logoalt Hacker News

augment_me • today at 4:12 PM • 2 replies • view on HN

This is a take that is outdated if not now, then in 3 months.

These has been this ladder of copium that "oh human still has input here for X, oh shoot it doesn't with Claude 4.5, well human is still needed for Y, oh shoot...".

You are suggesting that you can provide context that the model can't on its own. This is a search problem, every iteration it gets better.

We are at the stage in our company where "/goal optimize this" gets better result than any biased human context engineering can provide.


Replies

zahlman • today at 5:30 PM

> You are suggesting that you can provide context that the model can't on its own. This is a search problem, every iteration it gets better.

How exactly do you propose that the model should search my thoughts? Of course I can provide context that the model doesn't have. I can provide context that does not exist anywhere on the Internet (unless I got hacked).

> We are at the stage in our company where "/goal optimize this" gets better result than any biased human context engineering can provide.

First, a "this" must exist to optimize. Second, you have to be happy enough with it that optimization is the only problem.

shmolyneaux • today at 4:42 PM

I think it's reasonable that humans need to provide context for this, and situations like this, forever. The GP has a specific problem. The context of the problem exists in the real world, not something that the agent can search for. This isn't a "humans are still important" take. This is custom software for meeting the specific needs and preferences of an individual.

I fully agree with them that the easiest solution is to ask an agent for exactly what you want, and iterate from there. Evaluating and installing and learning off-the-shelf software is often more of a pain than doing it from scratch.

➕ show 1 reply