to remain in the technical work and not deal with the business and managerial side of things. But there's solid demand.
+1 for government. I work with state and local governments a fair bit, and while there’s a lot of red tape and weird politics, the overwhelming majority of clients I work with are smart, capable, highly mission driven people doing their best in a system designed to move slowly. Being able to look back on a project and see a positive impact in a community is way more rewarding than trying to move the needle on click through. And the pay cut for public sector consulting isn’t as much as one would think. Especially for federal.
I’ve been doing “big data” in the retail space for almost a decade now and it’s pretty clear to me how it affects the bottom line. It really can be as simple as “we need more of this milk” and “the spicy queso is very popular”
I'm at least somewhat skeptical. The face of the last data warehousing fad in the 90s was around things like optimizing retail through things like putting diapers and beer [1] close together. But pretty much, even when there was some correlation, stores never did much about it.
I can imagine that one reason is that you also need customers to be able to find things, which is difficult when not sorted into some meaningful categories. That said, I think at the Target near me the beer and the diapers are relatively close.
Actually you want customers to not be able to find things, so they have to wander around, and 'find' other things (in large displays for example) that you want them to buy. If the beer and diapers are always together then they won't discover the new high-margin sausage rolls. So what's better for the customer from a correlated data perspective may not be better for the business from a maximizing profit perspective.
> you want customers to not be able to find things, so they have to wander around, and 'find' other thing
This seems to be the Loblaw/NoFrills model in Toronto. They relocate products with alarming frequency.
Want milk? That's in the dairy section. Want organic milk? That's in the health food section at the opposite end of the aircraft hangar sized store. Nuts? Either in the snack foods, health foods, or baking ingredients depending on unknown critera - one literally has to check all the places.
Taken to the extreme though, I'm going to be taking up employee time asking where something is and eventually stop shopping at a store if it's too frustrating.
a) most people don't take things to extremes;
b) employee time is not a problem, as they're there anyway;
c) if all grocery stores act like this (and they do), you have to keep shopping there anyway (and they know this).
So what will actually happen, with you and me and everyone else, is that we'll grumble about how they moved the cheese again, we'll wander around the store until we find it, and we'll be exposed to some more products and implicit advertising in our search, and we'll forget about it as soon as we exit the store. Over time the store will increase its profits 1.3%, and the competing store "where everything is always in the same place as last time" struggles because their prices aren't maximally optimized with the same grocery store software everyone else uses.
Eh. I have several grocery stores--and grocery departments--that I more or less favor for various reasons depending on what I'm buying. If one of them makes my life less pleasant whether because they're always moving stuff around, they overly rely on self-service, they tend to be crowded, or whatever--I'll probably go elsewhere. I do consider pricing but I don't really comparison shop so that only loosely factors in.
That's not quite what the article says. Although it doesn't explicitly say so, there was apparently a (possibly valid) correlation. But Osco didn't act on it, instead removing a bunch of slow-moving SKUs instead.
Though that also makes the point that making things easier for customers to find is probably more important than any minor co-purchase optimizations.
I think there have been more recent fads. (If fad is the word for it.) There was the story about Walmart stocking up on Pop-Tarts during hurricane season that was in the news fifteen or twenty years ago.
There was the original "Big Data" and Hadoop, Chris Anderson's "The End of Theory" and Clive Humby coining "data is the new oil" around the same time. Probably others.
This is funny to me. In the old days, this was called: cost accounting to determine the profit of any particular item you sell and market analysis to see how many of which things your customers are likely to buy. Join the two results to optimize your profits. It's not rocket (/data) science.
Actually neither of these is particularly easy to do for unorganized companies.
My household would really like it if Amazon foods merely provide the same list we bought form last time to adjust for this delivery — we absolutely hate having to start from scratch every time (and that's not even mentioning their issues with substitution).
Heck, the local wine & beer outlet lets me just open up the last purchase list and buy that again (and adjust the quantities if desired).
WTF is wrong when a local store can get it so right, yet Amazon, with it's emphasis on "always be hungry like a startup", so totally forks it up?
Substitution was the real issue I had with online grocery delivery when I last used it 15 years ago. Things would get substituted that I didn't really think were equivalent and essential ingredients for some meal would be left off the order.
It was "OK" at the time; I was on crutches and could go to the store but not easily do a full grocery shopping. But I haven't done online grocery since.
And if stores had meals pre-shopped in bags you could just remove unwanted items from. Grocery stores seem ripe for simple optimizations even without big data
Grocery stores would be better served by not being giant chains with optimization problems, at least in North America.
Better zoning laws that allow mixed-use zoning would enable more, smaller grocers embedded in neighbourhoods. Personalized service would be much easier on a smaller scale. And you wouldn't have to travel far to pick it up.
Aside from specialty shops, I'm not sure there's any great virtue in smaller markets, often with higher prices and less selection, other than there are probably more of them you can walk to if you live in a denser urban area. And my observation at least in the UK is that most of those smaller grocers are giant chains like Tesco and Sainsbury.
Not smaller markets, just operating under a large one in a more decentralized fashion. Small grocers sourcing and selling local-first will of course be more responsive to local needs and more resilient against large supply chain shocks. Not to mention cut out a lot of traffic on trucking and shipping lanes.
Being the giant chain is the optimization. Grocery stores have razor thin margins; operating at scale is the only way to keep prices reasonable.
This is also one of the few domains where people are very price sensitive, and they regularly see and think about costs. I couldn't tell you what my local automotive guy charges for an oil change, but I can absolutely tell you the price of flour and cheese.
No one is going to pay $10 for a single apple, and if you raise prices hard enough people are going to start gardening.
It's ironic that low food prices are not for ensuring equitable access to food.
Update: Go buy an apple in a food desert and I guarantee that if they have one you will pay more than a dollar for one. Grocery chains are optimized for maximizing profit not distribution.
I’d beg to differ. The current system optimizes for large grocery trips for bulk items that require long shelf stability.
Most people would agree that produce should make up the bulk of ones diet, yet it’s pretty plain to see that the footprint of the produce section doesn’t scale with the size of the store.
Produce is something best shopped for frequently, and is something I’d personally be willing to pay a higher upfront cost for the convenience of closer smaller stores where I can get in and out quickly. Because while we may pay a lower sticker price, the elephant in the room with the massive hidden cost that we pay but rarely account for: food waste.
You're also likely someone with a higher than average income, and general closeness to potential store locations. If you were poorer, or had to drive farther it's less likely you'd use those as optimization points.
Also fresh food is great for the diet and health, but when some even occurs that interrupts things like distribution of fuel, food, or power just having it around may lead you to getting hungry really quickly. This is something you have to look at on both a personal and national level. Food security can quickly lead to destabilization.
Conversely buying in bulk implies having enough room in your house to store food items.
It’s not like the typical Costco shopper is low income. Our grocery system is optimized for moderately well to do suburbanites with SUVs and large houses. Low income people tend to live in “food deserts”.
Consider the experience of shopping for food in America if you’re dependent on public transit, which a much larger proportion of the low income population are. Requiring a car to get groceries is extremely expensive.
"... this is just the rise of this new general store, and how all specialty stores look the same now, so none of them are actually really that special."
I believe you, because the store is always out of the most popular products, and nobody ever seems to figure out that they could simply shift production towards the more popular products and make more money?
On a local basis I can go to two of the same store a few miles apart and one will out of products X, Y, and Z, but the other store may be out of X, T, and R. Localized buying trends can have significant differences.
Also, there may be many other effects here. For example, if a popular product is actually popular and the pipeline to make the product is months long, well there's going to be outages. Shifting production generally isn't easy and trends pass quickly.
In addition, perceived popularity can be used to manipulate consumer pricing. "X is always out. Oh look X is in stock now, I should buy it for 50% more"
Yes, I am sure. I go to grocery stores in multiple locations and they are all much more likely to be out of a particular version of a product than other versions. It has been this way for years. They do not price this version differently than the other versions. And this is true for multiple products even in different kinds of stores.
I think some suppliers just aren't that good at adapting supply to demand.
> My heuristic when evaluating potential jobs is only to consider positions where the output of statistical analysis and machine learning models has a clear and immediate impact within the organization, while avoiding adtech for ethical reasons and because of concerns toxic work could enable a toxic environment.
Currently starting to look for a new role and couldn’t have articulated my goal more clearly. The problem is that this narrows the field considerably, to finance/insurance/fraud (“pure money work”), healthcare/EMR (which as far as I can tell is also actually just profit optimization work for hospitals), and then you have manufacturing (closer to where I currently am), but the actual applications of data science are more limited and data Eng / analytics are more valuable, bioinformatics (which is really interesting, but also a large learning curve), and then just traditional BI (kind of generic and boring).
Also, if you’re thinking about long term career development, many of these functions fall under the IT/CIO umbrella as orgs grow, which means that career advancement would require learning cloud architecture, cybersecurity, IAM, networking, etc. which I’m not saying is a bad thing, just an observation. Just applying statistics and modeling can only rise so high in an org, unless like you said, their core business and value prop is being the best at modeling some phenomena.
I’m finishing my bioinformatics-centric PhD and I really resonated with the author’s sentiments. Except in the case of the bioinformatics field, I feel like there’s too much data for any person to make sense out of. And if you do anything beyond a volcano plot, everybody’s eyes roll back into their head and they get bored and ask for their volcano plot or their special gene/protein/transcript of interest.
Tbh I dread my future job of maintaining R and python scripts
anti-fraud is very unsexy work IMO, but as long as you believe that (insert some internet-enabled service) should exist, then it can't be safe and enjoyable to use without serious fraud mitigation and prevention work, you can be secure in knowing that you're making life a lot easier for people who otherwise would have suffered from that fraud
Curious why you think antifraud is “unsexy”? I worked in the field for a few years. My experience was that the bulk of fraud is easily detected with bog standard models (most people doing fraud are lazy). But, there’s a long tail of super sophisticated work for detecting more subtle operators. It’s also pretty fun to learn about the convoluted schemes people come up with for doing fraud and to imagine how someone would attack a new service.
Also theres quite a bit of money sloshing around and it’s relatively easy to quantify financial benefit from fraud reduction, so the departments tend to be well funded
> which means that career advancement would require learning cloud architecture, cybersecurity, IAM, networking, etc.
I was going to suggest: we have a bunch of data scientists working on intrusion detection problems in cybersecurity, tied to an engineering organization instead of an IT organization. On the balance of considerations, I think it has a net positive impact. And I'm pretty sure plenty of other companies have something similar. Unfortunately, we also have an ongoing hiring freeze, so I can't recommend my own organization.
When I was starting out my career, my father gave me a piece of advice that has turned out to be incredibly applicable and valuable - "Stay close to the money". What he meant by that is work in environments where the work you perform is very closely tied to the revenue of the company - or another way of saying it - "Do work that is directly part of the organization's mission". It sounds like you've applied this to your career with good results.
Yeah, I was basically filtering out jobs applications by whether or not I could see what effect ML was supposed to have on the functioning of the business, and importantly, whether or not they actually had access to the data to make it work.
Ended up at a place where ML doesn't just make things better, but is necessary to make things work at scale.
There were a lot of job ads I looked at that just seemed dreadful to me though. So many recruiting companies wanted ML Engineers/Data scientists. I could kinda see how it would work, but didn't think I would enjoy it.
As a non-ad tech example, a significant component to what we're trying to do at https://www.auxon.io is provide tech to companies they can use for testing & analysis of robots and cyber-physical systems. From our perspective internally what we're getting is the ability to perform something akin to "materials science" for the increasingly critical software parts of these systems and construct models of their behavior to use for predictive & comparative analysis purposes.
We have a research partnership with the University of Ottawa & University of Luxembourg specifically on the statistical analysis end of things to go deep in some areas to later incorporate the findings into our products. In fact the first go around of that research cycle has already happened (https://arxiv.org/abs/2301.13807v1) and the insights are being integrated into our Deviant product (https://auxon.io/products/deviant).
It's definitely not ad-tech. It definitely has a specific applied use case. Most of our marketplace traction is in aerospace, energy, automotive, and defense. We're not immediately hiring for roles on the data analysis end of things (we're in much more immediate need of visualization & frontend help), but we will be this year.
Focusing on jobs where your output makes an immediate impact is an incredibly smart move to make sure you matter. Of course, most roles matter but there are too few managers who know how to vouch and articulate a team’s business value properly.
I actually found myself working on a credit risk modelling project on the capital markets side and it’s been great as well.
This is actually the most important career advice for tech jobs.
First junior dev job it's not as important because you just need to find something. After that you will want to be very careful to make sure the job has actual business impact, otherwise you are likely to end up in some kind of bullshit vanity project that only has to appear to work.
If you're a baseball player, you don't want to work for a football team. You'll be undervalued and hate your job.
Less metaphorically, I've always looked for jobs where my skills are aligned with the "company mission" - first at a bunch of startups, and now as an academic researcher where I get to define that mission.
- energy management (shifting loads to times when energy is cheap) for consumer/commercial/industrial use cases
- energy markets, especially power trading: often highly algorithmic, and driven by models that turn fundamentals data (weather, calendar, …) into supply/demand predictions, and from there into price predictions
Yes, that's a common approach. You can take a dataset of consumer features at the point in time when loans were opened along with information about the loan outcomes and try to predict the loan outcomes. You can't just take a kitchen sink approach with the features though because there are regulations that demand a level of explainability. To get a sense of the basics, I think the book Intelligent Credit Scoring by Naeem Siddiqi [0] is very good.
i would also like to add that modelling in credit risk is not just about yes / no answers around loan outcomes.
There are lots of other goals that are regularly modelled such as default rates, profit optimization, loss minimization, delinquency and payoff rates at specific parameters ... endless options
There are also lot of different ways in which these models are implemented ( decision trees, statistical analysis, ML... )
Some examples of (real life) projects include:
- if our institution offers this card to clients with 750 credit scores vs 790 credit scores, how does my profit move vs my losses and what the factors to limit losses while maximizing profits
- how do I minimize my costs for servicing this card while keeping profits at the max ?
- what rewards options lead to the highest number of preselected / qualifying clients taking up a product at the lowest cost
- what contact strategies are best for specific types of clients if they are late on payments - call or email or text or legal letter? which strategies are the cheapest? which strategies give what this institution considers to be the best response ? which lead to fastest full payment? fastest partial payment? which lead to getting back to a regular payment plan?
- how can we identify clients who have a lending product with us who might be on the market for another lending product in the next 12 months? in the 6 months?, those who might need a limit increase pro-actively? those who whose might need a limit decrease pro-actively?
And, one of the largest area pf credit risk evaluation is real time decisioning on transactions: 'is throwaway201606 really buying $6000 of apple products, in person, at this mall in Toronto, Canada right now when I (the system) know I he bought a daily Wendy's Spicy chicken sandwich 10 minutes ago in Dallas" and should we allow this payment
Some example of how models are used here include ( note that modelling helps establish which transactions to look at more carefully and which to ban outright among other things )
- predict what type of terminals are being targeted: scammers -> we have left bank ATM machines alone and started looked at gas pumps:
- predict where transactions of interest might come from: scammers -> we do scams on site A at Christmas, scams on site B in the summer or we do site A scams with brand Y card and do site B scams with brand Z card
- predict behaviour patterns of transactions of interest: we always test the cards we will use by purchasing a $5 'brand x' gift card online 10 minutes before
Yes, completely agree, I omitted the part about not just modeling binary outcomes for the sake of brevity. I was even considering linking the article Capital One: Exploiting an Information-Based Strategy [0], showing that one of Capital One's innovations was successfully modeling more complicated outcomes like profitability.
I agree they do exist, and this heuristic sounds sensible to me. It's the good old patio11 "go work in a profit center" situation, and I wish I had done so. I'd probably be grappling with a new existential concern, but that's life (I wish I had a job that didn't matter so I wouldn't have to worry about performance all day!).
I find it strange that AI/ML people would avoid "adtech for ethical reasons".
I really wish the websites I visited made better use of my actual history on those sites to tailor relevant ads to my interests. It seems like an ideal application of AI to me. They could do a far better job than serving me the lowest common denominator stuff they keep throwing at me.
My Twitter news feed these days:
* 10% - posts of interest from people I follow, i.e. the stuff I actually go to Twitter for
* 90% - Ads and recommendations of topics to follow that I have zero interest in
If it's going to insist on showing ads and recommending topics of interest, you'd think those could be better personalized, given that Twitter has years of my tweet, reply, and like history to train its AI on.
But no... what I get is crypto ads, Hollywood events, celebrity news, sports news, etc.
The problem is that there are a loooot of ethical implication on using your own personal data in the first place, where that goes, who has access to it, how is it handled, and so on and so on. Then advertisements isn't anything but propaganda, which has its own set of implications. And then finally we have the ever present pressure to push more and more ads, thereby making the internet in general worse and worse, so the very field of ads is in itself unethical, as it is destroying the virtual environments we are building.
Also ads != recommendations. In a sense after a while these two are also at odds with each other. Cause there is again, the ever present need to sell you more stuff.
>I find it strange that AI/ML people would avoid "adtech for ethical reasons".
You founded and run an advertising company. Are you taking the piss? Surely you've aware of the ethical issues even if you don't give a fuck about them.
Or are you just saying you categorically expect AI/ML people's interest in the tech and/or money to override the ethics?
>Perhaps we can start from the points I'm making and assess them on their own merits instead?
As far as I can tell, the points you made are:
A) It's strange people make moral choices to stay out of advertising.
B) Advertisement could be more targeted.
B is boring and I don't care.
A isn't really a point with merits to be debated.
I enjoy eating meat. But I don't call it "strange" everytime someone turns out to be vegetarian. I understand why they refrain from eating meat. It's not like I disavow the existence of the question of "should we eat meat?". I just disagree.
I find it strange that anyone would want companies to accurately model their behaviour and desires so that they might influence you to behave in a way that benefits that company.
The rare time an ad slips past my blocker, (and rarer still when I stop to think about them) I take solace whenever they are irrelevant.
> I find it strange that anyone would want companies to accurately model their behaviour and desires so that they might influence you to behave in a way that benefits that company.
If the companies in question are billing the ads by impression, as is common, they get paid whether the ad served is relevant or not. If I'm going to be served up native content or ads anyway, I far prefer they be relevant to my interests than they're not.
> I far prefer they be relevant to my interests than they're not.
These companies are trying to extract my attention and money -- both limited resources -- and if I have to see ads better ones that I have no interest in, simply out of spite.
I'm failing to see how this addresses the concern of corporate entities trying to effectively shape my behaviour, with no regard for my well being, to favour their interests.
I'm not sure what particular interests you have in mind, but if I visit a site I don't have to pay for, subsidized by advertising, and that shows me content and ads relevant and tailored to my interests, I consider my own interests pretty well served.
Sounds pretty extreme. Telling anyone about a product you've built is really a form of advertising. I don't see how society advances if your viewpoint were to be implemented and taken to its rational conclusion.
It's almost impossible to be competitive whilst remaining ethical since being unethical provides a massive advantage. In the drive to optimize profits, any hard advantage soon becomes table stakes.
Not to argue for OP, but I think it’s more of a slippery slope thing. A friend of mine does data science at an online bet site but doing more of the number crunching; and from what I gather, the advertising side get close to pandering to gambling addicts
Edit:ignore, op posted. I should stop jumping to answer so soon!
You understand that you're not the one paying for the service, so you don't get a say, right? Even once The Machine has perfect knowledge of you, the ads will not be tailored to your preferences. It will be what an advertiser has paid The Machine to show you.
> what an advertiser has paid The Machine to show you.
As Cory Doctorow has point out so eloquently in his "enshittification" series, the end point isn't even to the benefit of the advertiser. At the end state, The Machine also have perfect knowledge of the advertiser, and the adtech company can turn the full power of the The Machine to the benefit of itself, extracting value from both the audience and the advertisers.
Jobs like this exist, though you may have to take a pay cut compared to the bullshit. My first job after finishing my PhD was as a staff scientist in an academic lab using ML engineering + data science to support scientific research. There seems to be a fair number of grant supported jobs like this and pay isn’t terrible. Just under or just at 6 figures. You can make more in industry, but scientific work feels very meaningful.
Now I’m working in credit risk modeling, something I never expected to be doing, but so far it’s been a good fit. The models are applied directly in decision making for the business, and there’s a real incentive to get everything right because mistakes could harm real people’s lives. The team I’m on is strong and ethically sound and I feel good about what I’m doing.
For anyone in the data space who’s despairing about the state of the industry, non-bullshit jobs do exist, you just have to look for them, and use your judgment when scoping out new roles.