We’ve come a long way since GitHub Copilot first showed us what AI-assisted coding could look like. These days, there’s a whole ecosystem of LLM coding tools out there, each with their own strengths and approaches. In this blog, I’ll give you a quick overview to help you figure out which one might work best for your workflow.

Level 1: Interactive Code Assistance

The entry point for most researchers is direct interaction with models like ChatGPT or Claude. While this approach excels at tasks like debugging PyTorch models or optimizing scikit-learn pipelines, it has notable limitations. The lack of context about your broader codebase means you’ll need to carefully frame questions and provide relevant code snippets. This becomes particularly challenging when working with domain-specific libraries like RDKit or Bio-Transformers.

Best for: Quick debugging, code optimisation, and learning new APIs

Limitations: No context awareness, manual copy-paste workflow, session memory constraints

Level 2: Integrated Development Environments

The next tier integrates LLMs directly into your development workflow. GitHub Copilot was first here, but tools like Cursor offer greater flexibility in giving you control over what LLM to use, as well as more advanced ‘agentic’ modes. Cursor’s ability to add docs from any python library to the context of LLM calls is particularly useful, giving more relevant edits.

Level 3: Autonomous Development Platforms

The frontier of LLM coding tools promises complete automation, handing over all coding to LLM agents, letting you stay completely in natural language. Tools like V0 are great for rapid prototyping of user interfaces – potentially useful for deploying model web apps or visualisation dashboards. However, they’re currently not great for scientific / ML coding. If you need to quickly build a front-end for your model though, I’d definitely recommend giving them a try, I was able to build a functional and very clean web app within a couple hours despite having minimal front-end experience. Other options, including loveable and windsurf, offer full-stack agentic coding, but I haven’t played around with them yet.