The world's most-powerful AI model suddenly got 'lazier' and 'dumber.' A radical redesign of OpenAI's GPT-4 could be behind the decline in performance.

L4sBot@lemmy.worldmod to Technology@lemmy.world – 274 points – The world's most-powerful AI model suddenly got 'lazier' and 'dumber.' A radical redesign of OpenAI's GPT-4 could be behind the decline in performance.
businessinsider.com

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.

90

You are viewing a single comment
View all comments

The model has become inbred because it’s now impossible to scrape the web without AI content getting ingested, which is full of “hallucinations” and other weird artifacts. The last opportunity to get “uncontaminated” training data was sometime in mid 2022.

Not to say that it’s causing this particular problem, but this issue will emerge eventually. Garbage in = garbage out. Eventually GPT-19 will grow a mighty Habsburg chin.

Maybe not yet, but...

  • Spez will turn Reddit into a bot farm and sell this as training data
  • Musk turns Twitter into a bigoted cesspool and will sell this as training data, which will subsequently be flagged for low quality (also: a botfarm)
  • Threads is a corporate ad dashboard (and we already know how easy it is to GPT copy) and Zuck will sell this as training data
  • Facebook is either dead or only good for boomers and Poles
  • blogs are dead
  • Fediverse is out there waiting to be scraped but possibly too small to sustain a big model

We'te getting there, hopefully.

..is Facebook popular with Polish people? Or was this a weird polish joke I don't get?

Very. Twitter never took off among general population (only politicians, journalists, botfarms and people who troll politicians and journalists), tiktok is for kids, Instagram is popular but again, rather among influencers and people who need to show off pictures not as a default SM app. I don't really know where did Americans and west Europeans move from Facebook.

That hasn't happened yet. Most likely they quantized GPT-4 more. It's still based on the same training data.

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.

7 more...