Surely the only rational reason to care about liability is you get paid in the bad outcome.
Let's assume you get paid 0$ for AI mistakes -- no liability at all. Then your EV is roughly:
(AI accuracy - human accuracy) * (cost of accident) - (1 - human accuracy) * (human payout)
Now I don't know about you, but even with pretty big numbers on the human payout scale and relatively small differences, I would not accept 10 million (presumably paid to my next of kin) to die. Does lead to weird situations where I may prefer humans for minor medical procedures where I can take the risk -- but only in edge cases.
Or is there some weird punishment thing going on here?
Data point of one, but after a few months of uselessness, I have managed to get some pretty serious, measurable performance improvements with AI optimizations -- order-of-magnitude speedups of business critical processes which took days as well as significant latency reductions.
But you need a really solid workflow, solid benchmarks which run quickly, and a lot of tokens -- plus a rigorous profiling workflow.
I think the issue is precisely the rationalising part.
Let's assume I have pre-decided something: if the fruit weighs over 100g, I'll buy it, otherwise I won't.
I can then measure and find out. My emotions -- based on looking at the fruit -- may mislead me here.
Now what about "likely to weigh more than 100g"?
Here it's really easy to lie to yourself.
But if you set conditions for what counts as likely before you know what they'll be used for, and you stick to them, you can make a decision based on reason.
There is no way to decide 100g is the threshold divorced from emotion, however -- unless it's rationally derived from something else emotionally based.
The "rationalisation" bit comes from us not making decisions for the reasons we think we are -- or in the way we claim to be.
A better technical solution imo is more P2P hosting -- say content addressing. Something designed to deal with the scrapers -- and maybe even make them bear infrastructure costs.
Honestly, I'll take it a step further: what's wrong with training?
Having principles means applying them uniformly -- even to large entities or those you hate (it's fine if the principles themselves have size bounds in them, though -- versus them being implicitly glued on -- but then you need a universal justification for why that size. Which is possible and valid.)
I think that one should use the best algorithms, the best information, the best knowledge they have access to -- period. I don't like using gimped machines, I don't like making gimped machines, and I certainly don't like being sold them.
So I'm not going to turn around and say LLMs need to be gimped via arbitrary restrictions on their training data.
I agree. As soon as I understood the gist of how modern models do what they do, I found it unreasonable to apply any stricter standards to their "learning" than we do to humans. Humans learn by reading ideas and looking at images, then go on to have ideas and draw images based on that training.
We have definitions of plagiarism and copyright infringement that apply to people based on what they put out into the world, not how they trained themselves to get there[0]. Artists literally trace art and study specific examples in detail to learn, and that's not a bad thing. Humans can also accidentally plagiarize or make things identical to past works - it's easy to think a great guitar riff just came to you when it was really based on a song you heard years ago that stuck in a part of your mind but you don't even consciously remember the influence, for example.
It would certainly be nice for AI output to provide citations if it realizes it's using a significant chunk of an idea from its training that has a clear source (or many), though this is difficult in the same way it would be difficult for me to cite where I learned about the Towers of Hanoi. The AI frequently does web searches for specific resources to get ideas from these days, which are easy for it to cite.
However, it would have a chilling effect on progress as a whole if we all started jealously guarding our ideas so close to our chest that machines couldn't read them and only a select few humans who passed some kind of gate (or even paid us) were allowed to see them. We would be nowhere close to where we are if we had always had such a mindset.
---
[0] We use proven absence of viewing certain material as a legal shield against copyright infringement e.g. clean room engineering, but this isn't strictly necessary, and possibly even discouraged with modern precedent: https://reactos.org/forum/viewtopic.php?t=21740
Content-addressed storage and error correcting codes mean that one can make bitrot astronomically unlikely with honestly minimal infra investment.
It's copyright that causes anything to disappear from the web IMO -- torrents never die.
EDIT: I am aware that unseeded torrents do in fact die. But it really doesn't take much to seed a whole hard drive's worth of rarely requested data -- this also detects bitrot and so corrects errors automatically if you're not the only copy.
If you are, there's ECC, as well as making another copy.
There are mitigations in both software and hardware, but most consumer machines, by default, do almost none of that. No ECC RAM, no error correction in the filesystem.
Let's assume you get paid 0$ for AI mistakes -- no liability at all. Then your EV is roughly:
(AI accuracy - human accuracy) * (cost of accident) - (1 - human accuracy) * (human payout)
Now I don't know about you, but even with pretty big numbers on the human payout scale and relatively small differences, I would not accept 10 million (presumably paid to my next of kin) to die. Does lead to weird situations where I may prefer humans for minor medical procedures where I can take the risk -- but only in edge cases.
Or is there some weird punishment thing going on here?
reply