Self-driving cars still need steering wheels because complete automation is very difficult to achieve. Experts thought we would have fully autonomous cars by now, but there are still many challenges to overcome.
Software engineering is even harder to automate than driving. As we create tools that simplify coding, the demand for software will only continue to grow, rather than decrease.
Small tools that help human engineers will likely be more valuable and widely adopted than fully autonomous coding systems. They make the coding process easier without completely changing how we work.
Design problems require more thought and effort compared to straightforward problems. It's about finding the best solution among many options, which is not always easy.
Good designers think ahead about how their work will be used in the future. They prepare solutions that can adapt to changes instead of just solving today's issues.
Scaling compute at inference time helps create better designs. It’s like having someone who combines experience and planning to come up with smarter solutions.
LLM coding can be easy at first, allowing users to operate without deep understanding, similar to driving on autopilot. However, this can lead to mistakes and poor coding practices over time.
Understanding complex systems is hard, and it's often not all written down. People rely on context and shared knowledge, which LLMs can miss out on, making it challenging for them to fully grasp what’s going on.
If you don't understand your project's requirements or the underlying system well, you'll run into problems and make mistakes. Using LLMs requires a critical eye to avoid getting lost in error accumulation.