"Every great cause begins as a movement, becomes a business, and eventually degenerates into a racket"

- Eric Hoffer

I've been playing the Ocarina of Time remake and whenever I get stuck trying to get a certain item or progress on a quest, I'll fire up Claude voice chat and ask it for guidance, and it regularly gets confused and gives me wrong answers (somewhere around ~30% of the time).

Given how long these models have been out and that it's a highly constrained game world from almost 30 years ago which is documented top to bottom, I struggle to understand how Claude is still so bad at straightforward questions with straightforward answers. This was a solved problem (albeit with a less modern UX) in the legacy internet and it doesn't actually feel like getting the right answer is necessarily easier.

▲

elevation an hour ago | parent | next [-]

> I struggle to understand how Claude is still so bad

Claude is absolutely killing it for me. Man-years and man-years of backlog work, hobby projects, optimizations, research, experiments, finally implemented. Bloated apps finally optimized to fit on micro instances, slashing my infra costs. Recovery from operations incidents in a third of the time.

If I could make Claude forget everything it every knew about Ocarina of Time to make it 1% better at what it already does so well, I'd press that button in a heartbeat.

▲

threetonesun an hour ago | parent | prev | next [-]

This feels like a very odd use for Claude as opposed to any LLM focused on web search, but more to the point 30 years ago you could hit up Gamefaqs and get a plain-text document to download and manually search it locally to your heart's content in milliseconds. A lot of LLMs are hitting the mountains of garbage in, tiny pile of garbage out problem more than being bad LLMs.

▲

hhh an hour ago | parent | prev | next [-]

Why would an LLM be good at giving you answers about specifically the ocarina of time remake? it seems like the worst case scenario for confusion

▲

dragontamer an hour ago | parent [-]

Because documents like this exist and it's an implicit assumption that LLMs have read them

https://gamefaqs.gamespot.com/n64/197771-the-legend-of-zelda...

Google used to return these documents, or at least forums and communities who knew where these guides were. Google no longer connects communities together however.

LLMs don't either but the hope was that they at least read all of the old internet and have it somewhat memorized.

	▲	arscan an hour ago \| parent \| next [-]
		Haven’t we moved past an assumption/need that these (frontier) llm (-backed services) have memorized their training data now that tool calling has proven to be so effective? The model knows enough about the question to be able to search for sources that contain the answer and then summarize the findings back? I’m shocked Claude voice ever makes a mistake on this, does it not have agentic capabilities backing its responses?
	▲	pjc50 an hour ago \| parent \| prev [-]
		God I hope all that lot is backed up somewhere. Maybe I need to take my own backup, all that ascii should be cheap.

▲

hackersnooze1 an hour ago | parent | prev [-]

llms seem really bad at giving information about existing video game for whatever reason. it just makes shit up half the time.