mittermayr

I know what you're thinking... and I still can't believe it, but...

This morning, our database flagged a duplicate UUID (v4). I checked, thinking it may have been a double-insert bug or something, but no.

The original UUID was from a record added in 2025 (about a year ago), and today the system inserted a new document with a fresh UUIDv4 and it came up with the exact same one:

b6133fd6-70fe-4fe3-bed6-8ca8fc9386cd

We're using this: https://www.npmjs.com/package/uuid

I thought this is technically impossible, and it will never happen, and since we're not modifying the UUIDs in any way, I really wonder how that.... is possible!? We're literally only calling:

import { v4 as uuidv4 } from "uuid";

const document_id = uuidv4();

... and then insert into the database, that's it.

Additionally, the database only has about 15.000 records, and now one collision. Statistically... impossible.

Has that ever happened to anyone?! What in the...

Couldn't find anything online, but just wondered what the chances would be, if, much like SETI at the time, people would share some of their GPU processing time (small slots, here and there, when underutilized) to train an open-source community-owned model? Basically a model by the people, for the people :)

I wonder for how much longer we'll get to see (high quality) open models be shared freely, it's already quite a limited landscape.

Would there even be enough GPU power? Would something like it be doable with a reasonably realistic community effort? The human part of training could also be a part of this effort, which is yet another thing a single company is having to pay a lot of money for (well, I guess not anymore, as everyone is now helping out for free, by using the hosted LLMs).

Then, of course the problem of the data, not sure how far off the data private companies have bought, stolen, ...found is from what's available out there.

Curious if anyone more involved in with the topic has a thought or two about it.