Communism doesn't belong only to those who can afford to pay for it. We remain committed to our mission of offering proletarian education without discrimination.

Unfortunately no ETA yet on our RAG files but they have also worked on the MIA alongside the PW archive they are making: github.com/percy-raskov...

But it's also going to be FOSS and usable by anyone for any purpose.

I don't know about you but this seems much more beneficial to the workers than another SaaS you have to pay for to regurgitate only things the service deems acceptable.

This means you will be able to make RAG models out of local models too without having to go through us every time. Or fine-tune them with our content. Or anything you like.

Sure, this is less "sexy" than a neat react.js webapp with animations and a ready-made commercial model.

We ourselves are making ProleWiki available for RAG.

We are working with a dev who has received our entire wiki corpus, and will make it available open source once it's parsed and cleaned up.

One last thing on their AI - since it communicates with API, there is also not only the model's bias (both GPT and Deepseek) but also a lighter system prompt injected by both. APIs are generally less censored than the web interfaces, but still have *some* system prompt in it.

Is this really what a "major breakthrough" looks like (to reuse their super lengthy self-congratulatory press release), or is this yet another profit-driven SaaS?

But they did not publish this documentation anywhere to help other comrades build their own systems. Instead, they want to sell paid tiers on a cloud service that talks to OpenAI.

Guess if you can't afford it you can't be a communist.

The takeaway is the WSWS is very cryptic about their technical implementation and very protective of their system. You don't just feed the AI the blog links, at least, if you do RAG correctly. You clean it in a way the AI can parse it - whether they did that we can't say.

As for how we found the info - it's all there in the HTML code since it needs to call JS and JSON files to work. Nothing was "reverse-engineered" (which their ToS says is verbotten 😡)

Disclaimer, it is possible our investigation is wrong in some way. We looked at the JSON and JS files they call in the HTML of their homepage, so the information is there - whether it's used in the way it's presented is something else.

RAG is legitimate, and it's also the least-effort service to make LLM models more specific. You start with a fully-trained model and then just tell it "look at these pages first before answering".

In this method the LLM looks up a corpus of data before answering the user query. It can indeed look at the WSWS whopping 250k magazine articles of the past 25 years (do trots do nothing but write all day long?) and makes an answer on it.

This gives some indication as to how their AI works. First, your chats are sent to OpenAI. So maybe don't discuss things too openly with Socialism AI.

Second, it seems to be using RAG, or retrieval augmented generation.

Our investigation shows they call two models: GPT4.1 and Deepseek.
GPT is OpenAI and is a closed-source model. So they won't have downloaded it to fine-tune train it - not possible.

Instead you have to communicate with GPT through OpenAI's API endpoint.

They're also very protective of how exactly their AI works, how it was made, etc, but we looked into it. Let's untangle it.

Let's start the privacy policy. Lots of mentions for paid tiers and copyright retention.

They claim everything "Socialism AI" outputs belong solely to WSWS.

The trotskist revolution will not be televised - because the workers won't have paid the broadcasting fees.