Present & Future of AI Coding Assistants

The current landscape of AI coding assistants can be put into 3 major categories - LLMs that can generate code (OpenAI, Anthropic, Poolside, Magic.dev), Copilots that warp these LLMs(Github Co-pilot, Cursor, Codeium) and Agents that can use tools, memory & reasoning to autonomously perform coding tasks(Replit, Cognition Labs). 

The strengths of AI coding assistants lie in their ability to enhance productivity and help developers of all skill levels. Beginners benefit from instant help in understanding programming concepts, while experienced coders can offload mundane tasks and focus on more creative aspects of development. By offering quick suggestions and explanations, these tools contribute to making coding more accessible and efficient. The LLMs behind these assistants leverage vast datasets of publicly available code, providing developers with time-saving suggestions, thereby accelerating the development process. And that's where we are currently - AI automates learning, research & summarization part of coding. 

At their core,  LLMs are designed to get you answers fast. To that extent, the answer is heavily dependent  on the question. A lot of it also depends on the developer to frame the prompt accurately. And mis-framing the prompt can get you entirely different results. Recovering from an LLM deep dive in a wrong direction can leave a trail of redundant, poorly formed, and in-efficient code. 

Coding process is deliberate, and slow. A developer on an average writes 150-200 lines of code per day. There is a lot of thought behind  every line, because there is time to think, reason and recover. LLMs get you the answer very quickly. And this works for fairly basic, beginner level coding tasks. When you onboard a junior developer, you ask them to work on less impactful, new code that doesn't require  deep understanding of the code base. Most of the new code will either be net-additions that are heavily re-worked, or thrown away. That's exactly where AI coding is today - at the level of a junior developer. A study done by GitClear of over 150 million lines of code indicates that most of the AI written code are net additions, and proportion of lines discarded within two weeks of being written doubled since last year. 

One major shortfall of today’s AI coding tools is their lack of deep contextual understanding. They excel at generating code snippets, but they often struggle to fully understand the broader purpose or architecture of a project, engineering best practices and vision for an end state. A developer might want a particular function written with efficiency in mind, or with future extensibility as a key feature. Current AI tools can't reliably take these nuanced intents into account. They generate code that works, but it often needs significant tweaking to align with the broader goals of a project.

Additionally, security remains a concern. AI coding assistants can inadvertently suggest insecure code, exposing projects to vulnerabilities. Their reliance on public code repositories raises concerns about the reuse of potentially flawed or outdated practices. Moreover, their inability to fully comprehend high-level architectural decisions or long-term project goals means they lack foresight in terms of performance, scalability and maintainability.

To truly unlock their potential, AI coding tools need to move beyond surface-level code generation. They need to understand context, learn from developer habits, integrate deeply with existing tools, and consider the human and real-world aspects of software development. Improving these facets could lead to tools that not only generate code but become trusted, indispensable partners for developers. Foundational LLMs are trying to solve this problem by expanding context windows to ingest large code bases or build a coding specific LLM. Gemini has context window of 1.5Million tokens & Magic.dev is targeting 5M token context window for coding. Poolside recently raised $500M to buy 10K NVIDIA GPUS to train coding models. Whether this up levels these models from junior to principal level engineer with context remains to be seen. One thing is for certain - coding LLMs need a lot of compute for training and inference, and hence, massive funding rounds.

An ideal solution  would be an LLM that sits in your development  environment, ingests all of your codebase at once, in a privacy compliant way, understands the end-end architecture, engineering goals and evolves with your business. It's hard to build a general purpose coding LLM that can generate perfect code, every time, for update use cases.

GibsonAI can write perfect code by narrowing the problem space to deterministic, repetitive tasks on the backend. GibsonAI also maintains the style by writing  code in industry standards. There you have it - a Principal level engineer that understands  your context and writes 70% of your backend code perfectly. Gibson recently raised a seed round of $3.5 million from Oceans Ventures & 1P ventures, with participation  from F7 ventures, Riverpark Ventures and Struck Capital. According to the CEO, Harish Mukhami, this new capital will be used to add world class engineers to the team. GibsonAI is specifically  focused on delivering production level code accurately, and consistently with contextual understanding of the codebase. Here is more from Harish Mukhami on the future of AI coding assistants.