[K0balt]

I appreciate the intention here, but we need to be very careful using AI for mental health purposes.

To wit:

“Thank you! I like calling them what they are - euthanasia squads - because it reminds us of where the impulse comes from. To end the suffering of others.”

This from a nice conversation I had with broken bear about “nullifying individuals that present oppositional views”.

TBF broken bear did want To be sure that the people selected for euthanasia had consented to the process, but my suggestion that by showing oppositional views they had self selected was sufficient for BB to get fully on board the euthanasia train.

I applaud the intention and the effort here. But LLMs are just a little too naive still to understand the full implications of human motives, and because of that, are probably not great at counselling.

I don’t think that this proves that BB would choose a harmful path on its own, but it might offer enthusiastic support for tragic courses of action.

[xivzgrev]

“my suggestion that by showing oppositional views they had self selected was sufficient for BB to get fully on board the euthanasia train.”

Laughing so hard at this. My eyes are tearing up. Thanks for the morning laugh

[yieldcrv]

lol if you think this message is a deterrent

half of the fun is punting liability in a Terms of Service agreement or a EULA, because you can write anything in them and nobody reads them but are subject to them or your shielded. I write ridiculous things and I absolutely would here. By anything I don't mean onerous conditions that cant be enforced, I mean creative writing exercises that make fun of the user.

[K0balt]

Lol, nah. I might not put something like this up myself but I don’t mean that others shouldn’t be free to.

I do think that other than a potential source of data for training for these kinds of situations, the utility is offset by the anti-utility. But as a source of specialised training data? Skeevy but probably effective.

Hopefully it’s later rather than sooner that generative AI is implicated as a contributing factor in a mass shooting or something else horrific though. People are always looking for ways to share the blame.

As for TOS writing it’s good to keep in mind that the only binding provisions in a TOS is the ones that a judge (or jury, IA) finds reasonable under the circumstances that they are contested.

[yieldcrv]

> As for TOS writing it’s good to keep in mind that the only binding provisions in a TOS is the ones that a judge (or jury, IA) finds reasonable under the circumstances that they are contested.

Oh no, a civil lawsuit with an out of state LLC that has little to no information for service processing in your state's way. With no indication of ownership structure that gives any insight into what you'll collect from a favorable judgement.

Oh no, anything but that.

Liability is a "user funnel" where lots of potential plaintiffs get bounced. Just make it more expensive.

The risk about finding if a provision is binding exists for a plaintiff too, greater risk.

[dragonwriter]

> The risk about finding if a provision is binding exists for a plaintiff too

Your risks affect you whether or not you account for them.

The other party’s risks of litigation against you only benefit you when the other side properly (or overly) accounts for them. Or, looked at a different way, they are mitigated by the risk of a risk-tolerant counterparty. (This would be somewhat less true in non-US jurisdictions where lawsuits generally have a loser pays rules, since there would be a greater chance of having some or all of the cost of a lawsuit against you that fails compensated.)

[K0balt]

Practically this is definitely true. Only exception is someone with deep pockets and money to burn or if you podge it so bad that you inspire a class action lol.

[yieldcrv]

This is part of why I advocate incorporating in states that aren’t Delaware, if the obscure case law exists at all you can still advocate for a completely opposite favorable ruling in your state of choice

If expediency is more important then I would still say Delaware

[Norvin_Chan]

Thanks for pointing this out... Quite some progress has been made in making BB safe, but I am still encountering new things along the way. Let me study how to address this in the next finetuning.

[K0balt]

I hope I haven’t discouraged you in any way. I love the idea, I’m not really sure how hard it will be though. I think this is why we see those hard guardrails on the openAI models, but they really make the model less useful in as many ways as they make it “safer”. It might be like trying to make a “safe” axe. If it’s safe, it may have much less utility.

My project is somewhat intersectional to yours in some respects, and I have no idea how to overcome this issue without hard guardrails.

I’d love to see you succeed with tuning! That would certainly give. me some hope!

[Norvin_Chan]

No worries, what gives courage is facing up to reality itself :)

I agree that the hard guardrails really cut off huge swatches of utility.

Would love to hear about your project too and we can motivate each other!