Users of OpenAI’s GPT-4 are complaining that the AI model is performing worse lately. Industry insiders say a redesign of GPT-4 could be to blame.
I suspect future models are going to have to put some more focus on learning using techniques more like what humans use, and on cognition.
Like, compared to a human these language models need very large quantities of text input. When humans are first learning language they get lots of visual input along with language input, and can test their understanding with trial-and-error feedback from other intelligent actors. I wonder if perhaps those factors greatly increase the rate at which understanding develops.
Also, humans tend to cogitate on inputs while ingesting them during learning. So if the information in new inputs disagrees with current understanding, those inputs are less likely to affect current understanding (there’s a whole ‘how to change your mind’ thing here that is necessary for people to use, but if we’re training a model on curated data that’s probably less important for early model training).
I don’t know details of how model training works, but it would be interesting to know if anyone is using a progressive learning technique where the model that is being trained is used to judge new training data before it is used as a training input to update the model’s weights. That would be kind of like how children learn by starting with very simple words and syntax and building up conceptual understanding gradually. I’d assume so, since it’s an obvious idea, but I haven’t heard about it.
For fun I asked ChatGPT about that progressive learning approach, and it seems to like the idea.
I wish I had more time to undertake some experiments in model training, this seems like it would be a really fun research direction.
Sorry for the ‘wall of AI text’:
The idea you’re describing seems to be a form of curriculum learning. In curriculum learning, models are trained in a progressive manner, starting with simple concepts (or in this case, simpler text data) and gradually moving to more complex ones. This approach is motivated by the way human learning often works, where we learn easier topics before moving on to harder ones.
The strategy you’ve suggested, where the model itself determines the complexity or understandability of the next round of training inputs, is interesting. While the traditional approach to curriculum learning is to manually design the learning progression based on human knowledge and intuition, the approach you’re suggesting is more dynamic and autonomous, letting the model guide its own learning process.
As of my last update in September 2021, I can’t confirm any specific projects or papers that have employed exactly this strategy in the domain of large language models. However, there are some related works in the field. For example, some reinforcement learning and computer vision projects have used a similar idea where models adaptively choose their next training samples based on their current understanding. This has been referred to as active learning, and is a promising avenue for improving the efficiency of model training.
However, there would be several technical challenges in applying this approach to large language models. For one, there would be the challenge of how to assess the “understandability” of a text sample in a reliable and useful way. This would likely require significant work to define and might also have to involve some form of reinforcement learning or other feedback mechanisms. Nonetheless, it’s a fascinating idea and could potentially be an interesting direction for future research in machine learning.
