What Software Developers Can Learn from Linguists: The Rise of the Software Post-Editor

This might get me canceled by both tech circles and linguistic ones.

Well, the cat is out of the bag. I just could not believe Andy Jassy said the quiet part out loud yesterday. He revealed the gains Amazon made using Generative AI in software development.

tl;dr

• 79% of the auto-generated code reviews were accepted without any additional changes.
• Estimated gains = $260M !!!

For those who worked in Machine Translation over the past decade, I’ve got news for you. We were in the middle of an experiment that will now get replicated in other fields. The experiment running as we speak is in software development. I find it somewhat ironic that the industry automating others is now being automated itself.

What fascinates me the most is the psychological and the economic aspect of this rapid change and I think we can place some bets on how this will play out.

The Translation Industry

The Secret

For people who are not in translation industry, you’d be surprised to know that the majority of professional translations today start with AI-generated content. I estimate 90% or more. What professional translators and languages services providers (LSPs) do is not really translate. They don’t start with a blank page. What they do is:

1. Correct what the machine got wrong, a process called post-editing.
2. Fine-tune the AI to be more predictable.
3. Review the end result, much like a double check.

Experienced linguists and professional translators absolutely hate this process. Instead of flexing their skill from beginning to end, they need to edit what a soulless machine produced. It’s almost correcting sub-par work over and over and over, and being measured on how many words you post-edit. Interestingly, younger linguists have no qualms about it; they simply accept that their job is to correct and review what the machine produces.

In practice, this means the work often involves scanning the translated text diagonally and occasionally overlooking parts due to time constraints. So, errors could be overlooked.

The Money

The arbitration in this industry is around:

A. who captures the efficiency gains from what the AI does, and

B. who takes the responsibility when things go wrong?

By offering B, LSPs capture A. It's not the translator who reaps the efficiency gains but the one who bears the full responsibility for the work.

The Software Development Industry

The (new) Secret

A lot of the software development work is boilerplate set-ups and repetitive tasks. If you buy the service by the hour, you literally might be paying someone to copy paste things over and over, not really invent some novel streaming algorithm.

Because of these repetitive patterns, and because the “language” of software is more structured than English, LLMs are able to generate almost perfect code if properly prompted. That’s what Andy Jassy was talking about in his post. And that is the secret. Apparently 79% of the results were accepted and merged into the codebase.

So, developers can automate, or “outsource” to an LLM large parts of their tasks (maybe the ones they hate the most - like writing tests).

The Money

Software development as a service is 2-10x more costly than translation. The stakes of capturing gains are much higher.

Here’s what I think will happen, by drawing a parallel with Machine Translation and the language industry:

1. Software developers will push back against LLMs and highlight their mistakes, just as linguists have done. Maybe they will use the trump card of “security” and leaking IP, while storing the code on a cloud service like GitHub.
2. Behind the scenes, they will use LLMs to generate a good part of the code and gain time. This will replace endless browsing of StackOverflow.com
3. As such, many mediocre developers will get lazier and lazier and rely a lot on what the LLMs produce. This is a risk, but it’s a risk in every field that will be aided by AI.
4. CEOs and CPOs, reading Andy Jassy’s post will demand 2x productivity of their software teams, and likely reduce the headcount to send a message. I know a really strong CEO in the language industry that is demanding higher productivity of his product team.
5. Smart 10x developers will embrace LLMs and get even more of the tedious work outsourced so that they can focus on high value tasks like architecture, optimization, and ultimately being creative.
6. This is bad news for junior / or poor software developers demanding high salaries with 0 experience.
7. Software development service providers (a.k.a. outsourcers) will either have to reduce their rates and/or ship things faster. They can retain their margin by running a leaner operation. The smart ones will create more value that goes beyond shipping code, and maybe offer to run the service. In other words, change from software development as a service to software as a service. But all in all, I don’t think they will be able to retain the value as the LSPs did, given the change is faster and the stakes are higher here.
8. Startups on the other hand are the biggest beneficiaries of this trend. Less capital is required to build a scalable product. Or the same capital can get you to market faster. Small teams of 10x developers can run circles against established businesses.

This is my case. At Video Highlight, I am lucky to work with 10x developers and witness what high performance means. I am not one of them, but I’m proud to say I am a “software post-editor”. I ship more code than I could have ever imagined. I get tutored by LLM-generated code every day. I try to improve daily, but most importantly, I do this to serve the Video Highlighters faster and delight them with new features. I could not do it without the help of LLMs. Oh, and we also use Machine Translation to communicate with our customers in Latin America and Asia :)

Now, back to post-editing some Python

Let me know your thoughts.