Artificial intelligence systems like ChatGPT could soon run out of what keeps making them smarter — the tens of trillions of words people have written and shared online.
A new study released Thursday by research group Epoch AI projects that tech companies will exhaust the supply of publicly available training data for AI language models by roughly the turn of the decade – sometime between 2026 and 2032.
Comparing it to a “literal gold rush” that depletes finite natural resources, Tamay Besiroglu, an author of the study, said the AI field might face challenges in maintaining its current pace of progress once it drains the reserves of human-generated writing.
In the short term, tech companies like ChatGPT-maker OpenAI and Google are racing to secure and sometimes pay for high-quality data sources to train their AI large language models – for instance, by signing deals to tap into the steady flow of sentences coming out of Reddit forums and news media outlets.
In the longer term, there won’t be enough new blogs, news articles and social media commentary to sustain the current trajectory of AI development, putting pressure on companies to tap into sensitive data now considered private — such as emails or text messages — or relying on less-reliable “synthetic data” spit out by the chatbots themselves.
I mean, one way to improve an AI is by increasing the size of the training corpus. That’s…not the only way. I think that “just throw more data at it” is probably good low-hanging fruit as long as that data’s there and not hard to get ahold of. But you can filter out bad data in your existing training set, or build more-elaborate systems that make more-effective use of the data there.
As you increase your training set size, you’re also increasing the cost of running training, which isn’t necessarily desirable.
I am confident that there is probably more-than-enough text and (and images, and audio, and video) in the world to train something that operates at the level of a human. I know that because humans do it.
But humans have literally millions of years of social and cultural development in our brains. You think babies are a blank slate, but they are hardwired to be social and learn whatever their parents teach them.
You don’t have to teach a baby what a human face is. You don’t have to do anything but talk to a baby, or talk near them, to teach them language. Their brain automatically makes connections and creates new sentences.
Humans are experts at inference (moving from specific examples to general rules). That’s why a child will learn what a “doggy” is, and then refer to a cow as a “big doggy”. They’ve already internalized that a dog is an animal that walks on four legs. If you tell them ‘It’s a cow. The cow says “moo”.’ they will quickly get the idea.
I think the main problem with AI is we try to train it in a few months. Humans are experts at making connections and we expect it takes decades to train them.
The primary problem with LLMs as they stand today isn’t a lack of time to train them. It’s is that they have a really primitive structure. They’re like only part of what’s in your head.
What an LLM can do, what generative AIs are doing, is basically take a very large amount of data, where each instance of data is annotated with something, and find commonalities between between them, which lets them learn to associate the annotation with characteristics of that data. And they can synthesize data based on those annotations.
That’s quite useful for a number of things. It lets us do things that we couldn’t do before. We can do really potent voice and image synthesis, upscale images, stuff like that.
But it’s really simple and mechanical.
But what’s going to happen here isn’t that we take that approach and then throw more data or more hardware at it and it suddenly becomes a human.
And there’s a lot of stuff that that doesn’t do.
It doesn’t permit for learning to perform tasks, say. It can’t develop higher-level concepts of things than associations with the annotations. You can give Stable Diffusion 1000 years of continuous running, and it will never start sending you messages about living inside your computer.
So the existing things that people are calling “generative AI” – those aren’t going to be a human because they just don’t have the infrastructure to do it. Some of the things they do might be part of something human-like, but they’re missing a lot of parts.