Scientists have unlocked the black box of AI learning by mapping out something called the “concept space.” They found AI systems don't just memorize, they actually build a sophisticated understanding of concepts at different speeds and develop the ability to mix and match them in creative ways. Researchers found that AI models develop capabilities in two distinct phases and we need to rethink how we evaluate AI capabilities. This discovery fundamentally changes how we should think about AI systems.
AI models might already understand complex combinations of concepts we haven't discovered yet. Researchers trained an AI model on just three types of images and asked if it could create a small blue circle. The model struggled with normal prompts, but discovered two techniques that produced the output: 'latent intervention' and 'over-prompting.'
Latent intervention is like finding a backdoor into the model's brain. They found that by turning color and size dials in specific ways, the model could suddenly produce what seemed impossible moments before.
Meanwhile, the over-prompting technique is like the difference between saying “make it blue” versus “make it exactly this shade of blue: RGB(0.3, 0.3, 0.7).” This extra precision helped the model access abilities that were hidden under normal conditions.
When researchers tested these ideas on real-world data using the CelebA face dataset, they found the same patterns. Regular prompts failed, but using latent interventions revealed the model could actually create images of “women with hats” – something it had not seen in training. This implies the models develop capabilities at different speeds depending on how strongly concepts stand out in training data.
AI models develop their abilities in two distinct phases. First, they learn how to combine concepts internally, which is what happens around step 6,000. However, there's a second phase where they learn how to connect these internal abilities to our normal way of asking for things. We need to get better at unlocking their hidden abilities.
This discovery fundamentally changes how we should think about AI systems. Just because a model might not be able to do something with standard prompts does not mean it cannot do it at all. We need to get creative with how we interact with AI and ask whether the model truly lacks the capability or if we're just not accessing it correctly.