Raw LLM Responses

Developer with over 3 decades of experience in too many languages to list here without it sounding like a brag. If you mean that you cannot just let it loose without any senior supervision: yeah, I agree. However, it is really good at coding if you use the correct model, keep the scope reasonable, and do code reviews. You might as well claim that IDEs are a "dead end" when it comes to doing intelligent work. LLMs are a tool, that in the right hands can multiply a dev's effectiveness by several times. All this is without any assumptions about "AGI" or just setting agent loose on a codebase. I would also be very \*very\* careful about the "cannot reason" argument. It's popular, has some grounding, but can quickly overreach. There are some papers that do say this, but be careful of just cherry-picking research that confirms what you want to be true (and let's face it: most of us want it to be true so that our careers are safe; understandable, but still potentially dangerous) The real answer is that we do not really understand LLMs very well, we do not yet have a good grasp on what they do (as a total system, because clearly we know the algorithm), and it probably will not matter that much anyway, as LLMs are being combined with other AI systems to deal with any perceived weaknesses anyway. Any developer not effectively using LLMs in their daily work is going to quickly fall behind. And I absolutely get that you are not saying they should not be used (I agree that LLMs are indeed strongest in the areas you pointed out). But they can go beyond that, already, today. While executives might dream of eventually getting rid of their entire dev team, that is not the reality today. But equally true is the dismissal of the ability of LLMs to code is also not the reality today (again, you are not nearly at the level of dismissiveness that I have seen on this subreddit). Keep the scope reasonable, and they do really well.