The promise and dangers of artificial data

Is it possible for an AI to be trained on data generated by another AI? It may sound like a harebrained idea. But it’s one that’s been around for a long time – and as new, real things are harder to come by, they’ve gotten better.

Anthropic used synthetic data to train one of its features, Claude 3.5 Sonnet. The meta has been successfully updated Rule 3.1 examples using AI generated data. And OpenAI is said to be getting training materials from o1his “imaginary” model, of what is to come Orion.

But why does AI need data in the first place – and what kind of data needed? And this data actually instead of data processing?

The need for explanations

AI machines are computational machines. Trained on many patterns, they learn the patterns in these patterns to make predictions, such as the “to whom” in email often leads to “it may be difficult.”

Annotations, which often describe the meaning or components of the data entered by these systems, are a very important part of these models. They act as guides, “teaching” the model to distinguish objects, places, and ideas.

Think about the type of pictures that are shown in many pictures of kitchens with the word “kitchen.” While teaching, the model will begin to make associations between “the kitchen” and everything values of kitchens (such as refrigerators and countertops). After training, when presented with a picture of a kitchen that was not included in the first examples, the model should recognize that it is. (Of course, if the pictures of the kitchens were labeled “cows,” they would identify them as cows, which emphasizes the importance of good words.)

The desire for AI and the need to provide documented information about its development has created a market for explanatory materials. Dimension Market Research comparison that it is worth $838.2 million today – and will be worth $10.34 billion in the next 10 years. Although there is no exact estimate of the number of people working in writing jobs, 2022 paper puts the number in “millions”.

Companies large and small rely on employees hired by data analytics companies to create AI training class labels. Some of these jobs pay well, especially if the writing requires special knowledge (eg math expertise). Some can be frustrating. Descriptors in developing countries they are paid only a few dollars an hour on averagewithout any benefits or guarantees of future gigs.

A dry data well

So there are human reasons for finding alternatives to man-made labels. For example, Uber is expanding its fleet of gig workers to work on AI explanations and data entry. But there are also benefits.

People talk too fast. Explanations too favoritism which can be shown in their descriptions, and, after that, any examples taught on them. Translators create errorsor find was turned up when writing instructions. And paying people to do things is expensive.

More information in general they are expensive. Shutterstock is charging AI vendors tens of millions of dollars for access old recordswhile reddit they are made hundreds of millions from licensing data to Google, OpenAI, and others.

Finally, data is also becoming harder to find.

Many brands are trained on large collections of public information – which owners are very selective about accessing out of fear of loss. plagiarized or that they will not receive a loan or repay for it. More than 35% of the top 1,000 websites in the world now write OpenAI’s web scraper. And about 25% of the data from the “high-quality” data has been blocked from the large datasets used to train the models, which is recent. learning found.

If the current blocking strategy continues, the Epoch AI research team work that developers will run out of data to train artificial intelligence between 2026 and 2032. This, combined with fears of copyright cases and suspicious things developing their methods in open datasets, has forced the calculation of AI vendors.

Production methods

At first glance, artificial intelligence may seem like a solution to all of these problems. Need an explanation? Make them. Model information? No problem. The sky is the limit.

And to a certain extent, this is true.

“If ‘data is the new oil,’ synthetic data positions itself as natural oil, produced without the misrepresentation of the real thing,” Os Keyes, a PhD student at the University of Washington who studies the nature of emerging technologies, told TechCrunch. . “You can take a small starting point and compare and generate new information from there.”

AI companies have taken the idea and run with it.

This month, Writer, an enterprise-focused AI development company, released a model, the Palmyra X 004, trained almost entirely in manufacturing. It only cost $700,000 to build it, the Author says – in comparison compared to $4.6 million for a similar version of OpenAI.

Microsoft is Phi open source models were trained using artificial data, among others. The same goes for Google Gemma examples. Nvidia this summer unveiled a family of prototypes designed to create teaching aids, and the AI ​​Hugging Face startup recently released what it says is the largest of AI studies of creative writing.

Artificial intelligence has become a business in itself – which it can be important $2.34 billion by 2030. Gartner they predict that 60% of the data used for AI and analytics projects this year will be artificially generated.

Luca Soldaini, principal research scientist at the Allen Institute for AI, said that data mining methods can be used to generate training data in a format that cannot be easily obtained through scraping (or even licensing). For example, by training his video generator Movie GenMeta used Llama 3 to create annotations in the lighting, which people edited to add details, such as lighting descriptions.

Along the same lines, OpenAI claims that it has been improved GPT-4o using design data to create a sketchpad like Cloth ChatGPT interface. And Amazon did he said that it creates artificial data to supplement the real data it uses to train Alexa voice recognition models.

“Synthetic data models can be used to quickly increase one’s perception of what data is needed to achieve model quality,” Soldaini said.

Manufacturing risks

However, the data for Synthetic sinacea. It suffers from the same “garbage in, garbage out” problem as all AI. Examples make it generated data, and if the data used to train these models has biases and limitations, the results will be equally tainted. For example, groups that were not well represented in the original data will be represented in the synthetic data.

“The problem is, you can only do so much,” Keyes said. “Say you only have 30 black people in the dataset. Addition can help, but if all 30 people are middle-aged, or all light-skinned, that’s how the ‘representative’ information will appear.”

Until now, 2023 learning and researchers at Rice University and Stanford found that over-reliance on natural resources in training can create models in which “quality or diversity gradually diminishes.” Model bias – a poor representation of the real world – makes model differences worse after several generations of training, according to the researchers (although they also found that mixing in real-world data helps reduce this).

Keyes sees other risks in complex models such as OpenAI’s o1, which he thinks could lead to more difficult detection. it’s a nightmare in their production data. This, in turn, can reduce the accuracy of models trained on the data – especially if the source of the data is not easy to identify.

“Examples difficult to understand; “What is created with complex colors has authenticity,” Keyes added.

The combination of hallucinations can lead to magical models. A learning published in the journal Nature shows how models, trained on mistakes with errors, are made more the data is full of errors, and how the repetition of these responses destroys future generations of species. Models lose their esoteric knowledge over the generations, the researchers found – becoming more stable and producing inconsistent answers to the questions asked.

Follow up learning indicates that some types of models, such as image generators, are not protected from such crashes:

Soldaini admits that “raw” input should not be trusted, especially if the goal is to avoid training forgetful chatbots and image generators. Using it “safely,” he says, requires careful review, moderation, and filtering, and combining it with new, real content — just like you would with any other group.

Failure to do so in the end leads to model collapsewhere a model is less “creative” – ​​and biased – in its results, it ends up severely compromising its performance. Although these can be detected and arrested before they occur, they are dangerous.

“Researchers must analyze the data generated, review the process, and identify safeguards to remove low values,” said Soldaini. “Designed data pipelines are not automated; the results must be carefully monitored and controlled before they can be used in teaching.”

OpenAI CEO Sam Altman once said that AI will one day generating well-structured data for effective training. But – assuming it’s possible – the technology isn’t there. No major AI lab has produced a trained model on automatically generated data.

At least for the foreseeable future, it seems we’ll need people on the road somewhere to ensure that model training does not go astray.

TechCrunch has a newsletter focused on AI! Log in here to get it in your inbox every Wednesday.

Update: This article was originally published on October 23 and was updated on December 24 with additional information.