Show HN: I made an emotional support AI companion | Hacker News Reader

Show HN: I made an emotional support AI companion

sayfli.com

3 pointsremusnegrota10 comments

Hey HN,

Remus here, solo dev building side projects after my 9 to 5.

I've always had this idea spinning in my mind for a truly conversational AI that is curious, asks questions, and is compassionate. Something that understands and resonates with you can offer a safe space to talk about problems that may otherwise be harder to discuss with others.

In crafting Sayfli, my mission was clear: to complement, not substitute.

Sayfli isn't here to replace the irreplaceable: the warmth of human conversation and the invaluable exchanges with those we hold dear. Rather, it's here to be the first step on a path towards open, meaningful engagement with ourselves and, ultimately, with each other.

While Sayfli provides emotional support, it is not a replacement for professional therapy and is not designed to be used as such. It is meant to serve as a preliminary step towards self-awareness and can be used alongside professional counseling.

Open articleView on HN

What AI does this use? How can it be "end to end encrypted" if the server has to decrypt it to run an AI model?

It is based on GPT4. Also tried Mistral, but found it less capable. Thinking on testing LLama as well.

Communication is encrypted in transit and at rest. May have to change the copy a bit to avoid confusion

I hope this doesn't come off as a shallow dismissal, because I'm sure you have good intentions and you've obviously put a lot of thought into this.

But on a gut level, it sounds like you're trying to pass off a heaping spoonful of marketing doublespeak here.

In particular, the terms "emotional support" and "trusted companion" and "compassionate" are inherently misleading when used to describe these tools. And to characterize them as such is potentially quite harmful to people facing a serious mental health crisis. Worse, you say "it isn't here to replace the irreplaceable" but when you use such terminology you are implicitly conveying the impression that these tools will do exactly that.

Drop those and similar terms from your promotional materials, please. "A preliminary step toward self-awareness" sounds infinitely more plausible.

I understand your point, but the word selection can be quite subjective. I've taken a lot of care when finding the right words. Copyright will improve as it evolves.

I don't understand your reference to people with a 'serious mental health crisis.' I've specifically mentioned this should not be used as such and it is not marketed as such. I'm a firm believer that people in that condition should get professional therapy.

I don't understand your reference to people with a 'serious mental health crisis.'

It comes from the fact that a significant portion of your users will surely be in that category.

I understand your point, but the word selection can be quite subjective.

Right, but in such cases (and especially when dealing with mental health services) one should apply the cautionary principle. And be extra careful to not use words that suggest that your service provides something ("compassion", "emotional support") which it plainly -- does not.

I am well aware of the stigma behind AI and empathy, but a recent study showed that LLMs are starting to provide better emotional support than most humans. I know these are crazy times.

Look, I understand your points and am well aware of the thin line that the product must thread, but this is the product's mission: to educate people on the roles AI can have, and the main goal is to talk out loud your thoughts.

The most important thing is to take steps towards becoming self-aware of our thoughts by speaking out loud: This is what is meant to be, this is what is presented, nothing more.

I will revise the copyright to make this clearer.

A recent study showed that LLMs are starting to provide better emotional support than most humans

Can you provide a citation?

Sure, here is the paper: https://osf.io/preprints/psyarxiv/fzvd8

And here is a post from one of the contributors on the research, that summarizes it: https://x.com/Amit_Goldenb/status/1782076346929475602

Thanks.

The paper describes LLM performance on a single, very narrowly prescribed task - compared with (if I am reading correctly) average humans (not trained practicioners in the field).

This is of course very, very far from a validation of these models as being on the whole suitable for use in real-life counseling situations. Involving, you know, real people facing a diverse set of mental health situations and risks.

Look, it's a fine technical achievement, your site. I'm pretty sure also that what I'm saying above is not news to you, and that you're perfectly aware of the current primitive state of these models. And that you're also aware of the fact that none of technology has been formally vetted or approved by anyone in the public health field.

In that sense -- it is really quite irresponsible of you to promote this tool in the way that you're doing it. Especially the language implying this is finished, trusted product ("Trusted AI companion"), and with human-like qualities no less ("is curious, compassionate").

And on top of that, to take people's money for it -- given that these tools can be quite addictive. And you're potentially exploiting people in very difficult situations and with limited financial options.

I know you have other disclaimers, but that's what I mean: These disclaimers are useless when they are contradicted by other statements (such as those above) which say basically the opposite.

That's why the disclaimers come off as -- doublespeak, basically.

You need to really re-think what you're doing here, and why you're doing it.

Thank you for your view and for the civilized debate. Still, you are steering it to extremes with arguments such as: " And you're potentially exploiting people in very difficult situations and with limited financial options."

I have to disagree with you on that, you are basically comparing my product with gambling, which is unfair.

That being said, these kinds of discussions, in general, have a net benefit in the world.

Reply on news.ycombinator.com