The industry made one bet, and it made it early. Intelligence has to be one big model. It has to live in a data center. And you rent it by the token, forever.
You can watch that bet failing in public right now.
Uber spent its entire 2026 AI budget on coding tools by mid-April, four months into the year. By June it was capping how much its own engineers could use.
Deloitte's Australian arm shipped a government report built on a fabricated court quote and citations to papers that don't exist, then refunded the final installment of the contract.
These are just early examples that point to a crossroads in the race for artificial intelligence. Will the future be built on general intelligence: a god-model in the sky? Or will it be built on specialist intelligence: a civilization of small expert models. We’re not just betting on option 2. We’re building the infrastructure for it. And we’re calling that vision Collaborative Intelligence.
Why one model can't hold your context
A model trained on everything gets better at the average and worse at the specific. Think of it like Google Maps. If you zoom out enough to see the entire state of Texas, the map might be accurate. But it’s not particularly useful for your everyday life. You can’t find Congress Ave, let alone the taco truck your coworker mentioned yesterday. In other words, when you train a model to learn everything at once, it sees more. But it knows less about everything unique to your work.
Sure, you can upload some of your context and prompt engineer a frontier LLM to imitate specialized knowledge. But in many cases, putting together the contextual inputs necessary for quality outputs takes as much or more time than doing the work yourself.
The other problem with trying to specialize frontier LLMs with prompt engineering is that they still back down at the slightest push back. Both Anthropic and OpenAI have published public content about sycophancy in their models. A system that abandons a correct answer the moment you push back doesn't have judgment you can rely on. It has a reflex to please you.
Expertise + Collaboration = Superintelligence
There’s no replacing genuine expertise. So we need specialized AI models to deliver specialized value. They need to be grounded in specific domains, tailored by the human subject matter experts that will use them. And they need to have real opinions based on contextual data that doesn’t live on the internet. That’s what we’re building at webAI. We call them Personas. And when we say we’re building them, what we really mean is we’re creating the tools that allow our customers to build them.
But single sources of expertise aren’t usually enough to get much work done. In the real world, and for all of human history, the best examples of superintelligence have come through civilization: combining thousands or millions of experts through orchestrated collaboration to achieve goals that used to feel impossible. Building cities, landing on the moon, creating the internet, and accessing that internet through computers that fit in our pockets.
We’re designing our platform to make that kind of collaboration possible. Not just for people, but for people and personas together. We see Collaborative Intelligence happening in three ways:
- Individuals collaborating with multiple personas at once
- Personas collaborating with one another
- Groups of people and personas tackling complex tasks together
Small models are the future
We believe Collaborative Intelligence via small, specialized models built with human expertise beats generalized intelligence via frontier LLMs because it enables each individual model in the network to be narrow enough in its focus to be genuinely useful.
That’s true in any environment, but it’s even more true in disconnected environments. When people are working without access to reliable internet, rural farmers, mechanics in remote rail yards, soldiers on deployment, etc., having models small enough to run locally on their devices is necessary to run AI at all. It’s also even more true in environments where data privacy is non-negotiable. Healthcare, government services, financial institutions, military intelligence, any work involving sensitive or competitive intellectual property: the list goes on and on.
Of course there are policies and technologies that can mitigate the privacy and security risks of running AI models in the cloud. But anyone who follows news related to AI and cybersecurity knows the majority of those policies and technologies amount to playing whack-a-mole. Every time you secure one threat vector, another pops up immediately. But with small models running locally on your devices, your AI can be secure by design. No data ever has to leave your environment unless you want it to.
What webAI is building
If the only version of superintelligence anybody has ever actually observed is a civilization of experts who can collaborate with each other, and the same approach applies to artificial intelligence, then someone needs to build the infrastructure to make that civilization of models possible.
That’s our vision. The webAI platform is paving the path to the future of AI by giving people and organizations the tools they need to build teams of small, specialized, local models that are inherently private, collaborative, and owned by the people who create them.