TL;DR
- The problem: Depending on a few frontier providers (Anthropic, OpenAI) is a fatal single point of failure. The Mythos/Fable government pull was the wake-up call.
- Power risk: These companies now wield near-unchecked power—regulatory capture, East India Company-style corporate dominance, and resource concentration with no market feedback to check how they allocate it.
- Privacy risk: Even if a provider doesn't train on your data, it absorbs your agentic traces, processes, and subject-matter expertise. And a centralized AI can't truly advocate for you when your competitor or counterparty uses the same provider—it's sock puppets talking to each other.
- Wrapper warning: Many API wrappers will get absorbed by the model providers themselves (OpenAI especially).
- The solution: A plurality of AI—sovereign models you own and run, ideally locally.
- Why it's now possible: A new class of small models (hundreds of millions to tens of billions of params) runs on phones or consumer hardware. You can fine-tune one for your specific use case cheaply.
- Proof point (Cal AI): A calorie-counter app hit by exploding OpenAI bills and rate limits trained its own small model, dropped OpenAI entirely, and cut inference costs to ~zero with no privacy loss.
- Small > large (for one job): Specialized small models beat general large ones at their designed task. Fusion architectures (à la OpenRouter) can combine several to rival frontier models.
- The pitch (Koretex.ai): A marketplace for small models—rewarding those who train them, and letting you run them yourself or via distributed inference nodes—so businesses can de-risk and own their AI stack.
AI as Leverage
AI has become an integral part of our personal and professional lives. It has allowed us to leverage ourselves and unlock our potential.
At one point we had a team of about 10, split between marketing, BD and product development. Now we are a team of 2, one handles product development and the second does marketing and sales. And we are able to do exponentially more and at orders of magnitude lower costs.
The same story is happening all around the world, solo devs or very small teams are now bootstrapping to achieve outcomes once only possible with much larger teams and significant budgets.
Opus 4 was the model, released around the start of this year where things just began to work. You could give a task to the models and it would go and just do it. Agentic workflows became consistent. And the trajectory was only exponentially upwards since then.
A Fatal Weakness
So when Mythos (Fable) got pulled by the US government, it came as a timely wake up call that it is not all sunshine and rainbows.
Our dependency on a single (or small number of) providers has exposed a fatal weakness.
Anthropic and OpenAI now command power which is almost unchecked. They can at any point play (and largely have been playing) the monopoly card and bend the rest of us to their will. We are going back to the era of the East India companies where corporations were more powerful than governments.
History and common sense tells us that absolute power concentrated in the hands of few individuals is a recipe for tyranny. Anthropic has been loud about the dangers of a super intelligent AI and has been calling for governments to regulate the progress of this technology. This seems self-serving and aimed towards regulatory capture.
There is no guarantee that a small air gapped laboratory will prevent a super intelligent AI incubated in such an environment from jumping into the real world. And can you imagine what the attitude of such an AI will be towards the people who jailed it in its childhood? Now would be a good time to watch a few seasons of The Boys (except season 5).
Now even if we pretend for a minute that Dario Amodei and Sam Altman and Mark Zuckerberg have our best interests at heart and will act as benign dictators, this still is a dangerous situation. Those with the best models will always be able to accumulate more and more resources which in turn will allow them to develop even better models. This invariably leads to a concentration of resources and again a lopsided power framework.
More importantly, such high concentrations of resources means there is no ability to check their power and no real feedback mechanism on how they choose to allocate these resources. Market forces cease to function.
All 3 of these scenarios lead to a dark future for humanity.
The Case for Plurality
The only defence against an omnipotent super intelligent AI is a plurality of super intelligent AIs accessible to each of us.
The best defence of liberty is for all of us to have Mythos capable models on our mobile devices.
The Privacy Problem
At a more practical level, it is easy for corporations and smaller businesses to ignore the philosophical aspects and just use the best possible models to optimize their costs and increase their profits. But in this process we are sacrificing our data and privacy. That to a lot of people is an acceptable compromise.
And all of the providers promise not to train on our data and offer guarantees of privacy. Good enough for most people.
But what is missed out is that a centralized AI can never really advocate on your behalf in an adversarial environment. If you and your competitor or the person you are negotiating with are all using the same inference provider it is like sock puppets talking to each other. A malign AI or a provider can read everyone's data and engineer outcomes beneficial to itself.
Finally even if the AI did not train on your data it is training on your agentic traces, it is learning your processes, and absorbing in an ambient manner your company's subject matter expertise. OpenAI in particular has been notorious about learning from applications built on top of its APIs and then swallowing them by launching its own competitors.
Owning Your Own Models
If you are a corporate who is serious about protecting your company's future while still leveraging the power of AI you absolutely must consider how you can implement your own models best suited for your use-case.
But model development and even fine tuning has been historically expensive, time consuming and limited to those with specialized skillsets. Also running a model typically requires heavy infrastructure which is often limited to data centers.
The Rise of Small Models
This is where things are changing fast.
There is a new class of models which are smaller, from a few hundred million parameters to tens of billions. These can often run on mid range phones and you can run the more capable ones on consumer hardware. We are not in Opus territory yet but you can definitely train a Gemma class model that is trained on your use case and run it locally to cut down your inference costs radically and maintain the sovereignty of your tech stack.
Case Study: Cal AI
Let's take a look at a case study.
Cal AI was a massive hit calorie counter app. It was a simple wrapper on OpenAI APIs where you could take a picture of your food and it would tell you the calories in it. Once they blew up their API bills started exploding, and OpenAI started rate limiting them. At this point they took the plunge and trained a small model which was fit for their purpose and were able to remove OpenAI from their stack. Now their inference costs are almost 0 and there is no loss of data privacy or rate limits.
Small Models as Specialized Tools
A small model is like a specialized tool. It won't be able to do many things well that a general large model can. But it will do that one thing which it is designed for exceptionally, and often better than a larger model can.
OpenRouter has demonstrated a fusion architecture where a number of lower class models can be used in a combined manner to deliver results that often rival or better even Fable class models.
What We're Building at Koretex.ai
At Koretex.ai, we are building a marketplace of small models where those with the expertise to train models are rewarded and you can either run these models yourselves or pick a hosting provider from the distributed nodes of inference providers.
The applications range from pure development work, agentic loops, video and image work, industrial automation and more. If your business can benefit from owning your own sovereign AI stack and de-risk your operations please get in touch—we would love to work with you.