Show HN: Map of YC Startups
yc-map.vercel.appHey Everybody! Hope you had a merry christmas
Today I had a bit of fun with Claude.
Started by scraping YC's startups list, then ran them through OpenAI's embedding service, then UMAP'd the embedding to reduce the dimension to just two coordinates and then just forced Claude to write React that would compile to visualize that.
I had fun and I think it's interesting, so take a look!
Also note that you won't be able to zoom on mobile (found about this Plotly limitation way too late). If there's interest I can fix this issue by changing plotting libs tomorrow :)
Merry christmas
Love this! It'd be interesting if some builds this but adds more dimensions (similar to Company status) to it that you can query or group by. For example, if I look at S21 and W21 batches, then it'd be nice to know things like -
1. How many of these companies made it to series A, series B, etc
2. How many of these companies have > x employees (where x can be 5, 10, 20, etc)
3. How many of these companies had a founder that moved on to something else
This does require a lot more intelligent data scraping or manual data collection though.
There's no need to include an X & Y axis, labels and gridlines if they all have no meaning. A simple cluster diagram is enough.
I agree it would be less confusing if they weren't there. I'm sure I'm not alone in spending some time trying to work out what the axes were.
Cool project, but missed opportunity to name the arbitrary dimensions Y and C...
My dumb ass was trying to figure out what each dimension meant
That doesn't make you dumb; there is no intuitive meaning for the axes chosen; you can think of them, roughly, as statistically chosen to maximize clustering.
Statistically chosen to maximize *some particular loss measure, which in this case might be the t-SNE or UMAP criterion, and is computed only globally and not for different filters.
Right (I mean, I'm saying "right" but really I should just say "I'm taking your word for it"), but even more fundamentally this is dimensionality reduction from an OpenAI embedding vector, which seems almost like the asymptotic limit of inscrutability.
same
OP made the change
haha awesome, shipped!
I figure why not plot them with an X and Y (Y,C) of some sort
It’d be nice to just see the name of the company on click instead of going to the website (I’m on mobile). Trying to find our company
This is awesome! Are you able to also add F24?
i'd like a filter by target market (US, EU, APAC...)
Coming for v2
Cool concept! What are the X and Y axes?
Oh, and your website has an unchanged Wordpress favicon...
They're semi-arbitrary, dimensionally reduced from OpenAI embedding vectors.
Love that, what are Axes Y and C?
Apparently inspired by a comment on this very post! (Above yours, right now.)
> Cool project, but missed opportunity to name the arbitrary dimensions Y and C...
Really nice to see - also, It would be great when filtering if there was a tabular view at the bottom as well.
Really neat! We were Tule, in the industrials part of the map in grey.
There's something wonky when I zoom in on Chrome on my laptop. It abruptly shifts to another part of the map.
fun, though I also got stuck on what the Y and C axes represent initially. IMO just hide the axes altogether, since the goal is just some visual clustering/similarity
Maybe I'm slow, but clustering on what dimension? The lack of axes and labeling makes it pretty confusing to me, but I'm a dinosaur.
Visuals that are not self-explanatory make me feel dumb.
We don't know what to label those features/dimensions, because they're a reduction form higher dimensions that we also didn't bother to interrogate.
It's possible to figure them out. I wish OP would.
OP here, Is there a way to figure that out?
(Not OP) I can think of a convoluted and expensive pair-wise comparison method, but I hope there's also a way to figure this out during the application of principal component analysis in a way I don't understand.
Edit: I'm thinking it can't be done without experimentation on the embedding model.
Edit2: Ah, even that might not yield results, because as the basis is derived interstitially through computation, there's no guarantee the features of the final coordinate system will have any accessible relationship to those of the initial basis.
Animated Gif of each category:
https://imgur.com/a/ycombinator-startups-map-iNX8k6M
Filters are unreadable on mobile.
should be fixed now, thanks!
Company status isn't up to date... I know there's more than 1 public company that went through YC.
Check the filters, not all batches are selected as default. Only the latest ones. If you select all of them, then there are many public companies
hella nice mate very interesting
what's the x and y axes?
they don't have meaning by themselves. they are two dimensions that umap projected the original embeddings down to in order to show a combination of local neighborhood similarity or closenes
Well, they do have meaning by themselves, but it's more work to figure that out. All regular, predictable relationships "have" meaning because all meaning is prescribed. And since we've captured many such prescriptions in LLMs, they can do a decent job approximating those.
Nice! What's the tech stack?
For scraping and all the processing, typescript. Embeddings: openai
For visualizing react (nextjs) + plotly (though the lack of mobile zoom makes me question if I should chsnge it)
I didn't know YC does Government, Healthcare, Industrials, Real Estate and Construction. All these are great sectors and never made the headline.