My friend says that genAI would become more human like, and perhaps even smarter than humans if it were simply trained on more and more data. I say that this would overtrain the models, and we would perhaps lose the generative aspect.

I looked up on the internet, and it seems overtraining is actually a problem... but could it be said that there is a limitation to the amount of data we could feed to a model which would kill it's generativeness off completely?

  • 1
    $\begingroup$ theoretically speaking, your friend is correct, with theoretically meaning that you have a powerful enough universal function approximators, but for neural networks, it's been observed a phenomenon called double descent, where increasing the dataset size in certain cases, actually hurts the overall performance arxiv.org/pdf/2303.14151.pdf $\endgroup$
    – Alberto
    Jul 24, 2023 at 21:05
  • $\begingroup$ also, overtraining on the internet usually means training for more epochs, not on more data, leading to overfitting $\endgroup$
    – Alberto
    Jul 24, 2023 at 21:06

1 Answer 1


I'll try to deconstruct your question and give you the most informative answer:

Is there a limitation to the amount of data that a genAI model could be trained upon?

In the way that this question is framed, the answer is clearly no. You can continue training on data indefinitely, given that you have the data and monetary resources to do so.

However, I do not think that this is the answer you are looking for. Let me attempt to rephrase the question to align with what you actually want to know:

Is there a threshold/point after which training a genAI model on more data is no longer beneficial towards the quality of the output?

This is a much more intricate question.

  1. There is the very well-known problem of overfitting models on training data. Small models are generally trained on a training set, which is iterated over many times during training. If the training loss decreases, but the model does not generalize well (the validation loss increases), then we typically (not always) say that your model is overfitting. Training more would, generally, not be beneficiary for your model.
  2. Large LLMs are, however, not trained by iterating over a dataset multiple times. As far as I know, the training sets are so enormous that each sample is only seen once. It is highly unlikely that training on more data will cause the model to overfit.
  3. The post OpenAI made on GPT 4]1 state certain benchmarks over time. They show that these models do seem to slowly converge to a state where the benchmarks do not improve anymore. This can allude to two things, namely (1) the benchmark is completed optimally or (2) the model no longer learns. The plots seem to suggest the former, further reinforcing the second point I made.
  4. However, many people have suggested that ChatGPT is getting worse, while being further trained. Whether this is true, I will not judge. Nevertheless, these models can get worse over time if the data on which it is trained is not of high quality. High-quality data has been shown to significantly matter in the rate of convergence of these models. If you train the model on lower quality data than it has been previously trained upon, your model can get worse.
  5. What is the 'final state/optimal state' of these models? Will they be 'smarter than human' (whatever that may mean)? These models are restricted to the data that we feed it. They are able to construe sentences together based on the data is has been trained on. Whether that means they can be smarter than humans according to your definitions of intelligence is up to philosophers to debate.
  • 2
    $\begingroup$ Regarding "You can continue training on data indefinitely, given that you have the monetary resources to do so" - this is only true if the data exists, or you have truly large enough resources to create it. With the largest LLMs we may soon reach a "peak usable data" limit for text resources, where there simply is not enough text available to scale in the same way naively - lesswrong.com/posts/Couhhp4pPHbbhJ2Mg/… $\endgroup$ Jul 25, 2023 at 21:31
  • $\begingroup$ Adapted my answer to reflect your comment. I was thinking more along the lines of 'you can simply iterate over your existing dataset' which would still constitute 'training'. But the framing of the question is ambiguous, depending on how you interpret 'limitation on the amount of data'. Is that new data, or just data in general. I interpreted it as just data in general, which you can keep training on given monetary resources. However, I can see how you can interpret it as new data, given the specific mention of 'amount'. $\endgroup$ Jul 26, 2023 at 9:18
  • $\begingroup$ There is a proposed optimal data-to-paramaters ratio (20:1) for LLMs. Continuing to train on repeated data may not be effective. Separately to this question there is also the issue of "poisoning the well" in that the LLM owners may soon not be able to collect data knowing it was generated by humans. There is research showing LLMs re-training on their own output is not a good idea, either. $\endgroup$ Jul 26, 2023 at 17:22

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .