Tag Archives: Machine learning

Will GPT models choke on their own exhaust?

Until about now, most of the text online was written by humans. But this text has been used to train GPT3(.5) and GPT4, and these have popped up as writing assistants in our editing tools. So more and more of the text will be written by large language models (LLMs). Where does it all lead? What will happen to GPT-{n} once LLMs contribute most of the language found online?

And it’s not just text. If you train a music model on Mozart, you can expect output that’s a bit like Mozart but without the sparkle – let’s call it ‘Salieri’. And if Salieri now trains the next generation, and so on, what will the fifth or sixth generation sound like?

In our latest paper, we show that using model-generated content in training causes irreversible defects. The tails of the original content distribution disappear. Within a few generations, text becomes garbage, as Gaussian distributions converge and may even become delta functions. We call this effect model collapse.

Just as we’ve strewn the oceans with plastic trash and filled the atmosphere with carbon dioxide, so we’re about to fill the Internet with blah. This will make it harder to train newer models by scraping the web, giving an advantage to firms which already did that, or which control access to human interfaces at scale. Indeed, we already see AI startups hammering the Internet Archive for training data.

After we published this paper, we noticed that Ted Chiang had already commented on the effect in February, noting that ChatGPT is like a blurry jpeg of all the text on the Internet, and that copies of copies get worse. In our paper we work through the math, explain the effect in detail, and show that it is universal.

This does not mean that LLMs have no uses. As one example, we originally called the effect model dementia, but decided to rename it after objections from a colleague whose father had suffered dementia. We couldn’t think of a replacement until we asked Bard, which suggested five titles, of which we went for The Curse of Recursion.

So there we have it. LLMs are like fire – a useful tool, but one that pollutes the environment. How will we cope with it?

Data ordering attacks

Most deep neural networks are trained by stochastic gradient descent. Now “stochastic” is a fancy Greek word for “random”; it means that the training data are fed into the model in random order.

So what happens if the bad guys can cause the order to be not random? You guessed it – all bets are off. Suppose for example a company or a country wanted to have a credit-scoring system that’s secretly sexist, but still be able to pretend that its training was actually fair. Well, they could assemble a set of financial data that was representative of the whole population, but start the model’s training on ten rich men and ten poor women drawn from that set – then let initialisation bias do the rest of the work.

Does this generalise? Indeed it does. Previously, people had assumed that in order to poison a model or introduce backdoors, you needed to add adversarial samples to the training data. Our latest paper shows that’s not necessary at all. If an adversary can manipulate the order in which batches of training data are presented to the model, they can undermine both its integrity (by poisoning it) and its availability (by causing training to be less effective, or take longer). This is quite general across models that use stochastic gradient descent.

This work helps remind us that computer systems with DNN components are still computer systems, and vulnerable to a wide range of well-known attacks. A lesson that cryptographers have learned repeatedly in the past is that if you rely on random numbers, they had better actually be random (remember preplay attacks) and you’d better not let an adversary anywhere near the pipeline that generates them (remember injection attacks). It’s time for the machine-learning community to carefully examine their assumptions about randomness.