Is Time Ripe To Throw Your Engineers Under The Trolley?

On the 30th of the holy month of October, Sundar Pichai climbed on top of the ramparts keeping Silicon Valley safe from Wall Street, and pronounced “Over 25% of Google’s code is now written by AI”. It was natural for executives and managers worldwide, who have never built even a bottle opener by themselves, to have sleazy dreams about AI serfs replacing software engineers. It was also natural for small-time script kiddy-turned-thought leaders to believe they could team with AI and replace software engineering with their thought leadership.

So should you signal your “Chief Prompt Offices” to run over your engineering team with the AI trolley? If you are able-minded you should consider these key aspects:

First off, Google is once again late to the party. Since late 2023, tools like Copilot for code completion and ChatGPT for low-key troubleshooting have been assets enough that they should have been adopted within any organization for pair programming. More recently, tools like Replit Agents have shown bigger promise with their spec-to-app capabilities. At Google’s size, it shouldn’t have taken a year to quantify the gains, so the timing is more of a marketing strategy.

There is no contention that these tools are game-changing. With just a bit of nudging they save you days of drudgery of writing repetitive code, designing test cases, filling in test data, and building schemas. The spec-to-app feature, although limited, can create small but functional apps from just a description or a sketch. That is a lot less of mundane work but does it translate into substantial time savings? Not necessarily.

AI cannot be trusted to be correct though so, depending on the levels of diligence and integrity of engineering teams, additional time is spent, reviewing the code, beefing up the test cases, and troubleshooting small mistakes. Sometimes the AI inserts bugs or deprecated code and weeding them out can take more time than writing code from scratch.

The same goes for AI software components, like text extractors for unstructured documents, such as transcripts and invoices. AI can ingest documents and spit out structured information quite accurately at djinnatic efficiency. However, “quite accurate” is not an acceptable threshold of data quality, unless there are humans in the loop. So while a 2 line prompt can extract 100 disparate fields from a document into JSON, building the guardrails to keep tabs on the data quality is an enormous effort.

AI-armed engineers are also more likely to let in bugs, adding development iterations and wiping off any actual savings. Either the infra for building, testing, and deploying changes has to be ramped up to cope with the rapid iteration, or it's a downhill road into truckloads of technical debt.

Then there is a false presumption that AI tools will keep their pace of improvement. Today’s tools have already been trained on large chunks of human knowledge, from books, forums, blogs, and code repositories spanning decades and even centuries. The sources are already exhausted. As AI generates more new content, the sources will be saturated and then overwhelmed by AI-generated content, making any further learning from them useless or deteriorating.

Finally, any process that can be fully automated with five lines of prompts, will sooner or later be usurped by your downstream users or upstream AI companies. If that is all you do, pack up your bags otherwise adopt the AI tools, cut your hard-working engineers some slack, or have them work on one of those ideas from the idea pipeline every organization has.