this post was submitted on 29 Aug 2023
52 points (96.4% liked)
Technology
59219 readers
4771 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I mostly agree, with the exception of the last paragraph. The software world today is built on open source software that is distributed without guarantees and this is an amazing thing. If a liability like that is applied to AI, it would be easily generalized to any piece of software that might be used in malicious ways the author did not intend or predict. It would be akin to punishing a knife manufacturer for a stabbing.
It is ridiculously easy to write a genuinely well intentioned script that encrypts your files and a minor modification can turn that into a ransomware.
That's true, but thinking about AI that is made to generate speech, processing power is still expensive enough that developers are careful with it. But what happens as memory gets cheaper and calculations get faster, and ordinary developers are able to train their own generative AI?
For example, what happens when a developer decides to train a LLM extensively on scam emails, and spammers love to buy copies of it - but the developer markets it as just "a helpful generative AI"? Or, what if a person trains their LLM on an extremist forum full of hate speech and disinformation, then offers it to a suicide prevention center as a 24/7 alternative to human labor? (Treating these as hypotheticals, where we assume the difference isn't immediately obvious. Perhaps they also used some legitimate training data, so that most outputs seem innocent enough.)
To me it sounds more involved than selling just a word processor with autocorrect, but less involved than selling an instruction manual for committing crimes.
That happens all the time since GANs entered the scene, and before that since alexnet broke image classification records in 2012 using consumer hardware. Anyone can train neural nets.
Ok, let me be more specific so that it's not open to uncharitable interpretation.
What happens when it becomes easy to make something as reliable and complete as, e.g., ChatGPT-4 without the hardware costs and other costs currently associated with it?