Copilot, a coding assistant developed by Microsoft-backed GitHub using OpenAI's large language models, was launched in 2022 and is now utilized by millions of developers globally. In addition, many users rely on general-purpose AI chatbots like Claude, ChatGPT, and Gemini for various tasks. Alphabet CEO Sundar Pichai mentioned in an earnings call that AI now generates over a quarter of new code at Google, allowing engineers to be more efficient. Other tech companies are expected to follow suit, as new startups like Zencoder, Merly, Cosine, Tessl, and Poolside are emerging in the AI coding space, vying for market share. Analysts note that developers seem open to paying for copilot tools. These new startups aim to enhance generative coding assistants beyond just auto-completion. They can now prototype, test, and debug code, potentially allowing developers to focus more on oversight rather than hands-on coding. Many believe these advancements could lead to artificial general intelligence (AGI) through coding. Correctness in software engineering encompasses two aspects: syntactical correctness, ensuring code runs without errors, and functional correctness, which ensures code performs the intended task. The next generation of coding assistants aims to improve not just code generation but also the reasoning behind creating that code. Training models effectively requires a detailed understanding of the coding process, not just code samples. Zencoder focuses on contextual relevance when analyzing codebases, while Cosine collects data on coders' thought processes, mapping the typical steps taken to produce code.

Poolside employs a method called reinforcement learning from code execution (RLCE) to optimize code generation. Cosine and Poolside draw inspiration from DeepMind's AlphaZero, utilizing trial and error in coding to refine their models. While Cosine adapts existing models, Poolside is building a new, specialized model. Cosine claims its model outperforms others in coding benchmarks, while Poolside is still in development but shows promising initial results. Contrarily, Merly’s approach dismisses large language models, focusing instead on the logical underpinnings of programming by using an intermediate representation of code that captures its underlying structure. This method seeks to enhance logical reasoning in code generation. The advent of these AI tools may shift software development roles, allowing teams to produce software more efficiently. Coding assistants like Cosine's Genie can handle bug fixes and other tasks, letting engineers oversee multiple processes simultaneously, thereby changing traditional ways of working. As AI coding technology progresses, a new hierarchy of software engineers may emerge, with elite roles tasked with overseeing AI-generated code, while smaller teams manage overall tasks. This transformation resembles the changes ATMs brought to banking. To some, like Gottschlich of Merly, creating advanced software systems necessitates AI that can independently code. He critiques the idea of humans assisting AI, arguing the reverse should be true. Overall, the excitement surrounding generative coding assistants is accompanied by aspirations for more significant developments in AI, as companies envision their tools as steps toward achieving AGI. As the landscape of software engineering evolves, these tools are reshaping the roles and workflows of software developers, signaling an exciting yet uncertain future in the field.