LLMs Aren’t “Trained on the Internet” Anymore

I often see a misconception when people try to reason about the capability of LLMs, and in particular how much future improvement to expect.

It’s frequently said that that LLMs are “trained on the internet,” and so they’ll always be bad at producing content that is rare on the web. A recent example from Hacker News:

LLMs are mostly trained on Internet posts, so they are Internet simulators.

Wikipedia also frames LLM training this way:

LLMs are trained on increasingly large corpora of text largely scraped from the web.

Even former Apple SVP of Engineering – and famously sharp computer scientist – Bertrand Serlet explains them thus, in his recent video Why AI Works:

[An LLM has] been trained on the entire internet, so it has packed all this knowledge that’s in the internet, but also all the reasoning patterns that can be found [there.]

Admittedly, this used to be true! And is still mostly true. But it’s increasingly becoming less true.

Because it’s becoming less true, the idea of an “internet simulator” is not a useful way to predict how GPT-5 and beyond will work. New models are already exceeding this definition, and the shift is only beginning.

The Data Wall

Way back in 2020, OpenAI’s GPT-3 paper described their training dataset in some detail. GPT-3 was, basically, trained on the internet.

By 2022, LLMs were being trained to follow instructions using custom human feedback. Since then, frontier model labs like OpenAI have gotten very cagey about what they’re adding to their training datasets. We don’t know what GPT-4o was trained on, let alone Sora or GPT-5.

But we do know it’s not just the internet.

Recently, LLM trainers have hit a “data wall”. More data generally improves model training, but OpenAI already has roughly all the data on the web – even inconvenient-to-access data like Youtube video transcriptions. So other than training ever-larger models on the same internet data, how can they make better LLMs?

The answer, for labs with money, is acquiring and creating non-public data. At first, their focus was making existing training data more useful, or adding existing non-public data to the training pool. For example:

1. Annotation and filtering: Researchers have been creating annotations for their training data for years. Good annotations let them focus LLM training on the highest value data, which allows for better (or smaller) models.
2. RLHF: Labs also have humans rate model outputs. They use this data to fine-tune the models, and encourage useful behaviours like instruction-following and refusing to say naughty things.
3. Usage data: ChatGPT is said to generate on the order of 10 billion tokens of data per day – even before they opened their more compelling GPT-4o model to free users.
4. Data acquisition: While a lot of humanity’s data is on the internet, much of it is not. Emails, chat logs, proprietary manuals and procedures, JIRA tickets, phone recordings, internal reports, and contracts are all useful data. While examples exist on the internet, model trainers can cut deals to add more of these to their training data.

So that’s all helpful.

But none of these techniques are a complete solution to a famous weakness of current models: the “LLMs suck at producing outputs that don’t look like existing data” problem.

Here are some of the many things that LLMs are bad at doing, in part because there isn’t a lot of text online that demonstrates them:

1. Expressing doubt or uncertainty in one’s answer
2. Having long conversations without repeating phrasings or looping
3. Making high-level plans for LLM agents to pursue
4. Reasoning like a Principal Engineer about a massive legacy codebase
5. Reliably following extremely long or complex prompts

While improved architectures and more parameters might help with these limitations, you can bet your butt that OpenAI, Meta, Google, and/or Microsoft are paying big money to fill some of these gaps in a simpler way: creating novel examples to train on.

LLMs are now being trained on custom data

A recent example of the rise of custom data is Microsoft’s Phi-3 Technical Report, published in April. phi-3-mini is only 3.8 billion parameters – a waif in LLM terms – but claims performance competitive with the impressive but much-heavier Mixtral model. The paper credits some of this improvement to including high-quality synthetic data, generated by larger LLMs, in the training data. Synthetic data allows them to fill gaps in the internet-sourced data, and improves model performance for a given size.

Now, synthetic data is a hot topic in LLM research. It’s not yet clear how far you can go with training LLMs on their own output before things go horribly wrong, like a giant neural snake eating its own tail.

At the minimum, though, synthetic data will help fill the kind of gaps that arise simply from LLMs behaving like “internet simulators.” For example, if your model is hallucinating because you don’t have enough training examples of people expressing uncertainty, or biased because it has unrepresentative data, then generate some better examples!

Still, creating great synthetic data with LLMs is a challenging problem, and will have its limits. That’s why there’s one last huge source of non-internet data inbound: Humans.

How much data can you create for $1B/yr?

It turns out that if you pay people money, they’re willing to make data for you. Scale.ai, which bills itself as “the data foundry for AI,” operates a service where labs pay humans to do just this. Reportedly, AI companies are paying more than $1B a year for Scale’s services already. While some of this is for annotation and ratings on data that came from the web or LLMs, they also create new training data whole-hog:

The company has shifted its focus to more highly-specialized workers such as Ph.D-level academics, lawyers, accountants, poets, authors, and those fluent in specific languages. These workers, who help train and test models for companies from OpenAI and Cohere to Anthropic and Google, also work through a third-party, often another Scale subsidiary called Outlier, but are paid higher hourly wages.

Not only can companies like OpenAI pay to have professionals create new, good data that fills holes in the internet-sourced data, they then keep that data to train subsequent models. A dataset like “50,000 examples of Ph.Ds expressing thoughtful uncertainty when asked questions they don’t know the answer to” could be worth a lot more than it costs to produce.

So yes, LLMs were originally trained on the internet. Yes, we can understand a lot of their early weaknesses as stemming from the foibles of whatever crap people post on the web.

However. As the size and impact of custom training data increases, we should expect LLMs to markedly exceed “internet simulation”. In particular, they’ll continue to get better at things that aren’t on the internet, but can be demonstrated in $1B+ of custom data creation.

That is to say: this train will keep rollin’ for a while yet.