Cognition AI's Silas Alberti: training runs 'always on'

Cognition AI's approach reflects a broader shift in how AI agents are built. Rather than treating training as a discrete phase followed by deployment, the company keeps training running continuously, using real-world experience to improve subsequent versions of Devin. Alberti described this as always trying to find the next data and the next reward signals to improve the models.

That continuous loop places new demands on the infrastructure underneath. Cognition has spread training across data centers in multiple countries and continents, so uptime across thousands of graphics processing units becomes essential. Alberti noted that if just one replica goes down, the whole training run goes down, which explains why 99.99% reliability is such an important metric.

The outcome likely hinges on whether that reliability can be maintained as the system scales, and on how quickly new hardware translates into better price-performance. For teams building or buying AI agent infrastructure, the question is whether continuous learning becomes a standard expectation rather than a differentiator.