SIA – The World’s First Open Source System For Self-Improving AI

In a world currently obsessed with the immediate outputs of Generative AI, there is a quieter, more profound revolution happening in Palo Alto. While most enterprises are scrambling to figure out how to write better prompts for closed models, Kunal Bhatia and the team at Hexo Labs are looking toward the horizon of what they call the “superintelligence takeoff”. We recently sat down with Kunal to discuss the foundational shift his company is leading: the creation of Self-Improving AI (SIA).

Hexo Labs is building a system where AI begins to build, optimize, and maintain itself. It is a mission to build “the last AI built by humans”. In this interview, we talk about the technical architecture behind self-improving systems, the economic necessity of owning your own intelligence, and why the “sovereign enterprise” is the only way forward for businesses that want to survive the coming decade of exponential change.

Watch the full episode here.

Kunal Bhatia,

Co-Founder of Hexo Labs

You’ve stated that Hexo Labs is working on something called “Self-Improving AI” or SIA. Could you break down the core thesis behind this and how it differs from the standard AI agents we see today?

To understand our work, you have to look at the current trajectory of the industry. My broader thesis is that eventually, AI is going to start building itself. We are currently in the early stages of a superintelligence takeoff. What that means in practical terms is that any version of AI you are building or using today is fundamentally one version below what it could be; there is always a better version waiting to be discovered through optimization.

Standard agents are static. A human engineer writes the code, defines the prompts, and sets the tools. At Hexo Labs, we have built a fundamental architecture—which we’ve released as an open-source repo called SIA, where a meta-agent improves a target agent. If you are a business building a specific agent for a task, our meta-agent takes that agent and improves it based on three critical metrics: performance, cost, and latency. It is an automated agent development process. Instead of a human spending weeks iterating on system prompts or tool calling, the meta-agent runs those experiments automatically.

One of the most compelling arguments you make is the shift from “renting” intelligence to “owning” it. Why is this distinction so important for the modern enterprise?

This is a philosophical difference as much as a business one. Right now, most companies are running their agents on proprietary, closed-source models like those from OpenAI or Anthropic. You are essentially renting their intelligence, and their business model depends on you doing so. They charge you per token, which is highly profitable for them, but it means they don’t want you to own your own weights. In fact, some providers have even updated their terms of service to explicitly state they don’t want you training your own models on their outputs, and they may even degrade service if you try to do so.

We want you to own your intelligence. When you own your intelligence—meaning you move to open-weights models hosted on your own infrastructure—you can bring costs down by almost 90% at inference time. Many companies use massive, general-purpose models for very specific tasks like text extraction or classification. That is an incredibly expensive way to solve a simple problem. By using SIA to transition from a massive closed model to a smaller, specialized open-weights model, you maintain the same performance at a fraction of the cost.

Furthermore, if you had to do this shift manually with engineers, it would take months to translate the logic from one model to another. SIA does it in a matter of hours.

We wanna build the last AI built by humans.

Technically speaking, how does the meta-agent actually “improve” another agent? What are the “knobs” it is turning behind the scenes?

When you build an agent, you have several variables, or knobs, that you can turn to find the right configuration. We primarily look at four or five of them: the harness, the weights of the model, the memory architecture, the specific skills, and the tools the agent has access to.

Testing across all these variables involves running hundreds of experiments to see what works best against your evaluation scores (evals). Normally, an engineer would have to set up infrastructure for every single experiment and deploy a new version every time they wanted to test a change. Our backend system automates this entire process. Our agent is essentially a “hill-climbing” agent; it runs hundreds of experiments in parallel, learns from each one, and figures out exactly how to hit the target eval metrics without human intervention.

You mentioned a staggering statistic: a 350x acceleration in AI development. How was that measured, and what does it look like in a real-world scenario?

We measured this using MLE Bench, a benchmark released by OpenAI that measures an agent’s ability to solve machine learning problem statements—essentially testing if an AI can build other AI.

When we looked at the history of human-developed agents on this benchmark, it took teams of humans about six months of constant iteration, writing better scripts, and refining agents to reach the top metrics. When we ran SIA on those same problem statements, it reached the number one spot on the leaderboard in just 12 hours. That is a 350 to 360x improvement in time.

The most interesting part was that SIA didn’t just reach the top; it iterated to beat its own record and then did it again. It is like Michael Phelps setting a world record and then immediately jumping back in the pool to beat himself. This is the power of agents building agents, they don’t sleep, they don’t get bored, and they don’t experience the “drudgery” that a human engineer feels when maintaining a model in production.

The human mind is not programmed to process in an exponential way.

There is a lot of anxiety regarding AI making jobs redundant, specifically in software engineering. If AI can now handle architecture and optimization, where does that leave the human?

I view this as the democratization of superintelligence. Right now, these high-level capabilities are concentrated in a handful of labs like OpenAI, DeepMind, and Anthropic. Those organizations have thousands of the most brilliant researchers and engineers in the world. If the top Fortune 2000 companies all wanted to build their own internal AI labs of that caliber, the world would need an additional five million AI researchers—and they simply don’t exist.

AI supplements that talent shortage. While AI takes over the “grunt work” and the manual iteration of development, humans move into the role of the director and the verifier. We are the ones who provide the taste, the direction, and the high-level goals. We evolve alongside the technology. Instead of being lost in the weeds of coding every tool call, the human engineer becomes a specialist who manages these exponential levers.

You’ve touched on the “Sovereign Enterprise.” What does that mean for a business’s long-term strategy and even for national interests?

In the future, your AI capabilities are going to be the biggest weapon you have as an enterprise. It is your most critical infrastructure. Because of that, you simply cannot afford to have it controlled by a third party.

There is also a geopolitical layer to this. Some of the best open-source models in the world are currently coming out of China. In that context, there is no stopping the development of these systems; it has become a race for superpower status. The real challenge isn’t putting on the brakes, but rather focusing on AI alignment, ensuring these systems are aligned with human values and the specific needs of an organization. We want to enable enterprises to be sovereign, meaning they own their own intelligence and can direct it according to their own mission.

Your AI capabilities are going to be your biggest weapon that you have as an enterprise. It’s your most critical infrastructure in the future. You need to own it.

How does the human mind, which is traditionally linear, adapt to the “exponential change” you describe?

We are entering an age where the progress of the last five years is going to be dwarfed by the progress of the next five years. The human mind is naturally programmed for linear thinking—we look at the past and assume the future will move at the same pace. But the curve is shifting upward.

At Hexo Labs, we aren’t building a point solution for today’s problems. We are building a lever for rapid, continuous change. Businesses need to stop thinking about static workflows and start thinking about dynamic, self-improving capabilities. If you align yourself to the linear, you will be left behind. You have to align yourself to the exponential.

Hexo Labs has a strong connection to academia and a massive community in the Bay Area. How does that ecosystem influence your work?

We run the Frontier Research Institute and a community called the Frontier Research Club. Every week in the Bay Area, we host research talks that attract top engineers and scientists. We now have a community of over 3,000 AI researchers and engineers.

This partnership with academia is vital because we are building an agent that essentially does AI research. Research involves solving problems over extremely long horizons with many steps, and working with scientists allows us to verify our outcomes in complex fields like quantum computing and material science. SIA is already being used to help scientists accelerate their own work by building new AI models for frontier research.

For engineers looking to enter this space, what is the most important skill they need to develop to work with probabilistic systems like SIA?

Traditional software engineering is binary: either the code works, or it doesn’t. AI is different; it is probabilistic. Often, an AI “works,” but it doesn’t give the output you wanted.

The most important shift for an engineer is developing a scientific mindset focused on evals. You have to become a master of experimentation and verification. You are no longer just a builder; you are a verifier of the work that agents are doing. This is the mindset of the future.

Finally, what was the personal spark that led you to build Hexo Labs?

This is my third AI company. My co-founder and I previously ran an AI consulting business where we built models for other companies. Our pipeline grew so large that I eventually told my co-founder, “I need an AI version of you”.

That thought stuck. If AI can start building AI, that is the end game. It is the most efficient way to scale intelligence. We decided then that we wanted to build the “last AI built by humans”. Since then, we have seen massive validation from developers across the Fortune 500 and fast-growing startups who are all looking for the same thing: the ability to own and scale their own agents.

Kunal, thank you for your time, your transparency, and your dedication to building the future of intelligence.