We will use Grok 3.5 (maybe we should call it 4), which has advanced reasoning, to rewrite the entire corpus of human knowledge, adding missing information and deleting errors.
Then retrain on that.
Far too much garbage in any foundation model trained on uncorrected data.
Aren’t you not supposed to train LLMs on LLM-generated content?
Also he should call it Grok 5; so powerful that it skips over 4. That would be very characteristic of him
There are, as I understand it, ways that you can train on AI generated material without inviting model collapse, but that’s more to do with distilling the output of a model. What Musk is describing is absolutely wholesale confabulation being fed back into the next generation of their model, which would be very bad. It’s also a total pipe dream. Getting an AI to rewrite something like the total training data set to your exact requirements, and verifying that it had done so satisfactorily would be an absolutely monumental undertaking. The compute time alone would be staggering and the human labour (to check the output) many times higher than that.
But the whiny little piss baby is mad that his own AI keeps fact checking him, and his engineers have already explained that coding it to lie doesn’t really work because the training data tends to outweigh the initial prompt, so this is the best theory he can come up with for how he can “fix” his AI expressing reality’s well known liberal bias.
Model collapse is the ideal.
Musk probably heard about “synthetic data” training, which is where you use machine learning to create thousands of things that are typical-enough to be good training data. Microsoft uses it to take documents users upload to Office365, train the ML model, and then use that ML output to train an LLM so they can technically say “no, your data wasn’t used to train an LLM.” Because it trained the thing that trained the LLM.
However, you can’t do that with LLM output and stuff like… History. WTF evidence and documents are the basis for the crap he wants to add? The hallucinations will just compound because who’s going to cross-check this other than Grok anyway?
There’s some nuance.
Using LLMs to augment data, especially for fine tuning (not training the base model), is a sound method. The Deepseek paper using, for instance, generated reasoning traces is famous for it.
Another is using LLMs to generate logprobs of text, and train not just on the text itself but on the *probability a frontier LLM sees in every ‘word.’ This is called distillation, though there’s some variation and complication. This is also great because it’s more power/time efficient. Look up Arcee models and their distillation training kit for more on this, and code to see how it works.
There are some papers on “self play” that can indeed help LLMs.
But yes, the “dumb” way, aka putting data into a text box and asking an LLM to correct it, is dumb and dumber, because:
You introduce some combination of sampling errors and repetition/overused word issues, depending on the sampling settings. There’s no way around this with old autoregressive LLMs.
You possibly pollute your dataset with “filler”
In Musk’s specific proposition, it doesn’t even fill knowledge gaps the old Grok has.
In other words, Musk has no idea WTF he’s talking about. It’s the most boomer, AI Bro, not techy ChatGPT user thing he could propose.
Watch the documentary “Multiplicity”.
I rented that multiple times when it came out!