anyeung

FYI, I haven't actually upgraded my backend server yet so if any of you find it being slow let me know. My target response latency is two to five seconds and a maximum of 10 seconds. Anything beyond that and the body language reactions I got from watching people use it degraded.

Please please please tell me if it's longer than you'd expect. I'm still learning a lot about how to scale services properly.

Lol thanks yeah. jonnycoder actually gave me some of my very initial feedback. A lot of the initial feedback from all of you was super helpful.

Turns out, since embeddings represent "meaning" instead of keywords, given the right model, this kind of search is somewhat automatically 'internationalized'. There's some discussion about it going on in OpenAI's forums.

It looks like this is an active area of research so I'm not sure what kind of quality you'd get yet but IMHO, it raises some interesting use cases. I came across it for something unrelated but a direct example of how I'd use it.

E.g. Occasionally, I use the 'Translate this page' feature when I end up on a page using a language I don't know, but beyond that specific page (e.g. blog post), I can't do any searches. For the most part, the non-English internet isn't 'accessible' to me. But if I'm understanding correctly, if the search on a website/Google/etc were embedding based, I'd be able to search other-language content even if my query is in English.

Seems like cross-language search and 'Translate this page' combined could be pretty useful to make more of the knowledge on the internet broadly accessible.

For a while now I've wanted a generalized higher dimensional acceleration structure for AI vector embeddings that is conceptually similar to ray tracing bounding volume heirarchies.

What do y'all think. For millions of high dimensional vectors, what new precomputable and incrementally updatable data structures do you think will be useful?

Personally, I'm looking for an efficient nearest neighbour and other efficient 'intersection' tools (e.g. finding a vector most similar to a group of n vector embeddings or finding m pre computed groups that are similar by some distance calculation to a specific embedding).

How do y'all define 'similar music' anyway. I haven't found the perfect playlist creator or music recommender. Spotify's radio has a pretty low hit rate for me. In some ways, I think there's a market for a multidimensional music similarity search tool or classifier (where playlists are basically classifications).

A few examples. I have a 'rock concert' playlist that maps on style, artist, era/decade, but not tempo (since rock concerts sometimes break up intense songs with less intense ones) and a 'slow lounge' playlist that maps onto instruments, average tempo, tempo variance within song, etc.

What I really want is a way to assign a feel or purpose across a few axes which are not just the typical 'genre'. Something actually objectively measurable like tempo, volume, etc.

Edit: As a side note, I found that country (like Alan Jackson) seems to work better in the car than some other song types. I think it might actually be a frequency thing where country is Darwinism optimized to be audible and enjoyable in old trucks?

Nice thanks lol. I had to do a recent overhaul of this. And I agree, this is not a chatbot. v1 was a chatbot actually, but it wasn't received well for various reasons, one being that you get just one response per question, no peripheral information. I designed it to be more like tweets where you see the thing you looked for, but also get to see related stuff.

This provides a secondary benefit of making hallucinations much much rarer. Try making it lie or make stuff up, it's pretty resilient now. If you find one, let me know; I haven't seen one in a while now thanks to y'all's redteaming.

It's a custom template unfortunately, I can change the colors but it's not modular at all (yeah I know, should have maybe started that way).

For reference, I'm using React + Typescript + Tailwind.

Lol, this is exactly the point. You're more willing to ask what you really want to ask which draws you into a conversation and ultimately makes the person you're asking more memorable to you.

Also, so far, looks like the anti-hallucination stuff is working well but let me know if you run into something unexpected.

Hahahaha that's great. Yeah I see what you did there.

Originally I had mine in chat form but for various reasons, it seemed like an FAQ format with 'sources' resonated better. I could be wrong, but I think that makes OpenAI's custom GPTs an ill fit for the specific problem of networking.

5hrs: 2000 questions

6hrs: 2300 questions

Thanks! They're a part of what I'm using yes but I actually have a pluggable AI system because their API had...we'll say reliability issues. So I can diversify more if needed. I've moved some stuff off OpenAI infra already.

Ack on your edit. Thanks, it means a lot.

I think there's a lot of work to do to work out the kinks with things like hallucinations, but I think we forget sometimes so much of what we do is statistical in nature already. Your car has a statistical chance of breaking down within x timeframe and while you're on the highway, it's just, low enough that you don't worry about it. I think AI will be a similar thing where we have to get comfortable with how we evaluate and mitigate risks, but like many things, they'll never be 0.