LLMs and Programming: A Practitioner’s Perspective on Code Generation and Productivity

In the first days of 2024, it's worth reflecting on how large language models (LLMs) have transformed programming. This isn't a retrospective on the AI boom of 2023; rather, it's a personal account from an experienced programmer who has integrated LLMs into their workflow to accelerate code writing and reduce mental fatigue.

The Shift in Programming Practices

Since the debut of ChatGPT and the subsequent availability of local LLMs, I've found these tools invaluable. My primary goal is to speed up my coding process, but there's more to it than that. LLMs help me avoid wasting time on trivial tasks, such as searching for obscure documentation or learning overly complex APIs. These are often necessary evils that can be sidestepped with the right AI assistance.

• Documentation and API Learning: Searching through Google, which has become a minefield of spam, is no longer a necessity.
• Immediate Usability: Writing code that I might discard after a few hours is less burdensome with LLMs.

Balancing Independence and Assistance

I'm not a novice programmer; I can write code independently and often do. However, over time, I've increasingly relied on LLMs for high-level coding tasks, particularly in Python. My experience has taught me when to use these tools and when they might slow me down.

• High-Level vs. Low-Level: I find LLMs most useful for writing high-level code in Python, less so for lower-level languages like C.
• Strategic Use: Knowing when to leverage LLMs is crucial. They are not a crutch but a tool that complements my skills.

The Parrot Phenomenon

One of the most concerning aspects of this AI revolution is the limited ability of experts to accept their own limitations. Neural networks, which humans invented and optimized, have become increasingly opaque. Despite our advancements in hardware and algorithms, there's still a lot we don't understand about why certain architectures work better than others.

• Transparency Issues: The opacity of neural networks remains a significant challenge.
• Expert Reactions: Many AI experts initially dismissed LLMs as advanced Markov chains, capable only of regurgitating variations of their training data. This "parrot" notion has been largely debunked by empirical evidence.

Personal Insights

My personal experience with LLMs has revealed that they are most beneficial to those who already have a strong foundation in programming. While LLMs can assist anyone, they are particularly advantageous for those who have the will, ability, and discipline to use them effectively.

• Benefit Distribution: Initially, LLMs may only benefit those who already have an advantage.
• Learning Curve: Understanding when and how to use LLMs is a skill that develops over time.

Conclusion

LLMs are here to stay, and they are changing the way we program. They offer significant benefits in terms of speed and efficiency, but their true value lies in how they complement our existing skills. As with any tool, it's up to us to use them wisely.