badlogic

badlogic, 2 days ago

@sinbad sometimes useful things fall out of that pile of money that can be used by us plebs to do meaningful things. It's not all bad.

badlogic, 2 days ago

All the preprocessing is about 200 LOC (including PDF parsing, talking to the LLM, embedding & projection), thanks to the easy to use libraries available nowadays. You just need to know what you're doing. :D

Code here:
https://github.com/badlogic/wahlomat/blob/main/preprocessing/main.ipynb

(see also llm.py and pdftools.py, which I stole and modified from somewhere on the web)

badlogic, 2 days ago

"Mario, this could have been a chatbot". That's actually what I wanted to do first, but that 1) costs more to provide it on a level that doesn't hallucinate half the time and 2) is less exploratory in nature. You must know what you are interested in and wouldn't be able to find serendipitous info.

E.g. I didn't know most parties fucking LOVE trains.

badlogic, 2 days ago

Side note: I built the stupid pan/zoom scatter plot myself cause every single library out there that's supposed to do that is ... not great.

It's not perfect, but it is mine (and now you can use it too if you don't mind Lit Elements. You can easily extract the logic to vanilla JS tho)

Funny how that took 2x the LOC than was needed for all the fancy ML-y stuff. Sigh.

https://github.com/badlogic/wahlomat/blob/main/src/utils/plot.ts

badlogic, 2 days ago

It also wouldn't have allowed me to grasp, that the greens are all over the place with their statements and repeat themselves a lot in their program, while the Nazis have barely any program to speak of.

image/jpeg

badlogic, 2 days ago

And as always, if you find this kind of "work" entertaining/imformative, and if you have disposable income, consider supporting our charity:

https://cards-for-ukraine.at

We are a zero-overhead charity. Ever donation cent goes towards food vouchers for 🇺🇦 families (mostly women and their kids) in 🇦🇹. We pay everything else and do the labor for free.

All invoices, payment confirmations etc. here:
https://drive.google.com/drive/folders/1PxOL8A44bIRU1Hdoq87_2iXSLNmnMXQr

2 years and still going strong. Over €250k in vouchers delivered.

badlogic, 2 days ago

@Aaron I'm first taking the statement text (which is usually just a single sentence) and embed it into a high dimensional space (~1500 dimensions) woth OpenAI's text embedding model.

This vector basically encodes the semantics of the statement text.

Embedding vectors of statement texts with similar or related semantics, will end up in the same area of that high dimensional space.

The closer two vectors are in that space, the more semantically similar they are.

badlogic, 2 days ago

@Aaron now, we can obviously not draw 1500 dimensions. We can thus apply a method called projection (which is also a form of embedding). In this case, I use a method call UMAP.

Simplified: it takes the high dimensional vectors. For each vector it tries to find a few closest vectors.

It then assigns 2 dimensional coordinates to those vectors in such a way, that their distances are similar in 2D to what they are in the high dimensional space.

badlogic, 2 days ago

@Aaron This way, the 2D projection retains the "neighbourhoods" that exist in the richer high dimensional space.

The end result is, that semantically similar or related points end up in the same area in 2D as well, which is nice for visualization purposes. We can clearly see clusters of points for different topics.

Neither the original neighbourhoods in the high dimensional space nor the 2D neighbourhoods are perfect of course. But it's plenty good enough for this purpose.

badlogic, 2 days ago

@Aaron UMAP has kind of become the standard method when you want to project high dimensional vectors to 2 or 3 dimensions for visualization.

For embedding text (single words, sentences, paragraphs, etc.) you have more options. I was lazy, so I used OpenAI's embedding model through their API.

A popular alternative is sentence BERT:
https://sbert.net/

For more details on text embeddings, start with word embeddings, then move up to sentence embeddings.

https://en.wikipedia.org/wiki/Word_embedding

badlogic, 2 days ago

@Aaron TL;DR: it's not just word frequencies.

Word embeddings are vectors that encode possible semantics of a single word.

Sentence embeddings are vectors that capture the semantics of an entire sentence. It starts with word embeddings for each word in the sentence, which are the distilled, to resolve ambiguities/references between words, ending up with a single vector that stores all the "meaning" in the sentence.

badlogic, 2 days ago

@Aaron enjoy! Just get too hung up on understanding the nitty gritty details. For projects like the above, all you need to understand is:

• text goes in, vector comes out
• similar texts have vectors that will be close to each other

The rest is just measuring distances or angles between those vectors.

badlogic, 21 days ago

@sinbad Sure hope you have more luck getting Epic to even recognize the issue. We had to break every user's project...

badlogic, 21 days ago

@sinbad Hope!

badlogic, 22 days ago

@accidentlyAnton honestly no idea. Had to upgrade. Works. Most of the time.

badlogic, 25 days ago

@shtrom @wallabag @mastoreaderio neat! Happy it has some use.

badlogic, 26 days ago

@bartholin LOD == level of detail.

As a 3D object gets further away from the camera, it gets smaller on screen, e.g. only covers 3% of the screen.

In that case, it makes no sense to render it at full fidelity. So you use a simplified mesh. You won't notice the difference at that size.

You ship a model at various such "resolutions". Each is one level of detail. You select the level of detail based on the distance to the camera. Helps improve performance, as you do less computations.

badlogic, 26 days ago

@bartholin in the screenshot above, you see the LODs going from 95k "polys" to 75k "polys" to 55k "polys".

That's ... a silly step size. Usually, it's more like an order of magnitude reduction per LOD level. This indicates that Kadeim, the company offering AI generated 3D models, has no idea what they are doing (or preys on customers who are not knowledgable)

badlogic, 26 days ago

@bartholin here's a visual example of what LOD levels usually look like.

badlogic, 26 days ago

@Robin_Van_Ee good observation!

badlogic, 26 days ago

@godotengine