[ideonexus]

You can become more specialised; but that will only take you so far. Coding will soon become no more than a means to an end. Which it always was. The only way out for developers is through expanding their vision beyond that of the tools made available to them. Why does something need to be coded? What does the end customer expect? Did they vocalise their need properly? Contextual knowledge can hardly be reproduced and is the best way to beat a ChatGPT-like tool.

This is the content I clicked for and I feel like it's pretty vague and poorly-articulated, but there's a seed of truth here. Coding is a highly-technical profession and Chat-GPT does seem to trivialize it somewhat. With Chat-GPT automating out some of the developers, you will still need some developers around to understand and review the AI's output. The developers who remain will be the ones who are high-functioning: the ones who can communicate with both technical and non-technical staff, understand the business logic, and fill-in the blanks of what the customer needs.

In this future, Chat-GPT replaces some technical staff positions. So how does that work? In my experience with the tool, getting it to write anything multidimensional--like a board game--requires hours of back-and-forth articulating and re-articulating the interface, the business rules, and how those things interact. In this sense, it's like those developers who throw up their hands and complain about the requirements documentation while other developers stay in constant contact with the stakeholders during the development process to deliver the business needs. I don't think the later need to worry about job security.

[ilaksh]

Insightful comment, but you are assuming that the capabilities of this type of technology will remain static. Maybe that was also a premise of the article.

But we know that these systems will not stop improving. Just in recent weeks there are scientific papers documenting LLMs with improved capabilities as well as clear reports of more capable models in the business pipeline. And papers describing major upgrades to the models such a as adding a visual modality to the data.

But beyond the last few weeks, there is an exponential trend in the capabilities of the hardware and systems. And a clear track record of creating new paradigms when there is a roadblock.

In the next several years we will see multimodal large models that have much better world understanding grounded in spatial/visual information. Cognitive architectures that can continuously work on a problem with much better context, connected to external systems to automatically debug or interact with user interfaces in a feedback loop. Entirely new compute platforms like memristors or similar that increase efficiency and model size by several orders of magnitude. All of this is well under way.