And regarding the data question the other commenters are asking — you scrape everything you can (oh look I used an em dash, wanna run me through the cover-your-ass Pangram?). Anna's archive, The Pile, the various Huggingface data sets and aggregates, Common Crawl. You pay proxy farms like Bright Data to run residential and mobile gray area proxies and VPN and CloudFlare bypasses to do more scraping. I see a lot of HN users up in arms on the front page thread about LG TVs having VPN SDKs within them etc. And then five minutes later they will go back to their frontier LLMs to print their pay cheque. Fucking hypocrites of the highest order. It's oh noes how bad this tech is in my house but we will happily benefit from it.
Then you have a data cleaning team deduplicate and clean and annotate the data (with or without help of more AI)
Certain RL specific datasets for supervised fine-tuning and RLHF like coding and git commits and chat needs to be curated by hand depending on your use case.
The level of discourse on AI has fallen tremendously on HN if 5 years into the AI revolution people are still wondering why datasets aren't being released. They aren't being released because they are a fucking snapshot of the internet for fuck's sake. There are a few "sanctuary" nations where AI data scraping is somewhat legally unenforced but the United States is not one of them so stop asking why data isn't released on a Bay Area website. Use your head for once. Too many React and YouTube influencers and the brains have been rotted.
Out of all the so called "AI engineers" here pontificating about "alignment" and "AI safety" and "Recursive Self Improvement", I wonder how many can even formulate or describe what an ELBO is. I wonder how many product managers here yapping about "recalibrating their priors" actually know what a prior is. I truly wonder why LLMs seem so magical to people when it's only a few steps removed from the same neural networks people have been using since 2015, at least architecturally (except scaled up by a few magnitudes).
Get your head out of Roko's Basilisk's agentic ass and maybe actually read the technical reports and papers for once.
(One important paper post Attention is All You Need is the DeepSeek paper where they used RL to bootstrap the "thinking" chain of thought token chains. IIRC it's the DeepSeek R2 paper. That's one of the most important papers for understanding LLMs beyond basic ML neural networks. If you need a quick way to get up to speed, read that one).