Today, foundational models train AI to learn languages human programmers are familiar with.
This is solving the wrong problem.
What matters is developing languages with intrinsic properties that allow AI to generate and execute code that does not require post hoc review by humans because it correctly satisfies users' and systems' requirements.
AI is the fastest growing group of engineers. Programming for AI is not the same as programming for humans.
Not only is AI meant to generate code humans use, but AI also needs to generate code in order to autonomously operate in real-world situations. At Outstanding we develop programming languages for AI, not humans. Our focus is on developing AI-purpose-built languages.
Our R&D established that certain math and programming methods generate code that is correct by construction. When used by AI, AI is able to not only generate accurate code, but it can do so in real-time and in new situations without prior information and without human intervention let alone post-hoc review. Our infrastructure is critical to AGI as intelligence in the real-world must be able to compose new code in real-time.
We have solved auto-formalization.
With respect to formal methods, we have 12 years of experience in working with Lean, Agda, Roq, and Idris. Consequently, we know through actual application the strengths and weaknesses of each. Their limitations drove us to to develop a new structure that is addressing critical gaps needed to scale auto-formalization across domains.
Our approach addresses key flaws in many market approaches to date that focus more on generic tools. While useful in mathematics, automation in the real world requires domain-specific information in order to correctly address real-world needs. We have focused on scaling domain-specificity. Complex domains are the substrate from which our R&D emerged.
Our expertise was first honed in financial services an industry where the stakes are high and errors can be fatal to a business if not the entire economy.
We originated, serviced, and managed $6+ billion in financial assets, including using AI in its embryonic stage to recommend portfolio-level actions, then using agents to execute them. To cite one example, we managed a loan portfolio that outperformed the best competitor by 500%with one just one employee to 30,000 borrowers.
Agentic finance.
Since then, we've applied our craft to create truly agentic finance in which precision-critical tasks cannot be approximated. For agents to truly be capable they must be able to execute autonomously and correctly.
Our infrastructure not only bounds agents, but also enables them to compose new code in real-time in order to adapt to new situations and instructions, all while ensuring that they do not deviate from the objectives of the parties they are acting on behalf.
Expanding beyond.
Using advanced math, we were able to abstract layers of our infrastructure and find applications in other sectors. We started first with legal because at the core all non-payment fintech is really legaltech. We structured a platform in order to ensure complete adherence to contractual obligations without human or AI error.
We then turned to healthcare, building a live application that combined clinical data and regulations to get veterans the care they deserve, and later to formal differential diagnosis. We formalized the tax code, encoding the Internal Revenue Code as a body of knowledge whose liabilities are derivable. And we built a world model in which AI agents dynamically compose new code in real time — an exercise that left us wanting to apply live synthesis to real-world domains, and also, maybe, to composing a new class of live games.
These pilots aren't about rapid expansion. They are the same thesis at different scales: formalize a body of knowledge, expose its language to agentic synthesis, show the proof, and let AI drive scalable, verifiable, cognitive work. The problems look completely different on the surface; underneath, they are the same problem.