[ericskiff]

We seem to be at an interesting feedback loop point for ML models.

We have textual and visual models strong enough to look at a picture and describe it, or to read a body of text and categorize and label it well.

That work can then be used to feed the training of other models, whether it's larger models that can make use of an ever-growing well labelled corpus, or smaller and more efficient models trained on sets that were previously too niche or cost prohibitive to individually label.

With this coming out of Microsoft for copilot/Bing and the strength of recent smaller models which have had the benefit of GPT4 and other larger LLMs to assist with training, we appear to be at an inflection point of training quality at the same time as training compute is being massively scaled.

[sroussey]

Yes, but…

When GPT4 gets it wrong… all the other LLMs do as well. The error gets reinforced when you ask a quorum of LLMs a question and a majority can trace the same wrong answer back to GPT4.

And you thought getting something out of googles index was hard.

Try removing it from the weights of a bunch of LLMs!

[bigyikes]

Sure, doing surgery on the weights is not feasible, but doing surgery on the training data is possible (and is done for certain benchmarks).

That's where it gets interesting. If you tell the LLM what the bad data is, it could examine its own training set and fix the mistakes. Get enough LLMs with diverse enough training sets, and they will correct each other.

If models are stochastic parrots, you might expect training data to devolve into corruption via a negative feedback loop, but the opposite might happen if models are actually intelligent.