Can you simply brainwash an LLM? • Hacker News Board

Re: Can you simply brainwash an LLM?

#71

Post by jdiff » Tue, Aug 01, 2023, 8:22 PM UTC

laverya wrote: post #66

jdiff wrote: post #8

Earlier quoted context omitted.

No, it isn't. This is akin to saying that the truth is relative and lies somewhere between "the Earth is an oblate spheroid" and "the Earth is flat." Perception and perspective varies, sure, but fact exists regardless. Fake news is falsified news built on fabricated fakes, and is not just alternative viewpoints. Do not normalize this.

"Japan has a higher GDP per capita than Alabama" is fake news. It's also confidently repeated by most LLMs. https://twitter.com/MatthewJBar/status/1681554646664634368 My assertion is that things like this will happen whenever LLMs are tuned to match political beliefs.

Nobody fed that false fact into any LLM. Garbage in, garbage out isn't fake news.

Re: Can you simply brainwash an LLM?

#72

Post by davidguetta » Thu, Aug 03, 2023, 2:10 PM UTC

catchnear4321 wrote: post #70

davidguetta wrote: post #61

Earlier quoted context omitted.

They really just seem bad faith in this thread. Just publish the training data FFS (medical data excluded)

then it wouldn’t be the training data, it would only be a subset. if the issue is sensitive data in a training dataset, perhaps that should be addressed rather than accommodated.

The point is that they seem to pretend they want to redefine the meaning of open source BECAUSE of medical data.

Just say medical data is not open source and make the rest really open source

Re: Can you simply brainwash an LLM?

#73

Post by catchnear4321 » Thu, Aug 03, 2023, 6:39 PM UTC

davidguetta wrote: post #72

catchnear4321 wrote: post #70

Earlier quoted context omitted.

then it wouldn’t be the training data, it would only be a subset. if the issue is sensitive data in a training dataset, perhaps that should be addressed rather than accommodated.

The point is that they seem to pretend they want to redefine the meaning of open source BECAUSE of medical data. Just say medical data is not open source and make the rest really open source

even reference would be sufficient if access were not controlled and denied under the guise of protecting people.

the problem is the source medical data itself is insufficiently cleansed. (if it can be at all.)

ideally the medical data is open source, but only contains what’s necessary, and not what’s sensitive.

that is, obviously, messy…