Brilliant framing of multi-LLM workflows as Byzantine fault tolerance—that's the mental model I've been missing for why ensemble AI works so well. The Mechanical Turk comparision is spot-on too. When I first started using Claude to audit ChatGPT's code I thought it was overkill, but after catching a subtle race condition that both GPT-4 and Gemini missed independently, I'm convinced this is the only way to ship reliably. The productivity addiction piece hits different tho. I've definitely had that 4am "just one more refactor" feeling you described, and the inability to just exist without optimzing everything becomes exhausting. But honestly, dunno if there's a way back once you've experienced 30x leverage.
Totally, man. I think of it like artillery. After watching too many World War II documentaries on YouTube, I think of it like artillery. You don't expect it to be precise in the same way that a rifle or a sniper shot should be. You just have to blast it enough, and eventually you'll get everything you care about.
It's a number and stats game. Okay, you're not gonna get 100% with current systems, but maybe the chances the AI nails it initially are like 70 or 80%. With more LLM calls, you can get that number up to 95%. Not perfect, but the bar isn't perfection, the bar is paying slow humans who make the same mistakes.
And yeah, regarding the 30x leverage, it is very addicting. It's an incredible feeling of power. But much as with social media, learning to unplug is a new skill that you need to develop.
Brilliant framing of multi-LLM workflows as Byzantine fault tolerance—that's the mental model I've been missing for why ensemble AI works so well. The Mechanical Turk comparision is spot-on too. When I first started using Claude to audit ChatGPT's code I thought it was overkill, but after catching a subtle race condition that both GPT-4 and Gemini missed independently, I'm convinced this is the only way to ship reliably. The productivity addiction piece hits different tho. I've definitely had that 4am "just one more refactor" feeling you described, and the inability to just exist without optimzing everything becomes exhausting. But honestly, dunno if there's a way back once you've experienced 30x leverage.
Totally, man. I think of it like artillery. After watching too many World War II documentaries on YouTube, I think of it like artillery. You don't expect it to be precise in the same way that a rifle or a sniper shot should be. You just have to blast it enough, and eventually you'll get everything you care about.
It's a number and stats game. Okay, you're not gonna get 100% with current systems, but maybe the chances the AI nails it initially are like 70 or 80%. With more LLM calls, you can get that number up to 95%. Not perfect, but the bar isn't perfection, the bar is paying slow humans who make the same mistakes.
And yeah, regarding the 30x leverage, it is very addicting. It's an incredible feeling of power. But much as with social media, learning to unplug is a new skill that you need to develop.