One simple, but incredible thing that now can be done, and six months ago was not really as easy to do as it is now, it is just run your LLM, and start generati

One simple, but incredible thing that now can be done, and six months ago was not really as easy to do as it is now, it is just run your LLM, and start generating new content. You can now use your LLM to generate the data you need to train bigger, better models. This is the key for future better LLM, you can generate ashtoningly amounts of data, as fast as your pipeline can generate the text and feed it to the training, You could even automatically curate data somehow (using experts systems, manually fed by humans, double checking "facts" and "known knowledge" from garbage), and ultimately you'll be using the same original AI to generate a better AI, then you iterate. At each cycle, the original LLM becames better. And you won't even need Internet at some point, you just have to make relatively sure the first iterations are feeding the LLM with relatively "good" data. It is not necessary that the data, the text, be true, it just needs to resemble what you'd have found in the open internet, maybe one good comment, and 40 other comments with somehow flawed data. How fast can you do this? We're probably about to see it. How better do you think a LLM can get by increasing the available data for training them? And if you could accurately "summarize" the text, the data somehow - using LLMs or other systems - then, how faster could you train a newer, better LLM? Just by decreasing by little magnitudes the amount of text for training without losing the significance of the information to be encoded in the LLM, you'll probably be able to get the training way faster than before. And what happens if you can somehow "stack" several LLMs? re-feeding one output to the next LLM to make it process it, several LLMs running ones behind others, the first ones the "cheaper", then the "expensive", a lot more powerful. And use that stack to produce even "deeper" text, encoding even more information in simpler, faster to process, text. And if you could somehow encode text in a WAY simpler representation format? Using that format to train those advanced LLMs (you'd probably then need a "translator LLM" to re-compose the simplified text to actual readable by human text). You think this is scifi? nope, just use ol' good stenography, and you're good. Yes, using a representation format as stenography would required advanced LLM capable of generating complex original never seen images from text, fully capable translate text to images and images to text..wait we already have those... Some amazing stuff is coming, if the makers care to make those advances public information.