Hacker News Reader

Comment by simlevesque | Hacker News Reader

There's a lot of LLM haters, simple as that.

For posterity: Claude is no longer “just” an LLM you’re interacting with.

I use AI everyday. But some AI adopters are getting a bit culty as well.

But there is nothing culty about saying “an LLM could one-shot this” when it has clearly been demonstrated that an LLM can, in fact, one shot this!

You must have a different definition of "one-shot"

"One-shot" means you type one prompt into Claude, press enter, and then judge the results when the AI completes the task. The article says that Claude was able to take a single prompt and produce a pixel-perfect replica.

Is there another definition of "one-shot" I'm not aware of?

Yea, the one that includes all the previous "one-shot" failures.

If you throw 4 basketballs at a hoop and only the 4th one makes it in, that's not "one-shot". It's 4.

Your definition appears to be, after many failed attempts, a single prompt that gets the AI to complete the task was found. You called that "one-shot" but what really happened is there were several "shots" that failed.

If three of my friends that have never played basketball take a shot and they all miss, that doesn't say anything about my ability to make a shot. Obviously, my friends are inexperienced with basketball. There is nothing wrong with that, but it's not a fair or just argument to tell people who do have experience that "your success not a true one-shot because other, inexperienced people, failed".

You appear to be extending the definition of "one-shot" to "an LLM should be able to accomplish anything no matter how poorly it was prompted", which is, I hope you can agree, not very reasonable. It's certainly not something I would be arguing in favor of.

your shot is not dependent on your 3 friends. The "one-shot" prompt only happened because they took the info from previous failed attempts to guide their prompt

You seem to be saying that I never would have thought of how to prompt Claude until I became inspired by reading the post yesterday. But I use Claude with Playwright + screenshots basically daily.

I can one-shot a 10-footer, might need few attempts to hit a half-court shot. not all shots are created equal but I have seen many full-court one-shots by LLMs which would have taken human (other than Steph) tens if not hundreds of shots :)

This isn't in a vacuum though, it's after someone else failed and they took the learnings from that and put it into the prompt.

> I sort of do agree that Claude Code on its own would not be able to do this. But Claude powered by nori configs absolutely should be able to.

Yeah not to mention setting up additional tooling

Reply on news.ycombinator.com