Optimization Memes

So we're optimizing LLM token consumption now by... using external stylesheets? The same practice we've been preaching since 2005? Incredible. The AI era has brought us full circle to basic web development best practices, except now the justification is "save tokens" instead of "save bandwidth." The beauty here is watching people discover that separating concerns actually has benefits beyond making your code maintainable. Who knew that not dumping 20 lines of CSS into every prompt would reduce token usage? Next you'll tell me that minifying code and using compression also helps. The real galaxy brain move is training the LLM to reference external CSS so it "never outputs CSS again." Because nothing says efficiency like teaching an AI to avoid generating something it's perfectly capable of generating. It's like hiring a chef and then telling them to never cook vegetables because you bought them pre-cut.

So someone submitted an AI-generated assembly patch to dav1d (a video decoder), and it was slower than C. Let that sink in. Assembly—the language you write when you want to squeeze every last CPU cycle out of your code—got outperformed by C because an AI wrote it. The entire point of hand-writing assembly is to achieve performance that compilers can't match. You're basically telling the compiler "step aside, I'll optimize this myself." But AI-generated assembly? That's like hiring a robot chef to make instant ramen and somehow ending up with something worse than the microwave version. Turns out AI doesn't understand cache lines, instruction pipelining, or the dark arts of SIMD optimization. It just vomits out syntactically correct assembly that runs like it's stuck in molasses. Modern C compilers have decades of optimization wizardry baked in—AI has... vibes.

Oh, the sweet irony of trying to optimize AI token usage by talking like a caveman, only to realize you're actually BLEEDING tokens by explaining your caveman strategy! 💀 Someone discovered that instead of politely asking the AI to do a web search (~180 tokens), they could just grunt "Me tool first. Me result first. Me stop" and save 135 tokens. Genius, right? WRONG. Because now they have to spend tokens explaining their brilliant caveman protocol, which costs MORE than just talking normally in the first place. The breakdown is absolutely brutal: teaching the AI what "tool work" means costs 2 tokens, explaining the normal behavior costs 8 tokens, and each caveman grunt swap saves a measly 6 tokens. So after 8-10 swaps, you MIGHT break even with 50-100 tokens saved total. But realistically? You're burning 50-75% MORE tokens just to set up your caveman efficiency system. It's like spending $100 on organizational tools to save $20 on groceries. The math ain't mathing, but hey, at least you feel productive! 📉