Some additional thoughts on GPT-4.5: it provides BFK-9k experience - eats e̶n̶e̶r̶g̶y̶ ̶c̶e̶l̶l̶s̶ budget ($2 per call!) like there is no tomorrow, but removes bugs with a blast.
In my experience, the gap between Claude 3.7 and GPT-4.5 is substantial. Claude 3.7 behaves like an overzealous intern on stimulants. It delivers results but often includes unwanted code changes, resulting in spaghetti code with deeply nested conditionals and redundant null checks. Although initial results might appear functional, the resulting technical debt makes subsequent modifications increasingly difficult, often leaving the codebase in disarray. GPT-4.5 behaves more like a mid-level developer, thoughtfully applying good programming patterns.
Unfortunately, the cost difference is significant. For practical purposes, I typically combine models. GPT-4.5 is generally reserved for planning, complex bug fixes, and code refinement or refactoring.
In my experience, GPT-4.5 consistently outperforms thinking models like o1. Occasionally, I'll use o3-mini or DeepSeek R1, but GPT-4.5 tends to be noticeably superior (at least, on average). Of course, effectiveness depends heavily on prompts and specific problems. GPT-4.5 often possesses direct knowledge about particular libraries (even without web searching), whereas o3-mini frequently struggles without additional context.
Sometimes I could solve in 15 mins, a bug I had been chasing for days. In other cases, it is simpler to write codes by hand - as AI either does not solve a problem (even a simple one), or does, but at a cost of tech debt - or it takes longer than doing things manually.
AI is just one more tool in our arsenal. It is up to us to decide when to use them. Just because we have a hammer does not mean we need to use it for screws.
> Wouldn’t it be easier instead of juggling with [something] and their quirks to just write the code the old way?
This phrase, when taken religiously, would keep us writing purely in assembly - as there is always "why this new language", "why this framework", "why LLMs".
Most tools we have been using are not probabilistic. I use syntax highlighting, autocompletion, refactoring because they work close to 100% of the time. So my question really is if using a tool that works some % of the time is worth it.
In my experience, the gap between Claude 3.7 and GPT-4.5 is substantial. Claude 3.7 behaves like an overzealous intern on stimulants. It delivers results but often includes unwanted code changes, resulting in spaghetti code with deeply nested conditionals and redundant null checks. Although initial results might appear functional, the resulting technical debt makes subsequent modifications increasingly difficult, often leaving the codebase in disarray. GPT-4.5 behaves more like a mid-level developer, thoughtfully applying good programming patterns.
Unfortunately, the cost difference is significant. For practical purposes, I typically combine models. GPT-4.5 is generally reserved for planning, complex bug fixes, and code refinement or refactoring.
In my experience, GPT-4.5 consistently outperforms thinking models like o1. Occasionally, I'll use o3-mini or DeepSeek R1, but GPT-4.5 tends to be noticeably superior (at least, on average). Of course, effectiveness depends heavily on prompts and specific problems. GPT-4.5 often possesses direct knowledge about particular libraries (even without web searching), whereas o3-mini frequently struggles without additional context.