HackerNews中文版

我知道这听起来可能有些消极，也有点笼统，但……说真的，LLM（大型语言模型）到底能帮你做些什么，除了“节省时间”之外——也就是说，真正的新颖工作。我个人觉得，虽然我无法证明这一点，也不想坦白，但自去年十月以来，LLM似乎被故意削弱了。4.5-4.6版本的Opus…… 每家公司都在有意或无意地降低其模型的性能（可能更糟），而他们无法承认这一点，如果你试图提及，他们甚至会让你感到困惑……但根据我自己的实验，LLM实际上比没有LLM还要糟糕…… 我没有看到任何人正在构建他们之前就能做到的东西……而且我看到的东西不需要被创造出来的频率越来越高……以至于我无法记起在这里看到的最后一个好主意（而之前每个月这里都有很酷的想法被发布）…… 我甚至无法得到中等水平的输出……这简直就是垃圾——简直就是一个提示到垃圾桶的管道…… Claude是最糟糕的。毫无疑问。自4.5以来，他们在每一个模型上都退步了……以至于4.5发布时绝对比今天市场上的任何东西都要好……价格却只是一小部分……而且速度快了10到100倍…… 那么……我需要吸什么，什么药能让我看到LLM的效果？我需要放松一下……因为在这个时候……使用它们真是令人感到侮辱…… Claude会竭尽所能地避免做被要求的事情……而且为了一个看起来像狗屎的单一HTML网站，它会花费50分钟的时间思考……而我用现成的样式手动编码大约90分钟就能搞定…… 所以……是的……什么药？多少？多频繁？副作用是什么？

查看原文

I know this is going to sound negative, and probably a bit generalist but....seriously, what exactly is the LLM helping you with that is not just purely "time saving" - ie. Actual, novel work.I personally feel like while I cannot prove it, nor want to frankly, that LLMs have been neutered intentionally since around October of last year. 4.5-4.6 Opus....Every single company has degraded their models eirher intentionally or accidentally(probably worse) and they cannot admit it and will gaslight you if you even try to acknowledge it....but for my own experiments, LLMs are actually worse than no-LLM....I dont see anyone building anything they couldnt have done before hand....and I am seeing stuff that didnt need to be made...much more frequently....to the degree that I cannot remember the last good idea I have seen even on here (whereas there were cool ideas every month posted here)...I also cannot get even mediocre output anymore...Its literally just trash - Literally a prompt to dumpster pipeline....Clause is the worst. Hands down. They have regressed on every single model since 4.5....to rhe degree that 4.5 when it was released was hands down better than anythingg on the market today...for a fraction of the price....and was like 10-100x faster...So....what do I need to smoke, what drug makes me see the LLMs work? I need to get faded...because at this point....Its insulting to use them....Claude will literally bend over backwards to do everything in its power to avoid doing what was asked....and will take 50minutes of thinking....for a single html website....that looks like fucking dogwater and which I could have coded up by hand using off the shelf styling in probably 90 minutes....so...yeah...what drugs? How much? How often? Side effects?

大型语言模型（LLMs）。你事先吸了什么？