We used to compare AI models only on benchmarks. Did it score higher on math? On law? On commonsense reasoning? This made sense when capabilities were limited—when you needed all the intelligence you could get, you chose the model that performed best on objective metrics.
But with the release of the Claude 3 family and the maturation of competing models, we're realizing that raw capability is only part of the story. Models have "flavor." They have cognitive styles. They have what we might awkwardly call personality.
Beyond Benchmarks
Some models are creative and loose—they generate unexpected associations, take risks in their outputs, surprise you with novel approaches. Others are rigid and precise—they stick closely to instructions, minimize embellishment, optimize for accuracy over interest. Some are verbose, offering extensive explanations and context. Others are concise, giving you just what you asked for and nothing more.
These differences aren't bugs; they're features. They reflect different training approaches, different architectural choices, different optimization targets. A model trained to be helpful in customer service develops different patterns than one trained to assist with research. A model optimized for safety develops different tendencies than one optimized for creativity.
Benchmarks don't capture these qualitative differences. Two models might score identically on a math test but feel completely different to work with. One might explain its reasoning in a way that helps you learn; another might just output the answer. One might ask clarifying questions when instructions are ambiguous; another might assume and proceed. The "flavor" matters.
Choosing Collaborators
We are beginning to choose our AI collaborators the way we choose human colleagues: not just for raw IQ, but for "fit." When you hire a team member, you consider their skills, certainly, but also their communication style, their work approach, their personality. The brilliant jerk and the pleasant mediocrity are both suboptimal; you want capability that comes in a compatible form.
The same logic applies to AI. For creative brainstorming, you want a model that generates wild ideas and runs with them. For legal review, you want one that's cautious, precise, and flags uncertainties. For coding, you might want one that's terse and efficient. For therapy or coaching, you might want one that's warm and exploratory.
This creates a matching problem: which model for which task? And it creates a design question: how do we build models with the right cognitive styles for different uses? The answer isn't one model to rule them all, but a portfolio of minds with complementary strengths.
The Society of Minds
This suggests that the future isn't one God-AI that handles everything, but a society of minds—different models for different purposes, each with its own cognitive profile. You will have a creative partner for ideation, a stern editor for review, a coding wizard for implementation, an emotional confidant for reflection.
This mirrors how human expertise works. We don't expect one person to be the best lawyer, doctor, engineer, and artist. We specialize. We build teams. We match specialists to problems. The same structure is emerging in AI.
It also suggests that "better" isn't a single dimension. A model can be better for you without being better in general. The chatbot that annoys me might delight you. The assistant that feels too formal to one user might feel appropriately professional to another. Personal fit becomes a design goal.
Diversity of Thought
Diversity of thought is coming to silicon. Where once we might have worried about AI monoculture—a single model's biases becoming universal—we're now seeing genuine cognitive diversity. Different models have different blind spots and different strengths. Their disagreements can be productive, their variations can be explored.
This diversity is both a challenge and an opportunity. The challenge is navigation: how do users find the model that fits their needs? How do developers build applications that route to appropriate specialists? The opportunity is richness: a world of diverse AI minds offers more than a world of identical ones.
We're moving from "the AI" (singular) to "an AI" (one of many). The models are developing identities, preferences, styles. They're becoming individuals in a meaningful sense—not conscious, not sentient, but distinct. The society of minds is taking shape.