Remix clone Hacker News

The devil is in the detail, I think in the age of LLMs this will become harder as docs can now be better scanned before they’re handed out.

▲

rob74 4 hours ago | parent | next [-]

I think current LLMs would be completely lost with the task "scan these documents for information that could be risky for our case": they would either produce tons of false positives, overlook actual risks, or both.

▲

maeil 2 hours ago | parent | next [-]

I can personally attest that companies are already using those LLM-based workflows for those exact purposes successfully, performing better than humans, which often isn't a super high bar.

▲

mschuster91 3 hours ago | parent | prev | next [-]

At least you can reasonably use them for translation and in a second pass summarize or just have it detect if a document contains specific keywords / phrases.

▲

chx 2 hours ago | parent [-]

Yeah but this is where a human is indispensable.

Suppose the source language has words which could be translated into multiple English words -- look at all the power banks which are advertised as charging treasure for example. One can see how bank vs treasure are close, after all. Further, even in English multiple phrases could be used "don't preserve" "bin it" "throw it away" and countless others. Even worse, it could be a company specific phrase which would only stand out as odd "apply procedure 66 to it".

If given a little thought this is exactly the kind of task where a native speaker would shine and LLM might just miss or if given a wide enough net produce a million false positive.

▲

infecto an hour ago | parent [-]

Just tested out the treasure idea and seems to work fine in ChatGPT. I suspect if given context at the task at hand that the LLM would provide a pretty decent first pass.

▲

chx 20 minutes ago | parent [-]

Y'all are hopeless. I already asked dang to delete this account.

	▲	infecto 13 minutes ago \| parent [-]
		What a weird response. Just providing a counterpoint to your idea that I don't believe is correct. I hope dang does not delete your account because I am not sure whats going on here and I hope you find the help/peace you need.

▲

infecto an hour ago | parent | prev [-]

You might be caught off guard then because its already happening successfully. There is a lot of overblown hype but systems already exist that do a pretty good job on first passes and other work that paralegals/juniors would be doing. I know this goes against the anti-LLM narrative here but don't be surprised as it happens.

▲

ano-ther 4 hours ago | parent | prev [-]

In a discovery you don’t really have a choice.

On the other hand, perhaps the summer intern would have had an easier time finding something (and the defense more time to spot potential gaps in their arguments).