|
Edit: The comment I am replying to was rewritten completely, and originally asserted that the quality of LLMs was now undeniable. "Undeniably"? I will deny that they are good. I try to use LLMs on a near-daily basis and find them unbearably frustrating to use. They cannot even adequately complete instructions like "following the pattern of A, B, and C in the existing code, create X Y and Z functions but with one change" reliably. This is a given; the work I do is outside the training dataset in any meaningful sense, so their next-token-prediction is statistically going to lean away from predicting whatever I'm doing, even if RL training to "follow instructions" is marginally effective. The conclusion I've come to is that the 10x hypebots fall into two categories. The first is hobbyists who could barely code at all, and now they are 10x productive at producing very bad software that is not worth sharing with the world. The other category is people who use LLMs to launder code from the training dataset to wash it free of its licenses. If your use case is reproducing code it has already been trained on, it can do that quickly. These claims of "holding it wrong", one of which I already see in the replies, are fundamentally preposterous. This is the revolution that is democraticising software engineering for anyone who can write natural language, yet competent software engineers are using it wrong? No, the reality is that it simply doesn't have that level of utility. If it did, we would be seeing an influx of excellent software worthy of widespread usage that would replace much of the existing flawed software in the world, if not pushing new boundaries altogether. Instead we get flooded with ShowHNs fit for the pig trough. That's not to say LLMs have zero utility. They can obviously generate a proof-of-concept quickly, and if the task is trivial enough, save a couple of minutes writing a throwaway script that you actually use day-to-day. I find them to be somewhat useful for retrieving information from documentation, although some of this gain is offset by the time wasted from hallucinated APIs. But I would estimate the productivity gains at 5%, maybe. That gain is hardly worth the accelerating AI psychosis gripping society and flooding the internet with garbage that drowns out the worthwhile content. Addendum: Now that your post has been rewritten to assert that no, LLMs aren't there yet, but surely in the next 6 months, this time for sure it'll be AGI... welcome to the bubble. I've been told that AGI is coming in a couple of months every month for the past two years. We are no closer to it than we were two years ago. The improvements have been modest and there are clearly diminishing returns on investing in exponential scaling, not to mention that more scaling can never solve the fundamental architectural flaws of LLMs. |