“Should my children learn how to code?” a friend has recently asked for my opinion. With the widespread sense that AI is replacing software developers, I got where he was coming from. “It’s… complicated,” was the best short answer I could come up with. The longer answer is, well, the following article.
When a shock hits, value flows to both extremes. That’s what happened in industry after industry following a major technological change, and it looks like this is what’s happening to the profession of software development with the introduction of AI-assisted coding.
Thanks for reading Matan’s Blog! Subscribe for free to receive new posts and support my work.
The Smiling Curve
Acer founder Stan Shih was the one who coined the Smiling Curve term in the early 1990s, when he noticed that in the personal computer industry, money was made in the two ends of the value chain, and not much in the middle. I learned this from Stratechery, where Ben Thompson explained the concept back in 2014:
From Wikipedia:
A smiling curve is an illustration of value-adding potentials of different components of the value chain in an IT-related manufacturing industry…According to Shih’s observation, in the personal computer industry, both ends of the value chain command higher values added to the product than the middle part of the value chain. If this phenomenon is presented in a graph with a Y-axis for value-added and an X-axis for value chain (stage of production), the resulting curve appears like a “smile”.
What makes this observation particularly ironic is that Acer is the epitomical company at the bottom of the curve. They put PCs together, but it was the critical component makers like Intel and Windows that captured most of the value on the left, and systems integrators and value-added resellers like IBM or Accenture that captured the rest of the value on the right. Acer and its merry band of 8 OEMs competed themselves to single digit margins and ultimately stagnant growth; there simply isn’t any money in the undifferentiated middle.
Thompson’s insight was that the Smiling Curve framework also describes what happened to value chains in industries affected by the internet; in publishing, individual writers (such as Ben Thompson) or niche newsletters were able to make money by directly selling to a small audience – on the internet – with very low overhead costs. On the other end of the value chain, major platforms such as Google and Facebook are making tons of money through the Aggregation playbook. In the middle, traditional publishers are struggling to make any money. Just like Acer and the “merry band of OEMs” were struggling in the PC industry, 30 years back. It’s obvious at this point how the internet killed the newspaper business model, but it was quite insightful when Thompson analyzed the economic drivers behind it back in 2014.
Source: Stratechery
The Stratechery page dedicated to the Smiling Curve links to articles that apply the concept to a variety of other industries and cases. It’s a fascinating read. Once you look through these lenses, you start seeing the smiling curve everywhere. For example —
Human Translators and AI
Generative AI has proven useful for coding in 2021, when GitHub Copilot was announced; about four years earlier, the transformer was originally first used for translation. AI’s impact on the translation industry – which Timothy B. Lee described in How human translators are coping with competition from powerful AI – can be interpreted via the smiling curve model; it may also provide hints as to how AI's impact on the software development profession is going to play out. Here’s Lee:
“We’ve been ‘in danger’ of being taken over by AI for 10 years now and it still hasn’t happened,” Eybert-Guillon said. “But we keep getting told that it’s going to happen.”
There are two big reasons AI hasn’t put many human translators out of work. First, human translators still do a better job in specialized fields like law and medicine. Translation errors in these fields can be very expensive, so clients are willing to pay extra for a human-quality translation.
Second, there has been rapid growth in hybrid translation services where a computer produces a first draft and a human translator checks it for errors. These hybrid services tend to be about 40 percent cheaper than a conventional human translation, and customers have taken advantage of that discount to translate more documents. Translators get paid less per word, but they’re able to translate more words per hour.
But while AI software has not put human translators out of work the way pessimists might have predicted, this isn’t an entirely positive story for translators either.
Can you spot the smiling curve? On one extreme, experts specialized in critical fields, where mistakes cannot be tolerated, and AI is not good enough. Lee’s post elaborates on some interesting examples, such as video games looking to maintain a specific style; human translators can still make a living in this type of area. The other end of the curve – hybrid translation services – isn’t as straightforward.
Jevons paradox, a 160-year-old economic concept, became popular last February when Satya Nadella invoked it1 as a positive way to look at DeepSeek-R1, a high-performant Chinese AI model trained at a fraction of the cost of similar American models. The same Jevons paradox has been playing out with regards to human translators.
The arrival of AI drove down the cost per translated word, but the overall spend did not decrease; on the contrary, lower prices unleashed induced demand for translation services. The overall market has grown, but translators are making less money; according to the U.S. Bureau of Labor Statistics, there were 53,150 employed as translators and interpreters in 2017 (with the onset of transformer-based translation), making a median hourly wage of $22.69. Employment rose to 78,300 by 2024, but the median hourly wage – which increased by 26% to $28.58 – did not keep up with inflation. The median hourly wage across all occupations rose by 31%, while CPI grew by ~28%, during this period.
That’s Jevons paradox: prices went down, but overall spend – instead of decreasing – went up. Because demand grew so much faster.
What about the area in the middle of the curve? Those would be traditional translators who aren’t adopting AI, but also aren’t specialized in a field where AI isn’t good enough. Things aren’t looking particularly rosy around there:
… this kind of [hybrid translation] service typically costs about 60 percent as much as a traditional human translation. And many clients take advantage of this discount to have more text translated.
While translators don’t like that discount, Lommel argues that the technology will enable them to translate more words per hour. In the long run, this might even lead to higher incomes.
Many of the translators I spoke to remain skeptical of this approach. They told me that machine translations are often so bad that it takes more work to fix them than it would have taken to produce a translation from scratch. Some found the experience so frustrating that they’ve stopped accepting this kind of work.
Yet Lommel says his firm’s data shows that machine-augmented translation is growing faster than conventional human translation. Because customers who buy cheaper semi-automated translation services don’t care as much about quality, these services can probably get away with using less skilled translators.
The whole post makes for a fascinating read. The bottom line: as AI transforms the translation industry, the two ways to earn money are either specializing in a high-quality niche, or going to the other extreme and fully embracing AI for high-velocity output at a good-enough quality. In the middle, skilled translators, who wrinkle their noses at AI translation outputs, are having a hard time making a living. The software engineering industry may evolve in a similar way.
So I like this idea of barrels and ammunition. Most companies, once they get into hiring mode…just hire a lot of people, you expect that when you add more people your horsepower or your velocity of shipping things is going to increase. Turns out it doesn’t work that way. When you hire more engineers you don’t get that much more done. You actually sometimes get less done. You hire more designers, you definitely don’t get more done, you get less done in a day.
The reason why is because most great people actually are ammunition. But what you need in your company are barrels. And you can only shoot through the number of unique barrels that you have. That’s how the velocity of your company improves is adding barrels. Then you stock them with ammunition, then you can do a lot. You go from one barrel company, which is mostly how you start, to a two barrel company, suddenly you get twice as many things done in a day, per week, per quarter. If you go to three barrels, great. If you go to four barrels, awesome. Barrels are very difficult to find. But when you have them, give them lots of equity. Promote them, take them to dinner every week, because they are virtually irreplaceable. They are also very culturally specific. So a barrel at one company may not be a barrel at another company because one of the ways, the definition of a barrel is, they can take an idea from conception and take it all the way to shipping and bring people with them. And that’s a very cultural skill set.
Ben Thompson invoked this in the context of AI, and described how inference-time scaling turns models such as OpenAI’s o3 into ammunition when it comes to writing code; that makes the barrels (also referred to as 10x engineers, or A-players) so much more efficient: there’s more AI coding ammunition, that is faster and cheaper, compared to human ammunition coders.
So that’s one end of the spectrum: tools such as Cursor or Claude Code allowing the top engineering talent - technical leaders of whole teams and departments – to be so much more productive.
What about the other end of the spectrum? No-code AI-based tools such as Bolt or Lovable are allowing anyone – regardless of technical background – to quickly and easily generate web applications. Via a single-shot prompt, often made on their phone. If you haven’t tried these tools - go ahead and try one of them. It’s… as mind blowing as taking your first Waymo ride. (more on these tools in my next post).
There are many issues, of course: the apps might not be secure, might not scale very well, the codebase might be hard to maintain, and so on; on the other hand, though, low quality software is in many cases better than no software at all. Just like with hybrid translation services, lower barriers and costs can unleash previously-induced demand for software applications.
So that’s another potential way to thrive in the post-AI software world: excel at shipping and maintaining AI-generated apps. Traditional software engineers might not consider you a real software developer, but that’s fine; when I just started my career in software, grumpy old C++ folks used to tease us Java developers, saying we’re not real engineers. “you have no control over the memory”, “no idea how the system resources are managed”, is stuff they used to say. A few years later, however, the Java team was building the company’s new products, whereas the C++ folks were stuck maintaining legacy software. Some of them were flexible enough, and eventually reached out and asked me to teach them Java.
This time we might end up with AI software builders who have no idea what their code looks like. And, as it is now me in the position of the grumpy older software developer – I have many reservations about this! but maybe it’s fine. That’s simply the way progress happens.
So there’s our smiling curve, describing software development in the age of AI:
On one extreme, the 10x A-Player engineers, the barrels shooting AI-based ammo to develop complex and mission-critical systems. Think enterprise billing software, or training the AI models themselves. On the other end: “prompt engineers”. The equivalent of the AI-assisted translators. Jevons paradox might lead to them making lower wages compared to pre-AI software engineers, but there will be demand for many more of them.
What about the middle ground? These are traditional software engineers, not necessarily the uber technical leaders of the right-hand side, but also not excited about AI-generated code. There might not be much future there. I hate to break it to you, but the value seems to accrue at the extremes. That’s where we – myself, my friend’s children, older C developers – should probably all try to get to at this point. One of the extremes. Before we find ourselves sitting with the grumpy engineers from my first job, wrinkling our noses and complaining about what happened to software.
Loved this piece - I was surprised when I asked a friend who works as a production engineer if their company uses Cursor or any vibe coding tools and it was a hard "no".
I suspect some technology teams will encourage developers to be more AI savvy while those more anxious over security risks will back away from it.
Loved this piece - I was surprised when I asked a friend who works as a production engineer if their company uses Cursor or any vibe coding tools and it was a hard "no".
I suspect some technology teams will encourage developers to be more AI savvy while those more anxious over security risks will back away from it.
Very interesting!