When AI became mainstream, developers got excited about using AI to write code. This phase was birthed by the release of dedicated generative AI tools for software development. But with AI scaling to even newer heights, coding has entered another exciting new phase. This phase bears a promise to generate code with a welcome twist. The word on the block is that this wave might have something to do with Artificial general intelligence (AGI).
The Second Wave of AI Coding
AI coding tools are not new. GitHub’s Copilot, launched in 2022 and backed by OpenAI, is already used by millions. General-purpose AI tools like ChatGPT, Anthropic’s Claude, and Google DeepMind’s Gemini have also become essential aids for developers. Alphabet CEO, Sundar Pichai, even revealed that over 25% of Google’s new code is now AI generated.
But today’s tools are just the beginning. The second wave of AI coding is pioneered by startups like Zencoder, Cosine, Tessl, and Poolside. These companies aim to elevate coding assistants beyond basic autocomplete features. Their vision? Tools that can not only write code but also prototype, debug, and optimize it. This essentially automates significant parts of the software development process.
Why Coding Is the Perfect Playground for AI
Coding presents a unique opportunity for AI innovation. Unlike tasks requiring subjective judgment, programming operates on logical rules and defined outcomes. This structured environment allows AI to excel.
Jared Kaplan, chief scientist at Anthropic, highlights the excitement developers feel as AI tools become adept at debugging and error detection. The goal, however, is more ambitious: creating AI systems capable of mimicking the thought processes of human coders.
From Autocomplete to Intelligent Problem-Solving
Current generative AI models can produce syntactically correct code. They understand grammar, syntax, and the surface-level structures of various programming languages. Yet, producing functionally correct code- code that does what developers intend- is a bigger challenge.
Alistair Pullen, cofounder of Cosine, explains, “Large language models can write code that compiles, but they may not always write the program you wanted.” To bridge this gap, developers are training AI to replicate the decision-making process behind coding.
Building Smarter Models
Startups like Cosine and Poolside are addressing this challenge by collecting new types of data. Instead of training AI solely on finished code, they focus on capturing the process of software development. This includes:
- Recording developers’ workflows, such as why they open specific files or scroll through certain sections.
- Annotating code to highlight dependencies on other parts of a codebase.
- Creating synthetic datasets that simulate the steps coders take to produce functional software.
Cosine’s approach involves training models to recognize the “breadcrumb trail” left by human developers. Similarly, Poolside uses reinforcement learning from code execution (RLCE), allowing AI to learn through trial and error, akin to how DeepMind’s AlphaZero mastered games like Go.
Context Is King
Understanding a project’s context is vital for generating accurate and relevant code. Zencoder, for instance, focuses on “repo grokking”, analyzing entire code repositories to provide contextually accurate suggestions. This approach reduces AI hallucinations and improves output quality, making it a game-changer for developers handling large projects.
The Push Toward AGI
For many in the field, generative coding assistants represent a potential fast track to AGI. These tools could reach human-level capabilities in software development before any other economically valuable activity.
We read all the AI news and test the best tools so you don’t have to. Then we send 30,000+ profesionnals a weekly email showing how to leverage it all to: 📈 Increase their income 🚀 Get more done ⚡ Save time.
Eiso Kant, CEO of Poolside, believes that this milestone will mark a turning point: “A human coder tries and fails one failure at a time. Models can try things 100 times at once.”
However, not everyone agrees. Critics like Justin Gottschlich, CEO of Merly, argue that large language models lack the logical precision required for complex coding tasks. He compares their limitations to training a dog to code: an impossible task, no matter how much training is provided.
The Automation and Human Expertise Tightrope
As AI takes on more coding responsibilities, developers are transitioning into managerial roles. Instead of writing code from scratch, they review, refine, and guide A -generated outputs. While this shift increases efficiency, it also raises questions about the future of programming jobs.
Real-World Impact
Companies like Cosine and Poolside are already demonstrating the potential of next-generation coding tools:
| Company | Innovation |
| Zencoder | Context-driven code analysis for improved accuracy |
| Cosine | Synthetic datasets capturing coding workflows |
| Poolside | Custom-built models trained exclusively on coding processes |
Challenges Ahead
Despite their promise, AI coding tools face hurdles:
- Bias in Training Data: AI can perpetuate errors or biases present in its training datasets.
- Complexity of Human Logic: Capturing the nuance of human decision-making remains a significant challenge.
- Reliability Issues: Ensuring that AI-generated code functions correctly in real-world scenarios requires rigorous testing.
