We are on a mission to build Engineering SI for the physical and digital world.
As AI coding agents improve with each model release, I cant help but wonder if we are approaching a software singularity. In my view, I think the first signal would be models that can autonomously build the next layer of the digital world. Specifically, I'm referring to operating systems, game engines, CAD Kernals, Browsers i.e systems that took entire organizations and years to build. The very systems that became infrastructure for the digital world.
In many ways the current programming paradigm has shifted towards a latent trajectory optimization of developer prompts, an iterative & sequential process consisting of a developer and a single coding agent loop to shape an isomorphism between intent to model output.
One major bottleneck in this case is the ability for us to leverage models to their fully capacity and effectively. Breaking the single agent loop however is very difficult for several reasons including memory, the challenge of long horizon coordination that must also scale with the number of coding agents in a given swarm, and thefact that most models were trained for single agent loops.
Very quickly most multi agent systems fail with scale. To adress this, we have dedicated significant efforts to develop a new compiler for agent swarms from scratch, with the goal of optimizing the isomorphism between developer intent, and breaking the single coding agent loop.
Krissh_C•42m ago
We are on a mission to build Engineering SI for the physical and digital world.
As AI coding agents improve with each model release, I cant help but wonder if we are approaching a software singularity. In my view, I think the first signal would be models that can autonomously build the next layer of the digital world. Specifically, I'm referring to operating systems, game engines, CAD Kernals, Browsers i.e systems that took entire organizations and years to build. The very systems that became infrastructure for the digital world.
In many ways the current programming paradigm has shifted towards a latent trajectory optimization of developer prompts, an iterative & sequential process consisting of a developer and a single coding agent loop to shape an isomorphism between intent to model output.
One major bottleneck in this case is the ability for us to leverage models to their fully capacity and effectively. Breaking the single agent loop however is very difficult for several reasons including memory, the challenge of long horizon coordination that must also scale with the number of coding agents in a given swarm, and thefact that most models were trained for single agent loops.
Very quickly most multi agent systems fail with scale. To adress this, we have dedicated significant efforts to develop a new compiler for agent swarms from scratch, with the goal of optimizing the isomorphism between developer intent, and breaking the single coding agent loop.