3 ms·

What makes you think that the decision making will be hard to automate? In a couple of decades, keeping this pace of development, doesn't it seem plausible that

by ori_b 2y ago

What makes you think that the decision making will be hard to automate? In a couple of decades, keeping this pace of development, doesn't it seem plausible that AI will be able to extract requirements from observed behavior, and decide how to fill them better than a human can?

viraptor 2y ago
Because good decision making requires lots of external context and interpersonal assumptions. These things are not even documented. It's like every outsourcing project for code that I've seen: yes, what was delivered matched the written requirements, but it's useless and needs lots of changes. There's still going to be space for people who understand how things work both in code and in practice. (Cue Office Space "I deal with customers so engineers don't have to. I've got people skills!") There are also right now cases of: I can write a paragraph of what change I need, or change those two lines myself. I don't expect they will go away entirely and having an agent available will be closer to an extra skill than a full replacement.
- ori_b 2y ago
  It seems like observing usage and setting up a feedback loop, then incrementally a/b testing, is certainly not something that can be done today, but doesn't seem beyond the realm of possibility a couple decades from now. Edit: In fact, now that I think about it, I'd probably be more ok dealing with a requirements document and testing plan generated by chatgpt today from a summary than I would be dealing with code -- hallucinations tend to have less impact at that stage. Requirements tend to be wrong anyways until they're implemented; code is where the rubber hits the road. The missing piece is having LLMs get better at integrating and reasoning about real world data, so that it can iterate on user testing. Heck, at that point, models can decide what to build, and not just how to build it; there's a treasure trove of data that they can access much faster than humans to find gaps in the market and generate solutions to fill those gaps. Assuming we don't hit a wall with model capabilities, we can focus purely on consuming the outputs of AI, as it optimizes itself to best keep us happy.

by ori_b 2y ago

viraptor 2y ago

Because good decision making requires lots of external context and interpersonal assumptions. These things are not even documented. It's like every outsourcing project for code that I've seen: yes, what was delivered matched the written requirements, but it's useless and needs lots of changes. There's still going to be space for people who understand how things work both in code and in practice. (Cue Office Space "I deal with customers so engineers don't have to. I've got people skills!") There are also right now cases of: I can write a paragraph of what change I need, or change those two lines myself. I don't expect they will go away entirely and having an agent available will be closer to an extra skill than a full replacement.

ori_b 2y ago
It seems like observing usage and setting up a feedback loop, then incrementally a/b testing, is certainly not something that can be done today, but doesn't seem beyond the realm of possibility a couple decades from now. Edit: In fact, now that I think about it, I'd probably be more ok dealing with a requirements document and testing plan generated by chatgpt today from a summary than I would be dealing with code -- hallucinations tend to have less impact at that stage. Requirements tend to be wrong anyways until they're implemented; code is where the rubber hits the road. The missing piece is having LLMs get better at integrating and reasoning about real world data, so that it can iterate on user testing. Heck, at that point, models can decide what to build, and not just how to build it; there's a treasure trove of data that they can access much faster than humans to find gaps in the market and generate solutions to fill those gaps. Assuming we don't hit a wall with model capabilities, we can focus purely on consuming the outputs of AI, as it optimizes itself to best keep us happy.