Comment by saadn92
1 day ago
good question. maintenance is cheap with tools like astra/opus. also, these aren't production level applications, so I don't need to worry about it being "perfect" (though there's no such thing)
1 day ago
good question. maintenance is cheap with tools like astra/opus. also, these aren't production level applications, so I don't need to worry about it being "perfect" (though there's no such thing)
Ah yes, the ye olde "it doesn't need to be perfect" response to serious, legitimate questions.
It's always about cheaper, better, faster. And when the thing being built extends beyond a single person, and still when working alone, conflict shows up in balancing the risks between these. The Strawman of Perfection is just that.
"It doesn't need to be perfect" is incredibly dismissive of the reality of the situation. Ignoring reality seems like a poor problem solving strategy.
No one is ignoring reality. The reality is that my internal tools don't have users besides myself and as long as they do what I want them to do, I'm perfectly happy with them.
What happens when these companies decide to start increasing the cost/token, exceeding the cost of hiring an engineer or paying for an external service in the first place?
Or they optimize the model for just one more hit, er, prompt. All these AI companies are backed by the same companies optimizing revenue via addiction.
So, same end result (cost exceeds hiring a new engineer), but achieved differently.
The counter, I suppose, is the managers have to deal less with engineers telling them their ideas aren't good or well thought out.
Cost per intelligence-adjusted token has been in free-fall since GPT-3 landed five years ago [0]. I see no compelling reason to think we'll see that trend suddenly stop, much less reverse.
0. "The cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13× per year." https://epoch.ai/publications/the-plunging-price-of-thought
Yet OpenAI is losing $2 for every $1 they're making :-?
it's certainly possible and has happened before, but that's why I'm hoping for two things:
1.) open weight models gain more popularity 2.) i'm making more money from these models than paying for them
if either happens, then I'll feel good about it either way
No, they would never...
What do you mean? They need to recover hundreds of billions in sunk costs?