The economics of AI-powered dev efficiency
If you've used AI to program, you know it speeds you up. It's not replacing you anytime soon, but it is making you write code faster.
There are plenty of doubters out there questioning whether it's actually helping developers ship a better product and arguing that AI-generated code is of lower quality than what a senior developer would write. This critique is mostly valid.
I'm both more productive, but the quality of my code is slightly worse than if I handcrafted every line.
Is that bad? Well, not necessarily—especially if the entire business gets shut down (read On Good Tech Debt). If you view your job solely as writing code, AI might not seem like a game-changer. But if you see your role as making a business successful and sharing in its success, AI is a huge enabler.
At previous companies, I saw two lines of business shut down. If AI had helped us ship 10x faster, we likely would have just realized there was no viable market sooner, saving time and resources, and perhaps even identifying a different market or opportunity.
But what about in large, established companies? Shouldn't they be seeing gains in their number of Jira tickets completed per week?
My theory is that these productivity gains are mostly being enjoyed by individual contributors for a few reasons (though this is likely to change over the next few years).
1. Corporate Sandbagging
So this is a slightly pessimistic take, but why ship more for the same salary? Most developer jobs don't offer much upside if a product is massively successful, nor do they provide much downside (other than potential layoffs) if a product flops for years (aside — sales engineering is a great way to get a piece of the upside).
From an engineering manager's perspective, how can they suddenly expect 5x output in 2024 when the expectation was 1x output in 2023? Performance reviews typically baseline YoY performance, expecting a "reasonable" percentage increase. A 20% increase might be enough to earn an "exceeds expectations" rating, so many developers will take AI-driven gains, work a bit less, and still achieve great performance reviews.
However, performance reviews are more zero-sum than they seem. While you should be evaluated against your own previous performance, in practice you are evaluated relative to your peers. And now it's not just peers but against AI automation & peers leaning into AI automation.
Where does this lead? AI-assisted productivity will become table stakes, and the advantage will be competed away as YoY productivity gains round out. If you increase output by another 20% each year, within four years, you've doubled your output from the baseline. But at large corporations, the bottleneck is rarely just coding; processes, bureaucracy, and slow decisions grind progress to a hault.
2. Quality
AI does write worse code than a senior developer. It frequently gets abstractions wrong unless guided properly. It relies on outdated framework principles and tends to replicate common mistakes seen across existing codebases (which makes sense given its training data).
To mitigate this, companies need more guardrails: more tests, both unit and E2E, better monitoring, and improved AI evaluation metrics.
AI can assist in writing tests, but test coverage and quality assurance remain human-driven. Teams will need to develop more robust review processes to prevent AI-generated technical debt from accumulating.
3. Too large of a diff
Great programmers have always excelled at holding large portions of a codebase in their heads. However, there is a natural ceiling to how much change a developer can manage in a single day while still keeping the entire system in context.
AI allows developers to write 10x more code in a day, but it does not improve their ability to retain a mental model of a complex codebase. This is a fundamental limitation that prevents AI-driven coding from scaling linearly with AI-assisted speed.
This is probably why all IndieHackers aren't shipping 10x faster, though it does seem like the best are shipping quicker. Levelsio has many successful net new businesses. I see lots of profitable AI bootstrapped SaaS on my timeline quite frequently now.
Many developers have internalized a risk barometer for how many changes they can confidently make in a day. The best new AI enabled devs have a wider aperture and are more comfortable shipping a larger diff in a single day.
Example: the v0 and AI SDK team at Vercel. They are comfortable moving at a pace others would say is impossible. And they do it every day, not just around big releases (p.s. I'm hiring)
You Can Just Ship Things
In short, it's difficult for large companies to embrace productivity changes.
Do not go gentle into that giant scrum meeting.
You can just ship things. And quickly. And you should.