Keller Scholl and Robin Hanson on automation

25 simple job features explain over half the variance in which jobs are how automated.

The strongest job automation predictor is: Pace Determined By Speed Of Equipment.

Which job features predict job automation how did not change from 1999 to 2019.

Jobs that get more automated do not on average change in pay or employment.

Labor markets change more often due to changes in demand, relative to supply.

That is all from their newly published paper on the topic.


You'd think it would be "pace determined by speed of human" that would get humans automated out of a task, wouldn't you? But then someone whose "pace is determined by speed of equipment" is already surrounded by automation and there are likely to be people actively working to engineer them out.

Yes, I'm struggling with how "pace is determined by speed of equipment" says anything much more than that's already partially automated, such as a capital-intensive factory. Thinking of some examples: semiconductor fabs, assembly lines with robot welding, metal fabrication with CNC equipment.

But manual welding would be "pace determined by speed of human".

The really interesting point is when a task moves from one category to the other.

The speed of dismembering meat and other food processing has increased significantly, but the labor effort has not gone down, only the labor cost, but worker costs have increased so much most white native born are incapable of doing the work or are willing to pay the cost of the work. Better to be a drug addict.

I discuss that effect here:

That explanation makes sense. Thanks for the reply and link.

This already needs an update for 2020. Companies are now looking to automate like crazy after laying off a chunk of their workforce. They realize those workers were non-essential and doing this makes their business pandemic-proof.

Look at the stock market. Tech companies and their ability to either automate or sell automation completely rule the roost. Apple is now worth $1.5 trillion dollars! All the recent tech IPOs have been booming. Zoom. Zoominfo. Vroom. Stupid names but stupid amounts of money being made.

"makes their business pandemic-proof": except, presumably that neither their supply chain nor their customers will be pandemic-proof.

Companies are always looking to automate more work, but that doesn't mean they'll succeed in doing so. A lot of the low-hanging fruit (non-varying, repetitive work done in a fixed location) has long been picked. Physical work that involves even a small amount of variation, adaptation, dexterity, hand-eye coordination, motion through space, etc is proving nearly impossible to automate. People make the mistake of thinking that what feels hard for a human (playing chess at a grandmaster level, for example) that only a relatively few humans can do must certainly be a harder computational problem than, say, making a bed. But that's not the case. Chess has been automated. The work of a hotel maid hasn't been (and I don't expect will be during my lifetime).

"Physical work that involves even a small amount of variation, adaptation, dexterity, hand-eye coordination, motion through space, etc is proving nearly impossible to automate."

Forbes, Feb 26, 2019: Japan Aims To Automate All Convenience Stores By 2025 With A New RFID Technology

"RIFD uses electromagnetic fields to identify and track tags attached to objects automatically. RFID is superior to barcode, because it does not require the tag to be within the line of sight of the scanner."

This technology is expected to reduce the number of convenience store workers by over 50% according to a similar article I read two years ago because items in a basket can all be scanned at once within a couple of seconds.

"METI is hoping to extend the RFID-based automated system to drugstores and supermarkets eventually."

Cashless stores using RFID to skip the checkout is a possibility that's been kicked around for a while. It may work, but there's a political obstacle here in the U.S. Although I supposed you could still have a store without employees that took cash (essentially a giant vending machine you can walk into), although cash could gum up the works on the way out the door.

But drugstores still need pharmacists, and convenience stores need employees to clean up spilled coffee, and tend to the various donut cases and hotdog cookers and all of that. And grocery stores aren't going to get away from needing staff stocking shelves, bringing in fresh produce, doing the baking and cooking of prepared foods and slicing in the deli, etc. RFID might replace self-checkout machines, but would that really be a big win?

I wrote that 50% of convenience store jobs would disappear, not 100%, but there are estimates of 75% as well. Until we get HAL 9000 there would be a person or two per store but not three to five.

Much of what pharmacist do can be automated, which would reduce the number needed but not eliminate them for a while.

This ignores how automation has made the modern engineer compared to his or her counterpart of 30 years ago. With computer tools, a modern engineer is very likely 8-10X more productive than their counterpart from 1990 for a given task. That is, a 2020 mechanical engineer can design a much more complex product in a fraction of the time compared to their 1990 counterpart. In 1990, tools for mold flow analysis, FEA, surfacing, flex circuit forming, etc, weren't widely used.

The electrical engineering teams had their own tools, and the tools didn't talk to each other. Marrying a circuit board to a mechanical enclosure was a very manual process.

PCB design was painful as all get out. SPICE was hard to use. This is true for sofwtare too. A person writing cellphone code in C (or assembler) just wasn't very productive.

This is one of the big reasons engineers have seen their pay skyrocket. A modern engineer with modern tools can do 10-30X more in a day than their counterparts from 1990. Not only that, because the tools are so good, it's not uncommon to see electrical engineers doing circuit design, thermal analysis or writing software--much more cross discipline stuff.

Pick a product from 1990 and I'll bet a modern team could functionally recreate it in 1/20th of the time using modern silicon, processes and tools. And it'd be way better.

The best part is the tools are cheap, which means a skilled engineers keeps the productivity gains for himself. This isn't the case for a machinist. The modern machines is much faster than the old school machines because the modern machinist's boss spent $200K on a HAAS CNC machine. So, the boss captures most of the productivity win and finds lesser skilled folks to run it.

"This is one of the big reasons engineers have seen their pay skyrocket. A modern engineer with modern tools can do 10-30X more in a day than their counterparts from 1990."

Starting salary of a mechanical engineer in 2016 dollars:

1965 $55,700
1990 $57,000
2015 $62,000 (9% increase since 1990)

Starting salary of electrical engineers in 2016 dollars:

1965 $55,600
1990 $56,400
2015 $67,600 (20% increase since 1990)

Starting salary of chemical engineers:

1965 $56,300
1990 $62,300
2015 $65,800 (6% increase since 1990)

Starting salary for mathematics majors in 2016 dollars :

1965 $52,200
1990 $48,000
2015 $56,000 (17% increase since 1990)

Starting salaries for English and history majors in 2016 dollars:

1965 no data
1990 $39,000
2015 $39,000

If my thesis were flawed, how would showing the starting salary invalidate that? But even then, the average salary of experienced engineers doesn't disprove that there are engineers making $300K to $500K all over the coasts. And that's where the salary explosion has been among mostly tech companies on the coast. Starting salaries for fresh engineers out of top schools going to top companies have been north of $100K for a decade.

When I started out of a no-name school in the early 90's, I earned $75K my first year, including base OT and bonus.

There were not million dollar engineers 30 years ago--there was no company that would pay through any type of compensation that kind of bank--even very senior guys. There are a lot today.

Companies that make a lot of money are happy to share it back to the engineers. And engineers, in turn, are expected to iterate faster than ever. And the tools let them do it.

I'm just reporting. Here is more:

A mechanical engineer in the upper 10% earns $93,000 in 2016 dollars.
An electrical engineer in the upper 10% earns $105,000 in 2016 dollars.

It would be nice to compare that to 1990, and I bet it would be similar in 2016 dollars.

In my experience (Chemical process engineer, process designer, designer of chemical plant automation systems), single contributors working for manufacturers tend to top out at $120K-$140K, and that's for chemical engineers. You can make more as a consultant if you have some specialized knowledge, but consulting work tends to be feast or famine, so rate of pay has to balance inconsistency. People making more than $150K are managers, or have left technical work for finance or some other more technical field.

Are there exceptions? Sure, but not enough to move the needle.

+1, that matches my experiences in industrial engineering as well.

How about pandemics as a predictor of automation. Humans can't work if they are sick, but machines can. And machines don't spread contagions, but humans do. Au contraire: machines do get viruses (or the computers than run them do), and those viruses often spread to other computers. Indeed, a machine with a virus can do far more economic damage than a human with a virus. Did Scholl and Hanson consider the risk of "sick" machines?

There, there, Thiago. Mummy will be home soon.

They came up with a 25 factor model that explains half the variance and they are bragging about it?

Meteorologists are the worst, cramming all these different factors into their models and crowing if they can predict the weather an hour further into the future. I told them to throw out all the work they've done over the past 50 years and just come up with one big factor, but they didn't listen to me.

A metric based on 13 items was published in AER that explains maybe 10% of variance. We do much better than that.

Uh, that's 23 items.

“Labor markets change more often due to changes in demand, relative to supply.” This is a good point that’s often missed. There’s a running debate about whether manufacturing job loss is primarily due to trade or automation (both supply side factors), but when you look at the timing of when these losses happened, it seems the real cause of manufacturing job loss is neither trade nor automation, but recessions (a demand side factor). Almost all the manufacturing job loss since 2000 happened during the 2001 and 2008 recessions. Manufacturing jobs were flat between those recessions and increasing after the 2008 recession.

Other anonymous, here.

It seems obvious (as an ex-automator of jobs) that you replace the jobs you can. That is forever a function of how the wavefront of technology meets the employment landscape. Which makes it functionally unpredictable over long spans. No one knows the future inventions, no one knows the future market desires. (In the short term of you can just extend trends, which should work well enough.)

But bottom line, we've seen the futility of predicting the impact of driverless cars. Those cars are as far away as ever, and all that spilled ink .. well I hope it was fun.

The paper is free to all for a few weeks via this link:

Thank you. I was about to whinge: what is the point of linking to an article that costs $35 to read? It assumes that either readers here have a remarkable amount of disposable income, or they are deeply, deeply interested in the automation revolution hypothesis. I suppose either or both or some combination could be true, but still. Or maybe, they've sprung for an annual subscription: all twelve issues for the low, low price of $5381.

Levi Strauss was my father, I couldn't even kill him with the Rafale. For caste in France, look at Bourdieu himself! Who even talks of a Lot of jaffre!

Reversing the gaze - how come the Russian Sukhois are not conquering the Frenchies - en Inde - they are underwears.

This cochon called a lot of Jaffre - while you die in EPHAD - courtesy Levi Strauss, RIP !

We only give grants to us - ANR, ERC - do you see any figure looking different than Le Pen - Les Blanc - allez les franchouillard! Of course, the Jewish Ag-Pro confrederie is there. Hanging by the skin of their teeth! La Reine de France.

Le serviteur Pologne de France - Parlez vous Français fleuri?

My master is retired - I earn the SMIC - talent chercheur. This might cause plenty an itch in plenty of 'us'. Vive Chez Nous - La Republique de Cul.

A lot of Jaffre - le cul de france - do not keep the minitel ringing !

This useless pig doesn't even have any cul - who talks about couilles - parlez vous français - bien sur !

While the dalit is trampled by the French police - in the heart of France - Paris !

I designed and implemented automated chemical plants in the specialty chemical business for 25 years before entering academia (where I have started teaching how to automate chemical plants). My experience in the industry is that the rate of automation is limited largely by the availability of engineers who understand and are capable of automating manufacturing processes. It is not limited by the mathematical tools (AI, control algorithms). It is not limited by the availability of capital. It is not, for the most part, limited by the quality of the automation tools available (exception: computer vision has progressed a lot, making some new projects possible).

There are lots of automation opportunities out there, but few engineers capable of executing automation projects well. Automation projects tend to run overtime and budget, and have a distressing tendency to result in systems that are too fussy and erratic to be worth using.
Industrial automation is hard. We don't have enough trained engineers to make full use of the technology available. We don't train engineers how to automate processes very well, because most engineering professors only teach control mathematics, rather than design and practice.

My point is that industrial automation is rarely limited by economic factors. Our failure to exploit the technology available is largely a function of engineers being ill-prepared to adopt it. Automation projects tend to do poorly because of poor engineering. It's harder than it looks. AI isn't going to fix that.

Interesting. You should write this up in more detail somewhere.

I work in Automation engineering within the US. I think what you say is true but it's only the first half of the story. I think the other half of the story is that because automation is hard and complex that we charge a lot for what we do. So, we don't actually spend our time picking the low hanging fruit. Instead we spend it marginally increasing the automation of already highly automated facilities. When you do see an opportunity to quote work at a low tech facility, most of the time they either do a very small project or they balk at the costs. Low tech facilities tend to be low margin and can't afford the costs of paying for high end industrial integration. High margin facilities can afford the cost, know the value and keep us busy. Nor is it easy to expand. It's difficult to find the talented engineers, since if you raise salaries you have to raise the hourly rate which obviously cuts market demand.

While I just swung with the rat, vive l'arménie !

Comments for this post are closed