Every week – sometimes every day – a new one state-of-the-art AI model born into the world. As we head into 2025, the pace of releasing new models is dizzying, if not exhausting. The rollercoaster curve continues to grow rapidly, and exhaustion and surprise become constant companions. Each release highlights why IT The particular model is better than all, with endless collections of benchmarks and bar charts filling our feeds as we struggle to keep up.

Eighteen months ago, most developers and businesses were using a an AI model. Today, the opposite is true. It is rare to find a business of significant scale that limits itself to the capabilities of a single model. Companies are wary of vendor lock-in, especially for a technology that has quickly become a core part of both long-term corporate strategy and short-term bottom-line profits. It is more risky for teams to place all their bets on a large language model (LLM).

But despite this fragmentation, many model providers are still championing the view that AI will be a winning market. They admit that the expertise and computing required to train best-in-class models is scarce, defensible and self-reinforcing. From their point of view, the hype bubble is for building AI models will eventually collapse, leaving a single, giant artificial general intelligence (AGI) model to be used for anything and everything. The exclusive ownership of such a model means becoming the most powerful company in the world. The size of this prize started an arms race for more and more GPUs, with a new zero added to the number of training parameters every few months.

We believe this view is wrong. No single model will rule the universe, either next year or the next decade. However, the future of AI will be multi-model.

Language models are fuzzy commodities

the Oxford Dictionary of Economics defines a commodity as a “standardized item that is bought and sold in quantity and whose units are interchangeable.” Language models are commodities in two important senses:

1. The models themselves are becoming more interchangeable with a wider set of tasks;
2. The research expertise needed to create these models is increasingly distributed and accessible, with labs across borders almost entirely independent of each other and independent researchers in the open source community working together. – running on their heels.

But as language models commoditize, they don’t do the same thing. There is a large core of capabilities that any model, from the GPT-4 to the Mistral Small, is perfectly suited to handle. At the same time, as we move to the margins and edge cases, we see greater and greater differences, with some model providers apparently specializing in code generation, reasoning, retrieval-augmented generation (RAG ) or math. This leads to endless handwringing, reddit-searching, evaluation and fine-tuning to find the right model for each job.

And so while language models are commodities, they are more accurately described as fuzzy commodities. For many use cases, AI models are almost interchangeable, with metrics like price and latency determining which model to use. But at the edge of capabilities, the opposite will happen: Models will continue to specialize, becoming more and more different. As an example, Deepseek-V2.5 stronger than GPT-4o in coding in C#, despite being a fraction of the size and 50 times cheaper.

Both of these dynamics – commoditization and specialization – uproot the thesis that one model is best suited to handle every possible use case. Instead, they focus on a increasingly fragmented landscape for AI.

Multi-model orchestration and routing

There is an apt analogy for the market dynamics of language models: The human brain. The structure of our brains has remained unchanged for 100,000 years, and brains are more similar than dissimilar. For most of our time on Earth, most people have learned the same things and have the same capabilities.

But something has changed. We develop the ability to communicate in language — first speaking, then writing. Communication protocols facilitated networks, and as people began to network with each other, we also began to specialize to greater and greater degrees. We have been freed from the burden of needing to be generalists in all domains, to become self-sufficient islands. Paradoxically, the collective wealth of specialization also means that the average person today is a stronger generalist than any of our ancestors.

In a sufficiently wide input space, the universe always tends toward specialization. This is true from molecular chemistry, to biology, to human society. Due to sufficient diversity, distributed systems are often more computationally efficient than monoliths. We believe the same will happen with AI. The more we can use the strength of multiple models instead of relying on just one, the more we can specialize in models, expanding the boundaries for capabilities.

A particularly important pattern for exploiting the power of different models is routing — dynamically sending queries to the most appropriate model, while also using cheaper, faster models when doing so is harmless. in quality. Routing allows us to take advantage of all the benefits of specialization – higher accuracy with lower cost and latency – without sacrificing any of the robustness of generalization.

A simple demonstration of the power of routing can be seen in the fact that most of the world’s leading models are routers themselves: They are built with Expert Blend architectures that route each successive generation to several dozen expert sub-models. If it is true that LLMs are rapidly multiplying fuzzy commodities, then routing should be an integral part of every AI stack.

There is a view that the LLMs will be the plateau in their reach of human intelligence – that as we fully saturate the capabilities, we will come together in a general model in the same way that we come together around the AWS, or on the iPhone. None of the platforms (or their competitors) have the 10X capabilities of previous years – so we can also be comfortable in their ecosystems. We believe, however, that AI does not stop at human-level intelligence; it will continue beyond any limits we can imagine. In doing so, it becomes more fragmented and specialized, like any other natural system.

We cannot overstate how well the AI ​​model is segmented. Fragmented markets are efficient markets: They empower buyers, increase innovation and reduce costs. And to the extent that we can leverage networks in smaller, more specialized models instead of sending everything through the internals of one giant model, we’ll move toward a safer, more understandable and more manageable future for AI.

The greatest invention has no owner. Ben Franklin’s successors did not have their own electricity. Turing’s estate does not own all computers. AI is undoubtedly one of humanity’s greatest inventions; we believe that its future will be – and should be – multi-model.

Zack Kass is the former head of go-to-market at OpenAI.

Tomás Hernando Kofman is the co-Founder and CEO of Not Diamond.