[aocallaghan17]

I just don't see any way you can work like this and maintain comprehension of the system being built?

Perhaps for toy projects or research, but for a production system you're accountable for understanding, maintaining and continuing to develop?

[pixel_popping]

You don't need to "maintain" a comprehension as you can just ask (with loops as well) anytime you want something, or you can ask the tool to give you the current state/summary. Actually, no model should directly answer to you in a proper workflow, it should always be another agent digesting and verifying, don't check the response directly.

In reality, production systems will be released without inner depth of knowledge anymore, because no humans will touch tomorrow's codebases, solely AI, so everything has to be designed for AIs, not for humans at this stage, same for documentations.

Documentations don't need to be done ahead as well as they can be prompted live, docs should be just pointers to assist AI to help you gen the docs. In my team we stopped having dashboards pre-made entirely and if we need to know how many people signed-up today (just an example), then agent hit prod data directly (with read-only instant snapshots), we kept having this discussion and we ended-up understanding that inventing "tools" that we aren't even sure we need is useless in this era, you'd rather prompt everything (in loops, adversarial with the model zoo and so-on to reach 99% accuracy).

[aocallaghan17]

I'm just struggling with this, surely you need inner depth knowledge to reason about the system and make some level of decision, at least around system design and architecture if not lower level implementation details? But it sounds like you're generating that knowledge each time through a system of agents? How do you have so much trust in a non-deterministic system, or are you deferring ALL decisions to these "loops"? What if you and a team member generate a dashboard and it gives different results because the agent(s) used a different methodology?

And surely cost plays a part here. This is giving you such productivity gains to boost revenue enough to outweigh what must be huge token costs?

[saidnooneever]

AI is doing the comprehending full stack here. they dont care because if there is an issue or request AI will handle that too. (i dislike it but its what people do.)

[cyanydeez]

If you start working with local LLMs, you'll find they _do_ work at a speed you can comprehend and you can guide them to the thing you would understand more.

I've been building projects mostly by forcing the llm to use a more modular approach, so when it does get stuck or break something, it's isolated. This happens to be a often promoted means to an end. Additionally, I've gotten into building and inspecting tests/ that allow it to break problems into functional blocks.

but tl;dr: you're seeing AI psychosis because token gen is through the roof leading to sprawling code bases that no one's competent in being responsible for. But if you tweak how fast these things operate, you can manageably come along for the ride even if you still dont need to know everything. Think of it like a new form of boilerplate, but extended to a lot of banal ritual magic.

[aocallaghan17]

Yeah I mean this is closer to my use of LLMs, where I'm intentionally slowing things down enough to follow and course correct myself. This whole build the loop not the prompt idea seems to be advocating the opposite... I don't feel comfortable to not be the driver of the loop for a production system.