Symbol’s Greetings: Onboarding to Active Inference across backgrounds & languages

All right, welcome to Active Inference Stream number 8.1. It's December 15th, 2024. I'm going to start with a GitHub push to the GitHub repo. Active Inference Institute slash start. While we have, meanwhile, cursor, current version, 0.43, translating, curricula, the background, visualizing, learning paths, having our agenda up. So thanks for joining. If you're watching live, definitely write some interesting things in the chat and I'll look to re-enter it into the stream. All right, here we are. Symbols, greetings. Good evening, Jeff. Again, this is in GitHub repository Active Inference Institute slash start. The idea here was to build on some prior perplexity and LLM based methods that we've worked on in different projects earlier in the year. And kind of continue with this all by all theme across domains and backgrounds intersecting with onboarding and curricula for Active Inference. This is more than just a translation tool. It's a framework for understanding and optimizing how knowledge flows through different perspectives. Modular treatment of topics and perspectives enables flexible application. Okay. Okay. So, first, where are we headed to in terms of the output? Then, let's look at the scripts that are using perplexity even now to go through it. So, the written curricula are in English and then they're being translated in chunks to other language. So, here's a synthetic biologist background introduction curriculum to Active Inference translated into Japanese. And this is more than 1400 lines long. That's what it looks like in cursor. Let's reload the repo. It's still writing this very large initial push. We'll look at what the markdowns look like in a few minutes. That's the translated curriculum. The written curricula, there's two sets of audiences that we'll trace back to that it's drawing upon. The first with just their names are individual specific people or different audience profile. We'll get to how that's specified like a high school student and a middle school student. So, these are different audiences or different synthetic knowledge domains that are constructed. We'll, again, look into them like cooking. So, this is a complete curriculum implicitly in English for somebody in the cooking industry. Conduct workshops where participants develop new recipes using Active Inference algorithms. Predicting food safety risks. Those are the different domains. So, this pipeline that we're about to go through uses perplexity.ai, large language model, LLM interface, API, to do research on entities, specific or groups of entities, and different synthetic domains. We'll look at how they're defined. Pool those researches. Those are done in step one. Step two, write introductions. Again, going to call perplexity to write out those curricula in chunks in English. Three, does visualization of a few different kinds on those generated curricula. Four translates into languages. Right now, these languages are uncommented out. There's more languages you can just comment or uncomment. I will show my perplexity balance at the beginning of the stream, and then we'll see how much it changes on during the stream. Okay. Okay. $67.68 in the perplexity bank. Using the small here on their model card page. I'm using the LLAMA 3.1 Sonar small, 128 online. This is the same context length as these more larger parameter models. Just from short testing though, it's faster and it's cheaper. So, 67, 68. Okay. All right. So, where we're going is this. Now, let's look at the visualizations on those curricula from step three. Then, look at the scripts. Let's look at the scripts. Okay. First, metrics. So, this is for each of those. It's just translating chunks in the background. That can be multi-threaded, but it can also cause more API trouble than it's worth. Okay. So, this is for each. This is a comma separated file. Open it in LibreOffice. This is some word count information in the sections and the paragraphs. Descriptive statistics on curriculum…