capability-research

A test's questions, and often their correct answers too, quietly end up inside the test-taker's study materials before the exam ever happens — a model's high score may reflect that it memorized the answers, not that it genuinely learned the underlying <a href="https://claudeskill-me.com/en/glossary/workflow/skill/" target="_blank" rel="noopener">Skill</a>.

Having a reasoning model write out its "thinking" process in natural language, so researchers can read that text and watch for problematic intent or behavior — currently one of the few safety monitoring methods that doesn't require dissecting a model's internal structure, relying instead on simply reading what the model says about itself. But its effectiveness rests on a fragile premise: whether the reasoning written out actually reflects the model's real decision-making process.

A model that can process different types of data — text, images, audio, video — within a single architecture, and have that different information genuinely cross-reference and inform each other inside the model, rather than being handed off separately to a few independent specialized models whose outputs then just get stitched together afterward. This internal fusion capability is the most fundamental difference between <a href="https://claude-me.com/en/glossary/core-concepts/multimodal/" target="_blank" rel="noopener">Multimodal</a> models and the older approach of chaining separate AI systems together.

Instead of pouring all computational resources into training to make a model bigger, this approach spends extra compute at the moment a model actually answers a question — letting it "think longer" by exploring multiple reasoning paths, self-checking, and repeatedly revising — trading that additional compute for better answer quality. It's a separate axis driving AI capability progress, distinct from <a href="/en/glossary/scaling-laws/compute-scaling/">Compute Scaling</a>.