Might be nonsensical, but one thing I sometimes wonder is why I reach for reducing a list to a value more often than I need to generate a list from a starting value. I guess the asymmetry has something to do with the kinds of applications I work on.

But as others have mentioned this is only as intuitive as F=ma, or p=mv.

In my view, at least classically it's just a matter of definitions then. If our definitions of energy differ, the only thing we will experimentally agree on is the equation of motion, and even then up to a frame transformation.

This has pushed me toward cheaper, less capable models and made me hesitate to even explore certain ideas. I don’t want every side project to become another $20/month subscription, but I also can’t compete with VC-backed companies willing to subsidize inference costs.

Then I had this idea: what if users simply paid for their own inference?

This already happens in some apps through locally configured API keys, but could the model be extended? If users bring their own AI account, developers can build AI-powered products without taking on variable inference costs. How many AI applications shouldn’t be subscription businesses at all?

The challenge is that developers don’t want to handle user API keys, users don’t want to hand them out to every app they try, and nobody wants the friction of collecting payment methods just to pass through inference costs.

That’s the backstory to my latest side project, infer0.com.

It's a bit like SSO for AI inference. Users connect their AI providers once, and apps use auth tokens to request inference through infer0. Developers don’t manage API keys or pay model costs themselves, while users can bring the same AI accounts across multiple applications.

This may be a terrible idea, both because nobody will trust it and because I’m sure there are risks around handling user API credentials that I haven’t fully appreciated. But I felt the need to build it. So here’s a rough first pass.

Comments URL: https://news.ycombinator.com/item?id=48556523

Points: 6

# Comments: 1

[1]: https://github.com/oven-sh/bun/issues/21237

this is amazing feedback, thank you for taking the time to write it all out. I love the idea of audio feedback!

I accidentally linked to a specific note at the top which I guess Cloudflare is struggling to preview due to traffic.

Hacking fingernotes.com together over a couple of weeks was a creative outlet when work got stressful. I think of it as digital sticky notes. The goal was to make notes with a personal touch that are easy to write and share. I also wanted them to appear as their own link preview image on supported platforms. That way when you send the link to a note, the person sees the message without following the link.

Let me know what you think!

I drew inspiration from Apple's quick notes: low latency made scribbling a pleasure, and sending notes to friends felt warm and original compared to a typical exchange. It was also intriguing to see my handwriting printed in a message chat. In a time of rising artificial generation, spreading my clumsy handwriting feels like an act of rebellion.

But I dislike the light background in Apple notes, which I don't think you can change when sharing. More importantly, no one sent a note back. With fingernotes the low-friction interaction is meant to make creating notes simple. I also find the image previews aesthetically more pleasing.

For implementation, fingernotes are publicly accessible links to collections of strokes that have been persisted to a Cloudflare D1 database and rendered in SVG. Like pen on a sticky note, each stroke is immutable but anyone can add to a note if they have the link. You can't undo strokes, so if you mess up your note just throw it out and start a new one. Having append-only collections avoids handling order of operations when multiple people edit the same note. Hosting it as a Cloudflare worker made it easy to get up and running. There's some latency in Safari on iOS which is absent on desktop. It's noticeable compared to Apple notes and I'm afraid it's a limitation of the browser.

Comments URL: https://news.ycombinator.com/item?id=43451222

Points: 67

# Comments: 20