Despite the intense arms race in artificial intelligence, we face a multimodal future

Every week – and sometimes every day – is new Modern artificial intelligence model Born to the world. As we move into 2025, the pace at which new models are being launched is astonishing, if not exhausting. The roller coaster curve continues to grow exponentially, and fatigue and wonder have become constant companions. Each issue highlights the why this One particular model is better than all the others, with endless combinations of benchmarks and bar charts filling our feeds as we strive to keep up.

Eighteen months ago, the vast majority of developers and businesses were using One AI model. Today, the opposite is true. It is rare to find a company of significant size limited to the capabilities of a single model. Companies are concerned about vendor lock-in, especially for technology that is quickly becoming a key part of a company’s long-term strategy and short-term bottom line. It is increasingly risky for teams to place all their bets on one big language model (LLM).

But despite this fragmentation, many modelers still subscribe to the view that AI will be a winner-take-all market. They claim that the expertise and computation required to train best-in-class models are scarce, defensible, and self-reinforcing. From their point of view, the hype bubble Building artificial intelligence models It will eventually collapse, leaving behind one giant model of artificial general intelligence (AGI) that will be used for anything and everything. To have such an exclusive model means to be the most powerful company in the world. The size of this prize has sparked an arms race for more and more GPUs, with a new zero added to the number of training parameters every few months.

We believe that this opinion is wrong. There will not be one model that will rule the universe, neither next year nor in the next decade. Instead, the future of AI will be multi-paradigm.

Linguistic models are ambiguous commodities

the Oxford Dictionary of Economics A commodity is defined as “a uniform good that is widely bought and sold and whose units are interchangeable.” Linguistic models are commodities in two important senses:

1. The models themselves have become more interchangeable in a wider range of tasks;
2. The research expertise required to produce these models is becoming more distributed and accessible, with frontier labs barely outpacing each other, and independent researchers in the open source community closing in on their niche.

But as linguistic forms become commodified, they do so unevenly. There is a wide range of capabilities that any model, from GPT-4 to Mistral Small, is well suited to handle. At the same time, as we move toward the margins and edge states, we see greater and greater differentiation, with some model providers explicitly specializing in code generation, inference, retrieval augmented generation (RAG), or mathematics. This leads to an endless process of searching, researching, evaluating and fine-tuning to find the right model for each job.

Thus, while linguistic models are commodities, they are more accurately described as Mysterious goods. For many use cases, AI models will be virtually 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, and become more distinct. For example, Deepsec-V2.5 More powerful than GPT-4o for C# programming, although a fraction of the size and 50 times cheaper.

These two dynamics—commoditization and specialization—are uprooting the assumption that a single model will be best suited to handle every possible use case. Instead, they point to a progressively fragmented landscape for artificial intelligence.

Multimedia coordination and guidance

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 much more alike than they are different. For the vast majority of our time on Earth, most people learned the same things and had similar abilities.

But then something changed. We have developed the ability to communicate in language – first by speech, then by writing. Communication protocols facilitate networking, and as humans began to communicate with each other, we also began to specialize to greater and greater degrees. We are freed from the burden of needing to be generalists in all areas, to be self-sufficient islands. Paradoxically, the collective wealth of specialization also means that the average person today has become much stronger science specialists than any of our ancestors.

In a sufficiently large input space, the universe always tends toward specialization. This is true all the way from molecular chemistry, to biology, to human society. Given enough diversity, distributed systems will always be more computationally efficient than monoliths. We believe the same applies to artificial intelligence. The more we can leverage the strengths of multiple models rather than relying on just one, the more those models can specialize, expanding the boundaries of capabilities.

One increasingly important pattern for leveraging the strengths of diverse models is routing – dynamically sending queries to the most appropriate model, while also taking advantage of cheaper, faster models when doing so does not degrade quality. Routing allows us to take advantage of all the advantages of specialization—higher accuracy with lower costs and lower latency—without giving up any of the power of generalization.

A simple demonstration of the power of routing can be seen in the fact that most of the best models in the world are routers themselves: they are built using A mix of experts Architectures that direct each generation of tokens into a few dozen specialized sub-models. If it is true that LLM holders are deploying murky goods in a big way, then mentoring should become an essential part of every AI stack.

There is a view that LLMs will stabilize as they reach human intelligence – that when we fully saturate capabilities, we will coalesce around one general model in the same way we coalesced around AWS, or the iPhone. None of these platforms (or their competitors) have gained 10x their capabilities in the past two years – so we may take comfort in their ecosystems. However, we believe that artificial intelligence will not stop at the level of human intelligence; It will continue to transcend any limits we might imagine. As it does so, it will become increasingly fragmented and specialized, just as any other natural system would.

We cannot overstate how important AI model segmentation is. Segmented markets are efficient markets: they give buyers power, maximize innovation, and reduce costs. To the extent that we can leverage networks of smaller, more specialized models rather than sending everything through the internals of one giant model, we are moving toward a safer, more explainable, and more steerable future for AI.

The greatest inventions have no owners. Ben Franklin’s heirs don’t have electricity. Turing ownership does not own all computers. There is no doubt that artificial intelligence is one of humanity’s greatest inventions; We believe that its future will be – and should be – multi-model.

Zach Kass is the former Head of Go-to-Market at OpenAI.

Thomas Hernando Kaufman is the co-founder and CEO of Not diamonds.