←back to thread

282 points _vaporwave_ | 1 comments | | HN request time: 0s | source
Show context
PaulKeeble ◴[] No.44999980[source]
This is a really common problem with science reporting in general. Its often the case that the news will say things about the paper that aren't in the paper, often they say something that is completely the opposite of what the paper actually represents with its data. Its become such a common thing and its very common when you can't find the referenced study itself from the article. Sometimes its the authors fault and they said things that aren't supported by the data but the science reporters do this a lot.

My basic rule on all science is go at least look at the papers abstract, method and their graphs/data. In 5 minutes you'll be better informed than the pop science article and it gets easier the more you read them.

Interruption do impact getting back in but I find it very variable, I actually if I am doing very strict TDD I recover from interruptions well. If I am busy thinking about a design or doing some more complex algorithm performance analysis its all happening in my head and they take longer. I think it is measurable and you could set up experiments to see how long it took to start producing again and if there is a slow start or not on a well defined programming task.

replies(3): >>45000015 #>>45000071 #>>45000418 #
falcor84 ◴[] No.45000071[source]
> This is a really common problem with science reporting in general. Its often the case that the news will say things about the paper that aren't in the paper, often they say something that is completely the opposite of what the paper actually represents with its data.

I wonder if perhaps a part of LLM hallucinations can be explained by them being provided such reporting and having it (mistakenly) tagged as high-quality training data.

replies(2): >>45000793 #>>45006465 #
1. devmor ◴[] No.45000793[source]
> I wonder if perhaps a part of LLM hallucinations can be explained by them being provided such reporting and having it (mistakenly) tagged as high-quality training data.

Probably (haha) far more of a function of temperature than training data. If the corpus is large enough for your prompt and you turn the temperature all the way down, you will get almost no hallucinations. You then have what is essentially a search engine.