No One Knows If AI Will Replace Software Engineers
Seven days ago marked a year since Cognition Labs announced Devin, the AI software engineer. While Devin itself may not have achieved wide adoption, software development has changed considerably in the last year as AI coding tools have become more powerful.
Devin, the AI software engineer announced by Cognition Labs
To me, the announcement of Devin was a culturally significant moment. For the first time ever, software engineers were confronted with the possibility that a machine might soon do their job better than they can [1]. Since then, the future of software development and the software engineering career have become a widely discussed topic.
The reality is that no one really knows whether software engineers will be replaced by AI. There are a few possibilities:
| Bull Case | Bear Case | |
|---|---|---|
| AI Surpasses Humans | Increased demand for engineers due to Jevons paradox | Reduced engineering jobs and lower salaries |
| AI Plateaus | More engineering jobs from productivity gains | Moderate decrease in engineering roles especially at companies where software is not the product. |
Each scenario remains possible.
We build AI tools for developers, so we often think about what the future of software engineering will look like. So far, I only have one prediction.
My Only Prediction
My only high-confidence theory is that AI will dramatically lower the cost of producing software, allowing for highly specialized software companies to profitably service extremely niche markets.
History provides a useful analogy. Consider the video entertainment industry—like software, it consists of numerous organizations of various sizes producing permissionless, easily distributed media.
In the early days, the only types of video entertainment were TV and movies, both of which were highly resource and labor-intensive to produce and distribute, and were only made by a handful of large studios. They also needed to have wide appeal - with the costs being as high as they were, your target audience needed to be as large as "adults" or "children" to make a profit.
At some point, the cost of cameras, computers, and editing software got low enough that anyone could make an entertaining video. Today, YouTube is the dominant form of video entertainment. The median cost of producing and distributing a successful YouTube video is nearly zero. The target audience can be extremely specific. You can make YouTube videos for people interested in deep dives into World Barista Championship strategy, for Germans passionate about raw denim, and for fans of the show 30 Rock. Content this specific can be made profitably and is often highly lucrative for those who make it work successfully.
The big studios still exist, they still make expensive, widely appealing movies, and are still profitable, but last year, YouTube made nearly four times as much money as Hollywood.
I suspect something similar will happen with software, and I have some idea of how:
The FBAR and the Content-ification of Software
The FBAR form, one of many complex tax forms for US citizens with foreign accounts
A few days ago, I filed my 2025 tax returns. My taxes are relatively straightforward, pretty much just W-2 income from Greptile. However, I grew up in India, and I still have a bank account there. I am a US citizen by birth and now live here in San Francisco. One would think a US citizen with a foreign bank account is a relatively common tax situation, but apparently not common enough to be covered by TurboTax, the leading consumer tax-prep software.
I am reminded of this every year as I fumble through a mess of obscure forms like the FBAR (Report of Foreign Bank and Financial Accounts) to accurately report my dividends, capital gains, and balances. Every year I search for an alternative, and every year I am disappointed.
TurboTax likely decided there just weren't enough people in situations like mine for it to be worth the engineering resources it would take to build support for the longtail of IRS forms. What happens if the productivity of an engineering resource goes up ten times? Suddenly, many more scenarios are economically viable to cover.
It may seem that this would allow today's software companies to build better software for everyone because even specific edge cases would be covered. I think this is unlikely. The new bottleneck for these teams will be leadership, agency, management overhead, and product + design resources. Even in a post-AI world, TurboTax is unlikely to expand its offerings drastically faster than they do today.
It follows that small groups of technical, entrepreneurial people can build TurboTax-like software products for niche scenarios. For the most niche situations, their revenue will likely cap out at a few million dollars a year, but because you would only need 3-4 AI-supported developers to maintain the app, that would be enough to make a nice small business.
You can apply this logic in broadly two situations.
- There are software products that are ill-fitted for but still the leader for specific use cases because the use cases can't themselves yield enough revenue to support a software team. Think horizontal software platforms like ServiceNow and Salesforce, which have to be squeezed into shape to approximately accomplish many tasks that are outside of their core thesis.
- There are entire use cases that are simply unserved by the software industry because they can't yield the revenue to support a software team. Education and consumer utility apps for niche communities seem like good candidates.
My only high confidence prediction for the future of software is that:
- There will be many more software companies
- They will each be smaller in headcount
- They will tackle niche problems that are partially served or not served by current software.
Think indie film studios that do sketch comedy on YouTube, or full-time Twitch streamers.
The "big studio" equivalents will continue to exist - big tech companies. They will probably also be able to make better software, at least on a technical level. Just as the amount, variety, and specificity of the content we consume has exploded in the last two decades, as will the amount, variety, and specificity of the software we use.
Footnotes:
[1] Of course, not everyone thinks this is likely. Many, including myself, view AI as a tool or assistant and not as a replacement for human software engineers. That said, I'm incentivized to have this view. My company makes an AI code review bot and our customers are software teams that want to be more productive. If AI replaces software engineers, we would have no customers. It's also important to my business that software engineers view us favorably, and telling people that the craft they have spent years perfecting can be trivially solved by matrix multiplication is a great way to make them dislike you.