The Age of Verification
and we should change our thinink
The core idea is simple: its verification process was fast and cheap, and this allows AI to quickly iterate the compiler at an unprecedented rate. I see this capability not as a mere incremental improvement in code generation, but as a fundamental paradigm shift.
We are at the beginning of new era: the age of verification.
The complexity of a problem no longer depends on itself but on how fast and cheap we can verify the solution. A solution can start simple but quick iteration allows the solution to evolve into something complex and robust. In the age of verification, the complexity of the solution is a function of the speed and cost of verification.
There are a number of recent works that demonstrate this trend:
IMO Gold medal level: a paper that describes a model-agnostic, verification-and-refinement pipeline that achieves gold medal level at the highest math competition for high school students.
The Graph Networks for Materials Exploration (GNoME) discovered 2.2 million new materials. Verification methods include Density Functional Theory, a physics-based simulation.
Autoresearch (by Andrej Karpathy): An AI agent that can perform scientific research by itself. Each solution is measured by the final test data.
What we used to think impossible now could be solve by simple tools: a for loop and a verification engine. Frontier models are getting better at going in the right direction to the solution instead of doing random search.
There are lots of details and nuances in producing and iterating over a solution. And I am not saying it’s easy. But this seems be not a bottleneck anymore. This marks a shift in mentality thinking:
This applies to programming, math, and many other fields. You should not constraint yourself by the complexity of the problem. Instead you should start with an assumption that if there is a solution how you would verify it. If you look around, you will notice that many problems can be solved by this mindset.
I’m seeing the shift toward verification fundamentally alter software engineering. While Andrej Karpathy coined “vibe coding” to describe intuitive, conversational debugging, I see it as just the beginning.
For professional development, I’m seeing the rise of Verified Spec-Driven Development (VSDD), where the verification is added to the SDD part to solve the hallucination problem.
The future of software isn’t just about generating code; it’s about building the ironclad loops that prove the code works. The age of verification is here and I cannot be more excited to see what will come next.