AI models are choking on junk data

Source Domain: fortune.com

The progress of transitioning from language models like ChatGPT to sophisticated humanoid robots hinges on the quality of the training data.
Historically, more data led to smarter models, but this approach may not suffice for physical AI which requires richer, multifaceted data.
A looming crisis involves an overabundance of “junk data” — data that does not aid in AI model development.
The demand for data has fueled the rise of expensive AI data startups but often produces useless data for high-stakes physical AI applications.
Effective physical AI requires significantly more data, which is harder to obtain and often requires complex simulations due to the complexity of the physical world.
Using junk data can degrade performance, delay market-readiness, and lead to unpredictable outcomes, especially for applications like self-driving cars.
Examples like OpenAI’s sunset of its AI video app highlight issues stemming from inadequate training data for physical AI.
To fully utilize AI’s potential, machine learning teams must prioritize filtering out junk data to ensure the training of effective AI models.
High-quality data is now seen as the critical constraint, and those who adapt to this challenge first will likely lead in the development of successful AI systems.