I’m skeptical of the construct validity of these studies claiming to show GPT levels the playing field. Okay, so you can close a gap in some contrived game that’s played once and has low stakes; cool. In the real world, where people have skin in the game and are beholden to market forces, tools like GPT will have a multiplicative rather than additive effect.
That said, I’m skeptical the effect will be as big as some are claiming. It’s another tool, like stack overflow or syntax highlighting. It will make you faster but it won’t make you more clever.
I think of it like this: My friend is a copywriter for a fitness watch company. Those little one paragraph blurbs? He writes those.
Because humans get bored easily, he rotates being specific products and media regularly; sometimes it's for a radio bit, sometimes for an online ad, sometimes for print. He's a quality professional who makes a bit over 100k. With GPT, he could *easily* quintuple his work output. Or, if you're willing to accept work that's 90% as good, a college sophomore in an english or communications program could replace him. That's true *right now* when IMO the abilities of these tools are middling to decent. By 2030, they'll be as good as human-crafted options. The job will shift from writing and curating (he'll write 100 options, narrow it down to 5 to present to the higher ups - in the future, those 100 options will be written in just minutes) the skill needed to successfully produce copy will go way down. The pool of available workers will jump considerably. Non-english speaking natives will be able to compete for these jobs, even.
I fear your model of why there is an outsized return for nerds is wrong and it undermines your claims about increasing equality. What's going on is mostly an effect of people enjoying doing what their good at and therefore investing more in that. Rather than increasing equality, AI helpers will be like the computer was to many boomers: a source of power for those who were good at it and dreed for those who felt they were bad at it.
But while you can power through a dislike of lifting heavy things and force yourself to get good at it for information/STEM type abilities you need to find it fun to play to gain understanding and that's nearly impossible if you hate it because it makes you feel dumb.
AI assistance will be like Photoshop. In theory everyone has the ability to learn to use the tool but some people will enjoy it more while others will come to fear it.
Though, the ability to do more self-paced learning without comparing yourself to others might help.
These ideas certainly make sense, Mr. Gerdes, but the basis of Noah's post is a series of studies indicating that although we might think this is true, evidence--so far uniformly, if Noah's failure to find an exception holds up--indicates it is not. It's reasonable to think that the evidence won't hold up in the long run because our intuitions are strongly inconsistent with it, but that's not an argument. One starting point for the argument you seem anxious to make might be to show why the data or analyses in these studies are flawed, which would need to be the case for your intuitions to be correct.
Except the studies don't really address the issue. Suppose you went back to the days when animation was all done by hand and you gave animators a computer to provide assistance like interpolation.
Of course it would help the slowest the most on average. Presumably slowness at drawing is relatively uncorrelated with interest in/ability to use the computer. As such the slow ppl will be most benefited.
But that doesn't translate to greater equality in the long run because what determines pay in the new world is skill with the computer not speed at drawing.
Mr. Gerdes, Your analogy supports your point, but it seems to me that you've drawn it (quickly) out of a hat: I really can't make it work.
Here's a counter: the issue with animators wasn't speed, it was the quality of the art. Relatively untalented animators now just had to tweak AI-drawn frames rather than expose their basically subpar artistry, while the brilliant artists found that AI only interfered and continued what they were doing. For corporate, good-enough art was good enough, so the pay premium for top artists declined because what determines pay in the new world is chiefly functional adequacy.
Let me make the point w/o an analogy. We have people who currently perform a task using one set of skills. A new way to do that task more efficiently shows up that relies on a different skill set. There is no reason to believe being bad at the old skill set is very highly correlated with being bad at the new skill set so we expect that they will be pulled toward the mean regardless of whether the new skill yields more or less unequal productivity.
Suppose we assign everyone a number. The first way is to give them the last 3 digits of their SSN as a number and the second method is to give them their SAT score. When we switch methods the ppl who had the lowest scores from SSN show the most improvement bc they end up at the same place on avg (no better or worse on avg than anyone else) but started at the bottom. It's the same idea here though the two measures won't be totally uncorrelated but same direction of effect.
--
If you also assume that differences are amplified over time (an eventual power computer user isn't that much different than a computer phone on the first day either of them sit at a computer) you'll also expect the new distribution to appear flatter not just have the most improvement property.
I think your second paragraph inverted SAT and SSN.
It seems to me the implications of your two examples support Noah's approach rather than undermining it. For Noah, the old skill set put a premium on "nerdiness," so nerds had their day in the sun. The new skill set gives the nerds no new advantage but advantages non-nerds, so we expect the nerd premium to decline, and that's Noah's point. (If you meant start with ranked SATs--nerds on top--and switch to SSN, with nerds and non-nerds on equal ground, the result is the same.)
Your last paragraph is different and interesting in my reading (which may not have been your intent). Studies now may show a big catch-up advantage for non-nerds, but over time it may turn out that nerds gain an advantage if the initial differences between nerd/non-nerd may give way to a different dynamic because nerdism includes dispositions that will allow nerds ultimately discover and master ways to exploit AI in tasks that non-nerds won't. (I suppose it's also possible that AI exploitation will turn out to be maximized by those with a bent towards poetry and song, just as I understand there's an intuitively surprising moderate correlation between exceptional math and musical skills.)
It doesn't matter which way you start. Going BOTH directions will show the most improvement for those who were doing the worst on the prior measure.
Since it's symmetric it can't be evidence for his claim because we'd see the same thing if we were switching from AI assisted skills to the old version. It's entirely an artifact of the fact you take the ppl who are worst at one skill and then look at a different skill that is only partially correlated with the first.
The people with SSN ending in 000 won't be any better or worse at the SAT so will on avg have the avg sat of 1050. 1050-000 is an improvement of 1050. On the other hand the avg SSN last 3 of 500 also tells us nothing about SAT score so on average they have the same 1050 resulting in an improvement of 1050 - 500 or 550.
Going thr other way, the ppl with the lowest possible SAT score of 400 will have an avg last 3 SSN of 500 so improve by 100. Those with an avg SAT of 1050 also have an avg last 3 Ssn of 500 so 'improve' by -550 (less than 100).
BOTH WAYS have the worst performers showing the greatest boost!
Besides, im not making a positive argument on this point just saying Noah's studies don't imply we'll see greater equality. Even your interpretation seems to do that.
If you've ever had the chance to listen to true scientists speak re: their discoveries, they'll point to the essence of what you seem to suggest here in that what truly matters to the scientist (and science) are observable, empirical data. Assuming the studies mentioned by Noah were performed scientifically, I would side with Noah and expect those at the lower end to increase their productivity markedly moreso than those with already-existing, higher skill sets.
> It's reasonable to think that the evidence won't hold up in the long run because our intuitions are strongly inconsistent with it, but that's not an argument.
It actually is an argument, and moreover a pretty good one.
If a scientist (and especially a social scientist) publishes a study that claims to demonstrate something that contradicts common sense, this is a good reason to doubt the study, and to request further investigation. Public understanding of science would be greatly improved if people grasped this.
In this case, the studies all take place in contrived environments that are unfaithful to the competitive nature of top tier stem work. Until we have a more realistic study, a priori arguments like Gerdes’ are more convincing than ”evidence.”
What’s the DoE look like for a study that is faithful to the competitive nature of top tier stem work?
Either way, the argument being made is about how this will effect normies. Normie is in the title of this post. Top tier STEM work is not where the normies are. Why do you need studies demonstrating the effects on top tier STEM work to support an argument about this being of benefit to low skill workers?
Either way, I’d love to see what that sort of study looks like to you, because clearly something like then has to have existed previously if you have a standard in mind.
> I’d love to see what that sort of study looks like to you, because clearly something like then has to have existed previously if you have a standard in mind.
The best possible study we could conduct to understand the effect of smoking would be a large scale RCT in which the treatment group is forced to smoke a pack a day for ten years. ”Cleary” that exists too, right?
Some things are simply hard to study, because we are either unwilling or unable to manipulate the world in a way that would expose the causality we are interested in. We are stuck relying on observational data or apriori arguments. In 2023, the former is overrated and the latter underrated.
> Why do you need studies demonstrating the effects on top tier STEM work to support an argument about this being of benefit to low skill workers?
Noah is writing about ”the widening gap in the performance of the nerds versus everyone else.”
The effect of AI on top tier stem work is half of the equation here.
Also, i don't think your studies prove what you want them to prove. Rather, they seem to be just an unsurprising instance of regression to the mean.
Almost anytime you take a task and offer people a very different quicker way to do it you'll see an apparently larger benefit for those who are the worst under the current system. But that just tells us who will do best under the new system will differ not that it will be less unequal. To measure that you need to look at how much skill at the new method varies given time to master it.
If you gave a bunch of old time craftsmen a CNC machine I bet the slowest ones show the most productivity gain and it may even look like things are more equal because no one has yet had time to develop skills in the new system. Eventual power users and computer phobes look similar in day 1.
--
Formally speaking if you have a random variable of current skill that's uncorrelated with the random variable for skill in the new system then the loweat quartile in the old skill will show the most improvement because they are a random sample as far as the new skill goes.
Yeah, there may be reversion to the mean involved. However, it also seems almost like a tautology that the AI-assisted performance will not be as much of an improvement for a top performer since they were doing it better originally. Software that lets you dictate 100 words per minute won't improve your performance if you could already type 100 words per minute, without requiring any skill differential in using the software.
But we don't want to know if AI will help the people who are currently on top as much. Of course it won't, just like computers didn't help the scientists and mathematicians who were the best when it all had to be done by hand as much as others.
But what we want to know is whether the people who are best at the new technique with AI in the future will have more of an advantage over those who are worst at it.
As I point out elsewhere, being a nerd isn't hardcoded. It's what we call the people who are best at manipulating the cutting edge tech of our world. What that means always changes over time.
A very narrow set of points on the study re: law exams:
1. Exam performance has very little to do with the skillsets required to be a superior lawyer in most settings.
2. AI assistance during an exam does not equate to long-term closure of significant performance gaps between average and good students. So much of what lawyers do is oral and extemporaneous.
3. But if AI can produce results superior to average students on items like legal research or document review, then elite lawyers will simply use AI instead of those average students for a variety of tasks currently done by humans, just as they began to do years ago with document review.
We've already eliminated thousands of jobs in the legal profession using AI in document reviews. It's hard to reconcile this with the broader hypothesis.
I don’t see any of this making Bezos any poorer. If anything it’s more likely to proletarianise all wages. The college premium falling relative to the poor is good, if the bottom 20% are growing wages, not if everybody is stagnating or declining but the poor less so.
We are all, relative to billionaires and the very top earners, poor.
I think it's an exemption for us as well as for the rich. I don't compare myself to Michael Jordan when I play basketball either. If you compare yourself to the richest people on earth, then of course nothing is going to move the needle. In that case, why worry about finishing college or getting a raise - none of it matters. It's much more meaningful to make comparisons to yourself previously or a more realistic reference class.
Eh, it’s easy to throw your money at things like a vaguely phallic rocket ship enterprise when there’s low inflation and low interest rates. I’m hopeful that in a higher interest rate environment, stupid bets and poorly managed companies will start losing again. (I know nothing about whether Blue Origin is a good company or not, I just find Bezos’ investment in it egotistical.)
Nice article... I would encourage you to consider the following point-of-view...
1) Technology drives productivity which leads to GDP/capita growth [well known]
2) GDP/capita growth makes individuals more powerful as compared to the collective.
3) Major shifts in technology cause major realignments in society.... we are going through one right now as profound as the industrial revolution. Major shifts in technology cause NON LINEAR changes in society. Economists almost always miss the nonlinear changes because they study backwards looking linearized models.
4) As the individual becomes more powerful, coercive (non-persuasive) mechanisms for control become increasingly difficult.
5) Countries with the greatest individual freedoms over a sustained period of time have an ability to statistically out-innovate their competitors. There is no guarantee because this is much like picking stocks... a few massive technology wins make the difference.
6) The role of government is to reflect the energy shifts from these technology shifts. It is a balancing act. Democracies tend to do a decent job at it. Much like a bucking bull... a stiff dictatorship has difficulty. They either kill the economic/technology bull or are bucked off.
The march of history is much better explained as: Technology => economy => politics vs the way history is typically taught with politics as the start.
Prior to industrial machinery, economic success was generally correlated to physical strength. The largest beneficiaries of the Industrial Revolution were physically weaker men.
Today, the economic losers are not physically but cognitively weaker. However, just like the industrial Revolution opened doors for the physically weaker, AI may well open doors for the cognitively weaker. If history really does rhyme, AI should undermine the college wage premium and make even the less intelligent more productive and economically useful.
It will also create huge fortunes of those few people who control the algorithms. But Teddy Roosevelt gave us a model to correct that as well.
This isn’t an article just about energy but I think this has to be pointed out since Noah seems to be such an advocate of solar. Either he is being dishonest or he doesn’t understand the stupidity of levelized cost of energy. Levelized cost of energy is such a stupid concept because it does not take into reliability issues with solar. This does not include the cost of keeping more reliable sources of power online (ie. Coal, natural gas, or nuclear primarily). Therefore, in reality you are just duplicating energy infrastructure, not replacing it. This perfectly explains why electricity prices have doubled in CA in the last decade.Levilized cost of energy is based on a scenario in which solar would have perfect reliability which is ridiculous. Now the counter to that would be well with battery storage we can store the electricity and make the infrastructure reliable. There are two issues with that. First, today’s commercial battery tech allows for at best hours of storage. We are a long ways away from full development of battery tech for longer timeframes of storage let alone full commercialization. Second, the cost of battery storage will be extraordinary which will have to be passed along as higher electricity costs to customers...primarily negatively impacting the same normies Noah talks about the most.
Thank you. Its comparing apples to oranges with solar/wind vs coal/gas/hydro/nuclear. The options are either gas backup, which is what most utilities are doing by adding token amounts of solar, mainly spurred on by renewable standards or public opinion, to distract from the fact that they're still emitting huge amounts of CO2 with their workhorse natural gas plants, or using batteries to store energy, which are also made with components mined by child labor in Sub-Saharan Africa. Transmission losses mean that power plants can't be too far from where the power is consumed, but solar and wind require large amounts of land. So what's a utility in a densely populated area supposed to do?? They're stuck in between a rock and a hard place imo, because they know they can't maintain grid reliability with wind and solar, and local citizens won't allow large amounts of land to be taken over for power production.
I recently came across this and figured I would share. It isn't perfect but it is significantly better than LCOE. Go to pgs. 15-17 for what I am referencing.
Unfortunately, I have not come across a credible article that has shown the true calculation. My guess is that such a cost may not even be something that can be calculated at an individual plant level. It may be a systemwide calculation that is required because it’s not like if you build a solar farm you can just say that the plant that offsets the intermittency is this specific natural gas plant or nuclear plant. Also, the the cost per MW likely differs a lot by location, time of day and time of year. There are so many moving factors.
What is needed with wind and solar is long duration energy storage. You can see the difference that LDES makes by comparing electricity prices in Austria, Germany, and Great Britain. Each has about 10% wind and solar, but Austria has enough pumped storage hydroelectric capacity equal to 18% of their grid size, while Germany and UK only have 3%. Electricity prices are running half as much in Austria as Germans and UK.
Its actually not surprising that AI improves the output of the below average performers. After all, it is trained on vast amount of human data and is engineered the generate the "most likely" next word. Substitute most likely for mean, median or mode or even p>0.5 and we land up somewhere in the middle of a standard bell curve. So AI adds value to the people on the left side of that bell curve and improves their out put. Not so much for the rest.
What this also means is that with the rapid adoption of AI, the supply of people with average skills would shoot up. This means that wages for such skills would actually drop.
The people on the right side of the bell curve would be impacted differently. The difference between their output and the average output would shrink dramatically. Many would respond to the pressure by upgrading while many would stagnate. The open question is whether there would a higher wage for them. If yes, the net effect over time would be shifting the entire bell curve to the right.
That is a beautiful, but incorrect, description of the relationship between generative AI capabilities and training data set.
LLMs are not trained to generate the "most likely" next word. They are trained to generate the next word. The training algorithm iterates over the data set, adjusting the model parameters to minimize the error rate.
LLM training does not try to mimic what a median person would write. It tries to mimic the authors of the training data set. The people whose writing is trained on are not median intelligence individuals.
Also, details like training for code generation and the way regularization works, plus the fine tuning process to make it a chatbot, really make it different from "just the most likely next word".
I'd advance the opposite claim and argue that we will see even greater inequality absent government intervention. The reason is because income inequality is a consequence of how hard it is to speed up a project by adding more workers and AI will only make this worse.
I mean why is it that information work has resulted in higher inequality in the first place? Why didn't it just result in more people learning to program or whatever until wages evened out? Because, while you can hire twice the construction workers and finish your building in about half the time if you double the number of programers you might eat up more time in coordinating their efforts as you gain in output. Thus, companies like Google pay a very nonlinear return for greater ability since they can't make up for quality with quantity. EDIT: you can't always halve the time for one building...I meant that if the demand for housing doubles you hire 2x the construction workers but you double the demand for searches you don't hire double the workers to improve your product since there are diminishing returns.
AI turns this process into overdrive. At google for each person with a brilliant idea you need hundreds of people to implement that code. AI will reduce that by an order of magnitude. That makes the people with the best ideas and who can beat use the AI even more relatively valuable.
The pattern with technology has been the elimination of "mid-skill" jobs. Technology saves money by either eliminating mid-skill jobs or turning them into low skill jobs (which get paid like low skill jobs).
Sometimes the "skill" in question is just willingness to do difficult work nobody wants to do. But the same technology that made that skill valuable (the assembly line) can make it obsolete (the robot assembly line).
> you can hire twice the construction workers and finish your building in about half the time
Is this really true? I thought a lot of construction is bottlenecked by things like waiting for concrete to set, waiting for delivery of parts, the speed at which machines can be operated. And there is definitely coordination overhead on a construction site.
> At google for each person with a brilliant idea you need hundreds of people to implement that code
Well no ... I worked there, most teams were small. "Hundreds of people" would be, like, all of Google Maps or something. Brilliant ideas tend to start small and often never need to grow beyond 3-5 people.
About is doing alot of work here and (very confusingly) by building I didn't mean *a* building but the overall set of buildings you want to construct. Obviously, there is some limit but for the most part if the demand for houses in the us doubles you hire 2x the workers to deal.
OTOH if the demand for finding things online doubles (eg search market becomes worth 2x as much) you don't double the workers you assign to that product. You increase it a bit and spend money on hiring the best people there to outcompete your rival because you aren't making a bunch of seperate products but one product with extremely low marginal cost of serving more users.
Oh I see. Yes if you're building a housing estate then you can just scale up because it's a lot of parallel repeated units. That argument makes sense, thanks.
The first part feels more convincing than the second part. The claims about AI equalization are interesting and novel, I haven't seen them before. It feels intuitively believable, modulo the problem with social studies being 90% flotsam. Plus I use AI, am a happy ChatGPT subscriber and have lots of ideas of what to do with it.
Green power and battery boosting isn't so convincing. I've been hearing this my whole life yet energy usage hasn't gone up, and governments are still forcing people to use drastically less based on fear generated by unreliable academic BS. The bottleneck to energy usage increasing wealth isn't technology, it's a mix of the ruling classes predilection for quasi-religious pseudoscience combined with lack of demand. If I look at my own energy usage, the only thing that'd significantly increase it would be if my country allowed powerful building-integrated air con at scale (they don't, you're only allowed portable units, for climate reasons). Again, the issue isn't how cheap solar power is, the minimization of energy usage at the cost of comfort has actually been encoded into law so people's pricing preferences don't even matter at all.
But beyond that, I'm not sure what I'd spend more energy on. I could upgrade to a bigger house but I'm not bottlenecked on energy costs for that. If energy costs halved right now my consumption would barely change, I feel like I have enough energy for what I want at the moment within the constraints of stupid regulations.
I suspect the increase in energy usage in the 20th century doesn't reflect some fundamental truth about unlimited energy demand (an implicit assumption here), but rather reflects the huge improvement in transportation technologies in that era. A car or plane just uses so much more energy than anything else it dwarfs everything else. But demand for these things is saturated now. Everyone who wants a car has one and I wouldn't do more air travel if costs fell, I already do as much as I want (I'd need more vacation time to do more travel which isn't an energy problem).
You raise great questions about the oft observed growth in energy usage. To what extent do we even want more travel and transit, or other energy usage? I think a lot, but you're right to question it.
61% of Americans say they are living paycheck to paycheck even as inflation cools. Most cannot afford a $500 emergency. Central bank officials have already raised rates 11 times, pushing the Fed’s key interest rate to a target range of 5.25% to 5.5%, the highest level in more than 22 years. Now it feels to me like the Generative A.I. hype bubble is indeed more re-distribution of wealth patterns in the spirit of a Monopolistic Silicon Valley.
Furthermore there's a lot of data that's showing the U.S. is the next Japan in terms of fertility rates. Many GenZ and Alpha cohort women won't be having any kids in tomorrow's America.
A point you have wrong is on energy. Solar and wind have gotten cheap, but they lack storage which means they cannot match up with consumer demand and that is going to create big problems. Solar and wind still do well when they are burdened with storage costs. And by storage, I mean long duration, not short duration batteries.
Wouldn't applications where solar's limitations would be less of an issue be air conditioning (where demand is naturally higher in hot sunny weather) and workplace charging of EVs?
There is still an issue, in engineering there is what is called thermal lag, which in this case refers to the fact that while solar power is strongest at noon, air conditioning demand is highest at 3pm and continues after the sun sets so there is a strong demand between 6pm to 8pm as the sun is setting and solar falls off, the “head” of the duck curve.
With solar and wind you have to have storage, look up pumped storage hydroelectric if you want to learn about long duration energy storage.
I am biased as I am working on underground pumped storage and my website is cavernenergy.com.
My mental model is quite different - I think the key distinction is between who uses AI and who builds AI, not who AI helps or doesn’t help. And I think the question is who gets the spoils? The fact that ChatGPT is (currently) free is what makes it seem utopian.
Say you start with 10,000 normie-hours of work to make a product. You invest 1,000 engineer-hours of work to make the process more productive and save 5,000 normie-hours of work. You can either
a) Pay the normies the same per-hour and half as much per-output. Pay the engineers 5x what the normies make per-hour because that’s how much value they added. No extra profit for the company.
b) Pay the engineers 2x what the normies make because that’s what the labor market demands to fill roles. The company saves 30% of its labor costs as profit which it distributed to shareholders.
c) Same as (b) except you distribute the profit to workers and pay both the engineers and the normies 43% more. No profit for the company.
We have been in the land of (a) and (b) for a while. The fact that it’s difficult to monetize ChatGPT is what’s preventing all of the gains from going to shareholders and tech workers. Otherwise it becomes yet another SaaS tool that goes into the company bottom line and funnels money into Silicon Valley.
Software engineer/coder take: I've found copilot convenient to getting the general pattern i'm looking for but plugging it in correctly and the minor tweaking to fit the existing codebase is where some of the skillset is shifting to.
Alternate take, the coding tasks that copilot is solving for will reduce the need of the low performers rather than elevate them. The higher performers will shift from spending their time coding to doing code reviews and tweaking MRs from genAI. GenAI will shift to turning product tickets and bugs into code commits for the higher performers to work on.
Though all these takes are based off assumptions companies will encourage using genAI and that engineers will integrate it into their current workflow..
I'd like to offer a few observations, some of which conflict with and some of which concur with the essay.
Just to give some idea of where I've been, my first programming course in college used punch cards. Only qualified personal (not students) were allowed anywhere near the actual computer. And that big computer had only a fraction of the computing power of today's cell phones.
Anyway, one problem I have with the essay is the questionable use of the word "education". It seems to be a common misperception that education can be measured in terms of years in school. That is not so. One person can spend sixteen years going to school and end up being able to design efficient electrical grid systems. Another person will spend sixteen years, and only have learned to make mediocre art. Another person will never finish college, and end up creating Facebook or Microsoft.
And the amount of education any of us can attain outside of school is limitless. Some of us of course do learn far more than school ever taught us. Others learn little more. So, when we say "High School Diploma" or "College Degree", how much are we really saying? Add to that, that many jobs require some sort of degree, which may freeze out people who are actually quite qualified. In my experience degrees can be more like a ticket than a measure of competence.
In spite of my master's in education, I ended up spending most of my career starting and running an architectural woodworking business. Much of what was said in this essay translates to the trades, as was mentioned. A few centuries ago, nearly all work was done with hand tools. Skilled workers spent many hours doing essentially grunt work. With powered machines, the grunt work was greatly reduced, and skilled workers could spend more time utilizing their actual skills. Productivity improved massively, lesser skilled workers could feed the machines and be more productive than a skilled worker working manually. And of course, CNC production is, at this point, an older technology. Years ago, I took a course in Solidworks, which is incredible engineering design software. I think this dovetails with the essay: Solidworks improves a lesser engineer's productivity far more than it does for a higher skilled engineer.
Some (many) will disagree, but I think we're still in the Model T era of computing. The Model T is mostly notable for being affordable to the average person, not for being a great car, especially by today's standards. You could not just get in and turn a key. You had to be knowledgeable about how cars worked, so that you knew how and when to set the choke, set the spark advance, prime if necessary, turn on the ignition, and then get out front and crank. Probably almost no one today could get a Model T started, yet ten-year-olds can intuitively start and drive today's cars. I see today's consumer software computer as being very flawed and not as intuitive as presumed. In other words, you can't just get in and turn the key. Well, why not?
I've lived thru the computer revolution from the very beginning. The advancements are amazing, but not so very different from the advancements in skilled trades, farming and manufacturing. Yes, computers are part of those advancements. As of right now, I don't see AI as being that much of a game changer. It's just one more advancement, and not so very much more of an advancement than what existed a few years ago. True enough, mediocre artists can now be better artists, but great artists will not become greater. Mediocre writers will become better, but great writers will not benefit at all. And politicians, well, they've always been hopeless.
That's a very interesting article. I had noticed how my software will save files without ever asking me where I want them. I have to go track them down and move them to where they should be. And where they should be is on a "tree" which I can follow from the trunk to the branches, or from the branches to the trunk.
In both my computer and on my bookshelves, I have items that I have completely forgotten about. Yet they can be invaluable at the right time, when I look in the right place, and there they are! Using search terms can help, but it's no replacement for an intelligent filing system, which can ALSO use search terms.
As for Model Ts, you can easily start the computer version of a Model T, because you have the proper background. But for those who don't have that background, a computer can be as impossible as a Model T. They need a computer that's as obvious as a modern car.
I do some serious swearing at my computer, every day. The cursor will jump to the pointer position with no warning. The problem comes and goes, apparently at the whim of coders at Microsoft who have trouble keeping track of such things. Boxes pop up with spelling and grammar suggestions, and cover what I'm writing, leaving me to maneuver to another point in the text and then back. Model Ts start to look simple. I subscribe to the newspaper online, and there are issues with the display that have gone unfixed for YEARS. My feeling is, before anyone writes any more software, fix the OLD software.
So, that's my vent for today. Oh, and get off my lawn!
I’m skeptical of the construct validity of these studies claiming to show GPT levels the playing field. Okay, so you can close a gap in some contrived game that’s played once and has low stakes; cool. In the real world, where people have skin in the game and are beholden to market forces, tools like GPT will have a multiplicative rather than additive effect.
That said, I’m skeptical the effect will be as big as some are claiming. It’s another tool, like stack overflow or syntax highlighting. It will make you faster but it won’t make you more clever.
I think of it like this: My friend is a copywriter for a fitness watch company. Those little one paragraph blurbs? He writes those.
Because humans get bored easily, he rotates being specific products and media regularly; sometimes it's for a radio bit, sometimes for an online ad, sometimes for print. He's a quality professional who makes a bit over 100k. With GPT, he could *easily* quintuple his work output. Or, if you're willing to accept work that's 90% as good, a college sophomore in an english or communications program could replace him. That's true *right now* when IMO the abilities of these tools are middling to decent. By 2030, they'll be as good as human-crafted options. The job will shift from writing and curating (he'll write 100 options, narrow it down to 5 to present to the higher ups - in the future, those 100 options will be written in just minutes) the skill needed to successfully produce copy will go way down. The pool of available workers will jump considerably. Non-english speaking natives will be able to compete for these jobs, even.
I fear your model of why there is an outsized return for nerds is wrong and it undermines your claims about increasing equality. What's going on is mostly an effect of people enjoying doing what their good at and therefore investing more in that. Rather than increasing equality, AI helpers will be like the computer was to many boomers: a source of power for those who were good at it and dreed for those who felt they were bad at it.
But while you can power through a dislike of lifting heavy things and force yourself to get good at it for information/STEM type abilities you need to find it fun to play to gain understanding and that's nearly impossible if you hate it because it makes you feel dumb.
AI assistance will be like Photoshop. In theory everyone has the ability to learn to use the tool but some people will enjoy it more while others will come to fear it.
Though, the ability to do more self-paced learning without comparing yourself to others might help.
These ideas certainly make sense, Mr. Gerdes, but the basis of Noah's post is a series of studies indicating that although we might think this is true, evidence--so far uniformly, if Noah's failure to find an exception holds up--indicates it is not. It's reasonable to think that the evidence won't hold up in the long run because our intuitions are strongly inconsistent with it, but that's not an argument. One starting point for the argument you seem anxious to make might be to show why the data or analyses in these studies are flawed, which would need to be the case for your intuitions to be correct.
Except the studies don't really address the issue. Suppose you went back to the days when animation was all done by hand and you gave animators a computer to provide assistance like interpolation.
Of course it would help the slowest the most on average. Presumably slowness at drawing is relatively uncorrelated with interest in/ability to use the computer. As such the slow ppl will be most benefited.
But that doesn't translate to greater equality in the long run because what determines pay in the new world is skill with the computer not speed at drawing.
Mr. Gerdes, Your analogy supports your point, but it seems to me that you've drawn it (quickly) out of a hat: I really can't make it work.
Here's a counter: the issue with animators wasn't speed, it was the quality of the art. Relatively untalented animators now just had to tweak AI-drawn frames rather than expose their basically subpar artistry, while the brilliant artists found that AI only interfered and continued what they were doing. For corporate, good-enough art was good enough, so the pay premium for top artists declined because what determines pay in the new world is chiefly functional adequacy.
Let me make the point w/o an analogy. We have people who currently perform a task using one set of skills. A new way to do that task more efficiently shows up that relies on a different skill set. There is no reason to believe being bad at the old skill set is very highly correlated with being bad at the new skill set so we expect that they will be pulled toward the mean regardless of whether the new skill yields more or less unequal productivity.
Suppose we assign everyone a number. The first way is to give them the last 3 digits of their SSN as a number and the second method is to give them their SAT score. When we switch methods the ppl who had the lowest scores from SSN show the most improvement bc they end up at the same place on avg (no better or worse on avg than anyone else) but started at the bottom. It's the same idea here though the two measures won't be totally uncorrelated but same direction of effect.
--
If you also assume that differences are amplified over time (an eventual power computer user isn't that much different than a computer phone on the first day either of them sit at a computer) you'll also expect the new distribution to appear flatter not just have the most improvement property.
I think your second paragraph inverted SAT and SSN.
It seems to me the implications of your two examples support Noah's approach rather than undermining it. For Noah, the old skill set put a premium on "nerdiness," so nerds had their day in the sun. The new skill set gives the nerds no new advantage but advantages non-nerds, so we expect the nerd premium to decline, and that's Noah's point. (If you meant start with ranked SATs--nerds on top--and switch to SSN, with nerds and non-nerds on equal ground, the result is the same.)
Your last paragraph is different and interesting in my reading (which may not have been your intent). Studies now may show a big catch-up advantage for non-nerds, but over time it may turn out that nerds gain an advantage if the initial differences between nerd/non-nerd may give way to a different dynamic because nerdism includes dispositions that will allow nerds ultimately discover and master ways to exploit AI in tasks that non-nerds won't. (I suppose it's also possible that AI exploitation will turn out to be maximized by those with a bent towards poetry and song, just as I understand there's an intuitively surprising moderate correlation between exceptional math and musical skills.)
It doesn't matter which way you start. Going BOTH directions will show the most improvement for those who were doing the worst on the prior measure.
Since it's symmetric it can't be evidence for his claim because we'd see the same thing if we were switching from AI assisted skills to the old version. It's entirely an artifact of the fact you take the ppl who are worst at one skill and then look at a different skill that is only partially correlated with the first.
The people with SSN ending in 000 won't be any better or worse at the SAT so will on avg have the avg sat of 1050. 1050-000 is an improvement of 1050. On the other hand the avg SSN last 3 of 500 also tells us nothing about SAT score so on average they have the same 1050 resulting in an improvement of 1050 - 500 or 550.
Going thr other way, the ppl with the lowest possible SAT score of 400 will have an avg last 3 SSN of 500 so improve by 100. Those with an avg SAT of 1050 also have an avg last 3 Ssn of 500 so 'improve' by -550 (less than 100).
BOTH WAYS have the worst performers showing the greatest boost!
Besides, im not making a positive argument on this point just saying Noah's studies don't imply we'll see greater equality. Even your interpretation seems to do that.
If you've ever had the chance to listen to true scientists speak re: their discoveries, they'll point to the essence of what you seem to suggest here in that what truly matters to the scientist (and science) are observable, empirical data. Assuming the studies mentioned by Noah were performed scientifically, I would side with Noah and expect those at the lower end to increase their productivity markedly moreso than those with already-existing, higher skill sets.
> It's reasonable to think that the evidence won't hold up in the long run because our intuitions are strongly inconsistent with it, but that's not an argument.
It actually is an argument, and moreover a pretty good one.
If a scientist (and especially a social scientist) publishes a study that claims to demonstrate something that contradicts common sense, this is a good reason to doubt the study, and to request further investigation. Public understanding of science would be greatly improved if people grasped this.
In this case, the studies all take place in contrived environments that are unfaithful to the competitive nature of top tier stem work. Until we have a more realistic study, a priori arguments like Gerdes’ are more convincing than ”evidence.”
What’s the DoE look like for a study that is faithful to the competitive nature of top tier stem work?
Either way, the argument being made is about how this will effect normies. Normie is in the title of this post. Top tier STEM work is not where the normies are. Why do you need studies demonstrating the effects on top tier STEM work to support an argument about this being of benefit to low skill workers?
Either way, I’d love to see what that sort of study looks like to you, because clearly something like then has to have existed previously if you have a standard in mind.
> I’d love to see what that sort of study looks like to you, because clearly something like then has to have existed previously if you have a standard in mind.
The best possible study we could conduct to understand the effect of smoking would be a large scale RCT in which the treatment group is forced to smoke a pack a day for ten years. ”Cleary” that exists too, right?
Some things are simply hard to study, because we are either unwilling or unable to manipulate the world in a way that would expose the causality we are interested in. We are stuck relying on observational data or apriori arguments. In 2023, the former is overrated and the latter underrated.
> Why do you need studies demonstrating the effects on top tier STEM work to support an argument about this being of benefit to low skill workers?
Noah is writing about ”the widening gap in the performance of the nerds versus everyone else.”
The effect of AI on top tier stem work is half of the equation here.
Also, i don't think your studies prove what you want them to prove. Rather, they seem to be just an unsurprising instance of regression to the mean.
Almost anytime you take a task and offer people a very different quicker way to do it you'll see an apparently larger benefit for those who are the worst under the current system. But that just tells us who will do best under the new system will differ not that it will be less unequal. To measure that you need to look at how much skill at the new method varies given time to master it.
If you gave a bunch of old time craftsmen a CNC machine I bet the slowest ones show the most productivity gain and it may even look like things are more equal because no one has yet had time to develop skills in the new system. Eventual power users and computer phobes look similar in day 1.
--
Formally speaking if you have a random variable of current skill that's uncorrelated with the random variable for skill in the new system then the loweat quartile in the old skill will show the most improvement because they are a random sample as far as the new skill goes.
Yeah, there may be reversion to the mean involved. However, it also seems almost like a tautology that the AI-assisted performance will not be as much of an improvement for a top performer since they were doing it better originally. Software that lets you dictate 100 words per minute won't improve your performance if you could already type 100 words per minute, without requiring any skill differential in using the software.
But we don't want to know if AI will help the people who are currently on top as much. Of course it won't, just like computers didn't help the scientists and mathematicians who were the best when it all had to be done by hand as much as others.
But what we want to know is whether the people who are best at the new technique with AI in the future will have more of an advantage over those who are worst at it.
As I point out elsewhere, being a nerd isn't hardcoded. It's what we call the people who are best at manipulating the cutting edge tech of our world. What that means always changes over time.
It sounds like you have a question that you’re interested in, which is distinct from the question we were discussing.
A very narrow set of points on the study re: law exams:
1. Exam performance has very little to do with the skillsets required to be a superior lawyer in most settings.
2. AI assistance during an exam does not equate to long-term closure of significant performance gaps between average and good students. So much of what lawyers do is oral and extemporaneous.
3. But if AI can produce results superior to average students on items like legal research or document review, then elite lawyers will simply use AI instead of those average students for a variety of tasks currently done by humans, just as they began to do years ago with document review.
We've already eliminated thousands of jobs in the legal profession using AI in document reviews. It's hard to reconcile this with the broader hypothesis.
I don’t see any of this making Bezos any poorer. If anything it’s more likely to proletarianise all wages. The college premium falling relative to the poor is good, if the bottom 20% are growing wages, not if everybody is stagnating or declining but the poor less so.
We are all, relative to billionaires and the very top earners, poor.
> We are all, relative to billionaires and the very top earners, poor.
This is why we should be comparing against previous generations of middle / working class, not Bezos.
That’s a nice exemption for the very rich.
I think it's an exemption for us as well as for the rich. I don't compare myself to Michael Jordan when I play basketball either. If you compare yourself to the richest people on earth, then of course nothing is going to move the needle. In that case, why worry about finishing college or getting a raise - none of it matters. It's much more meaningful to make comparisons to yourself previously or a more realistic reference class.
Eh, it’s easy to throw your money at things like a vaguely phallic rocket ship enterprise when there’s low inflation and low interest rates. I’m hopeful that in a higher interest rate environment, stupid bets and poorly managed companies will start losing again. (I know nothing about whether Blue Origin is a good company or not, I just find Bezos’ investment in it egotistical.)
Nothing to do with my point really. I was thinking Amazon.
Nice article... I would encourage you to consider the following point-of-view...
1) Technology drives productivity which leads to GDP/capita growth [well known]
2) GDP/capita growth makes individuals more powerful as compared to the collective.
3) Major shifts in technology cause major realignments in society.... we are going through one right now as profound as the industrial revolution. Major shifts in technology cause NON LINEAR changes in society. Economists almost always miss the nonlinear changes because they study backwards looking linearized models.
4) As the individual becomes more powerful, coercive (non-persuasive) mechanisms for control become increasingly difficult.
5) Countries with the greatest individual freedoms over a sustained period of time have an ability to statistically out-innovate their competitors. There is no guarantee because this is much like picking stocks... a few massive technology wins make the difference.
6) The role of government is to reflect the energy shifts from these technology shifts. It is a balancing act. Democracies tend to do a decent job at it. Much like a bucking bull... a stiff dictatorship has difficulty. They either kill the economic/technology bull or are bucked off.
The march of history is much better explained as: Technology => economy => politics vs the way history is typically taught with politics as the start.
Prior to industrial machinery, economic success was generally correlated to physical strength. The largest beneficiaries of the Industrial Revolution were physically weaker men.
Today, the economic losers are not physically but cognitively weaker. However, just like the industrial Revolution opened doors for the physically weaker, AI may well open doors for the cognitively weaker. If history really does rhyme, AI should undermine the college wage premium and make even the less intelligent more productive and economically useful.
It will also create huge fortunes of those few people who control the algorithms. But Teddy Roosevelt gave us a model to correct that as well.
This isn’t an article just about energy but I think this has to be pointed out since Noah seems to be such an advocate of solar. Either he is being dishonest or he doesn’t understand the stupidity of levelized cost of energy. Levelized cost of energy is such a stupid concept because it does not take into reliability issues with solar. This does not include the cost of keeping more reliable sources of power online (ie. Coal, natural gas, or nuclear primarily). Therefore, in reality you are just duplicating energy infrastructure, not replacing it. This perfectly explains why electricity prices have doubled in CA in the last decade.Levilized cost of energy is based on a scenario in which solar would have perfect reliability which is ridiculous. Now the counter to that would be well with battery storage we can store the electricity and make the infrastructure reliable. There are two issues with that. First, today’s commercial battery tech allows for at best hours of storage. We are a long ways away from full development of battery tech for longer timeframes of storage let alone full commercialization. Second, the cost of battery storage will be extraordinary which will have to be passed along as higher electricity costs to customers...primarily negatively impacting the same normies Noah talks about the most.
Thank you. Its comparing apples to oranges with solar/wind vs coal/gas/hydro/nuclear. The options are either gas backup, which is what most utilities are doing by adding token amounts of solar, mainly spurred on by renewable standards or public opinion, to distract from the fact that they're still emitting huge amounts of CO2 with their workhorse natural gas plants, or using batteries to store energy, which are also made with components mined by child labor in Sub-Saharan Africa. Transmission losses mean that power plants can't be too far from where the power is consumed, but solar and wind require large amounts of land. So what's a utility in a densely populated area supposed to do?? They're stuck in between a rock and a hard place imo, because they know they can't maintain grid reliability with wind and solar, and local citizens won't allow large amounts of land to be taken over for power production.
Do you know of a good article that puts all the numbers together to compute the real cost?
Hi Jim,
I recently came across this and figured I would share. It isn't perfect but it is significantly better than LCOE. Go to pgs. 15-17 for what I am referencing.
https://am.jpmorgan.com/content/dam/jpm-am-aem/global/campaign/energy-paper-13/growing-pains-renewable-transition-in-adolescence.pdf
Thanks!
Unfortunately, I have not come across a credible article that has shown the true calculation. My guess is that such a cost may not even be something that can be calculated at an individual plant level. It may be a systemwide calculation that is required because it’s not like if you build a solar farm you can just say that the plant that offsets the intermittency is this specific natural gas plant or nuclear plant. Also, the the cost per MW likely differs a lot by location, time of day and time of year. There are so many moving factors.
Yes, thanks for the explanation, seems right that it would be complex to calculate!
What is needed with wind and solar is long duration energy storage. You can see the difference that LDES makes by comparing electricity prices in Austria, Germany, and Great Britain. Each has about 10% wind and solar, but Austria has enough pumped storage hydroelectric capacity equal to 18% of their grid size, while Germany and UK only have 3%. Electricity prices are running half as much in Austria as Germans and UK.
Thanks for directing me to the topic of pumped hydro in Austria. The story of Kohlbrein Dam is fascinating.
Its actually not surprising that AI improves the output of the below average performers. After all, it is trained on vast amount of human data and is engineered the generate the "most likely" next word. Substitute most likely for mean, median or mode or even p>0.5 and we land up somewhere in the middle of a standard bell curve. So AI adds value to the people on the left side of that bell curve and improves their out put. Not so much for the rest.
What this also means is that with the rapid adoption of AI, the supply of people with average skills would shoot up. This means that wages for such skills would actually drop.
The people on the right side of the bell curve would be impacted differently. The difference between their output and the average output would shrink dramatically. Many would respond to the pressure by upgrading while many would stagnate. The open question is whether there would a higher wage for them. If yes, the net effect over time would be shifting the entire bell curve to the right.
That is a beautiful, but incorrect, description of the relationship between generative AI capabilities and training data set.
LLMs are not trained to generate the "most likely" next word. They are trained to generate the next word. The training algorithm iterates over the data set, adjusting the model parameters to minimize the error rate.
LLM training does not try to mimic what a median person would write. It tries to mimic the authors of the training data set. The people whose writing is trained on are not median intelligence individuals.
Also, details like training for code generation and the way regularization works, plus the fine tuning process to make it a chatbot, really make it different from "just the most likely next word".
I'd advance the opposite claim and argue that we will see even greater inequality absent government intervention. The reason is because income inequality is a consequence of how hard it is to speed up a project by adding more workers and AI will only make this worse.
I mean why is it that information work has resulted in higher inequality in the first place? Why didn't it just result in more people learning to program or whatever until wages evened out? Because, while you can hire twice the construction workers and finish your building in about half the time if you double the number of programers you might eat up more time in coordinating their efforts as you gain in output. Thus, companies like Google pay a very nonlinear return for greater ability since they can't make up for quality with quantity. EDIT: you can't always halve the time for one building...I meant that if the demand for housing doubles you hire 2x the construction workers but you double the demand for searches you don't hire double the workers to improve your product since there are diminishing returns.
AI turns this process into overdrive. At google for each person with a brilliant idea you need hundreds of people to implement that code. AI will reduce that by an order of magnitude. That makes the people with the best ideas and who can beat use the AI even more relatively valuable.
+1
The pattern with technology has been the elimination of "mid-skill" jobs. Technology saves money by either eliminating mid-skill jobs or turning them into low skill jobs (which get paid like low skill jobs).
Sometimes the "skill" in question is just willingness to do difficult work nobody wants to do. But the same technology that made that skill valuable (the assembly line) can make it obsolete (the robot assembly line).
> you can hire twice the construction workers and finish your building in about half the time
Is this really true? I thought a lot of construction is bottlenecked by things like waiting for concrete to set, waiting for delivery of parts, the speed at which machines can be operated. And there is definitely coordination overhead on a construction site.
> At google for each person with a brilliant idea you need hundreds of people to implement that code
Well no ... I worked there, most teams were small. "Hundreds of people" would be, like, all of Google Maps or something. Brilliant ideas tend to start small and often never need to grow beyond 3-5 people.
About is doing alot of work here and (very confusingly) by building I didn't mean *a* building but the overall set of buildings you want to construct. Obviously, there is some limit but for the most part if the demand for houses in the us doubles you hire 2x the workers to deal.
OTOH if the demand for finding things online doubles (eg search market becomes worth 2x as much) you don't double the workers you assign to that product. You increase it a bit and spend money on hiring the best people there to outcompete your rival because you aren't making a bunch of seperate products but one product with extremely low marginal cost of serving more users.
Oh I see. Yes if you're building a housing estate then you can just scale up because it's a lot of parallel repeated units. That argument makes sense, thanks.
The first part feels more convincing than the second part. The claims about AI equalization are interesting and novel, I haven't seen them before. It feels intuitively believable, modulo the problem with social studies being 90% flotsam. Plus I use AI, am a happy ChatGPT subscriber and have lots of ideas of what to do with it.
Green power and battery boosting isn't so convincing. I've been hearing this my whole life yet energy usage hasn't gone up, and governments are still forcing people to use drastically less based on fear generated by unreliable academic BS. The bottleneck to energy usage increasing wealth isn't technology, it's a mix of the ruling classes predilection for quasi-religious pseudoscience combined with lack of demand. If I look at my own energy usage, the only thing that'd significantly increase it would be if my country allowed powerful building-integrated air con at scale (they don't, you're only allowed portable units, for climate reasons). Again, the issue isn't how cheap solar power is, the minimization of energy usage at the cost of comfort has actually been encoded into law so people's pricing preferences don't even matter at all.
But beyond that, I'm not sure what I'd spend more energy on. I could upgrade to a bigger house but I'm not bottlenecked on energy costs for that. If energy costs halved right now my consumption would barely change, I feel like I have enough energy for what I want at the moment within the constraints of stupid regulations.
I suspect the increase in energy usage in the 20th century doesn't reflect some fundamental truth about unlimited energy demand (an implicit assumption here), but rather reflects the huge improvement in transportation technologies in that era. A car or plane just uses so much more energy than anything else it dwarfs everything else. But demand for these things is saturated now. Everyone who wants a car has one and I wouldn't do more air travel if costs fell, I already do as much as I want (I'd need more vacation time to do more travel which isn't an energy problem).
You raise great questions about the oft observed growth in energy usage. To what extent do we even want more travel and transit, or other energy usage? I think a lot, but you're right to question it.
61% of Americans say they are living paycheck to paycheck even as inflation cools. Most cannot afford a $500 emergency. Central bank officials have already raised rates 11 times, pushing the Fed’s key interest rate to a target range of 5.25% to 5.5%, the highest level in more than 22 years. Now it feels to me like the Generative A.I. hype bubble is indeed more re-distribution of wealth patterns in the spirit of a Monopolistic Silicon Valley.
Furthermore there's a lot of data that's showing the U.S. is the next Japan in terms of fertility rates. Many GenZ and Alpha cohort women won't be having any kids in tomorrow's America.
> 61% of Americans say they are living paycheck to paycheck even as inflation cools.
According to surveys by companies that sell you financial education, not according to reliable sources.
(Also, there's the effect that people lie on surveys, which is why they still report this even when they also report making over $100k.)
A point you have wrong is on energy. Solar and wind have gotten cheap, but they lack storage which means they cannot match up with consumer demand and that is going to create big problems. Solar and wind still do well when they are burdened with storage costs. And by storage, I mean long duration, not short duration batteries.
Wouldn't applications where solar's limitations would be less of an issue be air conditioning (where demand is naturally higher in hot sunny weather) and workplace charging of EVs?
There is still an issue, in engineering there is what is called thermal lag, which in this case refers to the fact that while solar power is strongest at noon, air conditioning demand is highest at 3pm and continues after the sun sets so there is a strong demand between 6pm to 8pm as the sun is setting and solar falls off, the “head” of the duck curve.
With solar and wind you have to have storage, look up pumped storage hydroelectric if you want to learn about long duration energy storage.
I am biased as I am working on underground pumped storage and my website is cavernenergy.com.
My mental model is quite different - I think the key distinction is between who uses AI and who builds AI, not who AI helps or doesn’t help. And I think the question is who gets the spoils? The fact that ChatGPT is (currently) free is what makes it seem utopian.
Say you start with 10,000 normie-hours of work to make a product. You invest 1,000 engineer-hours of work to make the process more productive and save 5,000 normie-hours of work. You can either
a) Pay the normies the same per-hour and half as much per-output. Pay the engineers 5x what the normies make per-hour because that’s how much value they added. No extra profit for the company.
b) Pay the engineers 2x what the normies make because that’s what the labor market demands to fill roles. The company saves 30% of its labor costs as profit which it distributed to shareholders.
c) Same as (b) except you distribute the profit to workers and pay both the engineers and the normies 43% more. No profit for the company.
We have been in the land of (a) and (b) for a while. The fact that it’s difficult to monetize ChatGPT is what’s preventing all of the gains from going to shareholders and tech workers. Otherwise it becomes yet another SaaS tool that goes into the company bottom line and funnels money into Silicon Valley.
ChatGPT is monetized; GPT4 is so much better that I think they have plenty of paying users for it.
Software engineer/coder take: I've found copilot convenient to getting the general pattern i'm looking for but plugging it in correctly and the minor tweaking to fit the existing codebase is where some of the skillset is shifting to.
Alternate take, the coding tasks that copilot is solving for will reduce the need of the low performers rather than elevate them. The higher performers will shift from spending their time coding to doing code reviews and tweaking MRs from genAI. GenAI will shift to turning product tickets and bugs into code commits for the higher performers to work on.
Though all these takes are based off assumptions companies will encourage using genAI and that engineers will integrate it into their current workflow..
I'd like to offer a few observations, some of which conflict with and some of which concur with the essay.
Just to give some idea of where I've been, my first programming course in college used punch cards. Only qualified personal (not students) were allowed anywhere near the actual computer. And that big computer had only a fraction of the computing power of today's cell phones.
Anyway, one problem I have with the essay is the questionable use of the word "education". It seems to be a common misperception that education can be measured in terms of years in school. That is not so. One person can spend sixteen years going to school and end up being able to design efficient electrical grid systems. Another person will spend sixteen years, and only have learned to make mediocre art. Another person will never finish college, and end up creating Facebook or Microsoft.
And the amount of education any of us can attain outside of school is limitless. Some of us of course do learn far more than school ever taught us. Others learn little more. So, when we say "High School Diploma" or "College Degree", how much are we really saying? Add to that, that many jobs require some sort of degree, which may freeze out people who are actually quite qualified. In my experience degrees can be more like a ticket than a measure of competence.
In spite of my master's in education, I ended up spending most of my career starting and running an architectural woodworking business. Much of what was said in this essay translates to the trades, as was mentioned. A few centuries ago, nearly all work was done with hand tools. Skilled workers spent many hours doing essentially grunt work. With powered machines, the grunt work was greatly reduced, and skilled workers could spend more time utilizing their actual skills. Productivity improved massively, lesser skilled workers could feed the machines and be more productive than a skilled worker working manually. And of course, CNC production is, at this point, an older technology. Years ago, I took a course in Solidworks, which is incredible engineering design software. I think this dovetails with the essay: Solidworks improves a lesser engineer's productivity far more than it does for a higher skilled engineer.
Some (many) will disagree, but I think we're still in the Model T era of computing. The Model T is mostly notable for being affordable to the average person, not for being a great car, especially by today's standards. You could not just get in and turn a key. You had to be knowledgeable about how cars worked, so that you knew how and when to set the choke, set the spark advance, prime if necessary, turn on the ignition, and then get out front and crank. Probably almost no one today could get a Model T started, yet ten-year-olds can intuitively start and drive today's cars. I see today's consumer software computer as being very flawed and not as intuitive as presumed. In other words, you can't just get in and turn the key. Well, why not?
I've lived thru the computer revolution from the very beginning. The advancements are amazing, but not so very different from the advancements in skilled trades, farming and manufacturing. Yes, computers are part of those advancements. As of right now, I don't see AI as being that much of a game changer. It's just one more advancement, and not so very much more of an advancement than what existed a few years ago. True enough, mediocre artists can now be better artists, but great artists will not become greater. Mediocre writers will become better, but great writers will not benefit at all. And politicians, well, they've always been hopeless.
I’m pretty sure we entered the post-Model-T era of computing with the iPhone. Speaking of Solidworks, here’s a fun article for you.
https://www.theverge.com/22684730/students-file-folder-directory-structure-education-gen-z
That's a very interesting article. I had noticed how my software will save files without ever asking me where I want them. I have to go track them down and move them to where they should be. And where they should be is on a "tree" which I can follow from the trunk to the branches, or from the branches to the trunk.
In both my computer and on my bookshelves, I have items that I have completely forgotten about. Yet they can be invaluable at the right time, when I look in the right place, and there they are! Using search terms can help, but it's no replacement for an intelligent filing system, which can ALSO use search terms.
As for Model Ts, you can easily start the computer version of a Model T, because you have the proper background. But for those who don't have that background, a computer can be as impossible as a Model T. They need a computer that's as obvious as a modern car.
I do some serious swearing at my computer, every day. The cursor will jump to the pointer position with no warning. The problem comes and goes, apparently at the whim of coders at Microsoft who have trouble keeping track of such things. Boxes pop up with spelling and grammar suggestions, and cover what I'm writing, leaving me to maneuver to another point in the text and then back. Model Ts start to look simple. I subscribe to the newspaper online, and there are issues with the display that have gone unfixed for YEARS. My feeling is, before anyone writes any more software, fix the OLD software.
So, that's my vent for today. Oh, and get off my lawn!
I’m sorry, substack is having display issues, so it cut off the entire part of your comment where you talk about buggy software 😂