The world of coding is undergoing a transformation and it is due to the second wave of AI-powered tools. AI is not just a helpful assistant for developers today, but it is gradually becoming an indispensable partner in writing, debugging as well as understanding code. The change is highly promising and said to redefine the role of software engineers. It could even pave the way toward artificial general intelligence (AGI).
The Copilot project of GitHub and OpenAI was launched in 2022 and since then millions of developers have relied on it for smarter autocomplete suggestions. General-purpose chatbots like OpenAI’s ChatGPT and Anthropic’s Claude are simultaneously also being used in simplifying coding tasks. Alphabet CEO Sundar Pichai claims that more than 25% of new code at Google is first generated by AI and thereafter refined by engineers. This means that AI is not just an add-on now, but it is already an important part of the workflow for many.
Startups like Zencoder, Cosine, Tessl and Poolside are also creating tools which are beyond just the basics. These next-gen assistants can prototype, test and debug entire programs. All these let developers to focus more on managing and reviewing code rather than creating it from scratch.
Poolside is a good example and it has already reached a $3 billion valuation without launching a product. Founder Eiso Kant believes that AI coding tools could be a major step toward AGI.
However, building the tools is not just about training on massive datasets of existing code, but companies like Cosine and Poolside argue that understanding the way code is written is as important as the final product. Developers don’t just type lines of code, but they think through problems, reference documentation and debug. Capturing the process is what makes the tools more than glorified autocompletes.
Startups are employing unique methods as Zencoder has hired search engine experts to improve the context that their tools can analyze. Similarly, Cosine has been asking developers to document their coding process in detail like what files they opened, why they scrolled to certain sections and so on. The tools aim to replicate a logical journey of developers and not just the end result by mapping the steps.
Everyone is not convinced yet and critics like Merly CEO Justin Gottschlich argue that large language models (LLMs) are not suited for the logical precision coding demands. Merly trains its system on the underlying logic of programming and not with raw code.
