Smart Architecture on raw computing: Deepseek breaks the greatest ‘approach is better’ for the development of AI

The narrative of AI has reached a critical turning point. Deepseek’s advance, achieving avant -garde performance without depending on the most advanced chips, demonstrates what many in neurips in December had already declared: AI’s future is not about throwing more calculations into the problems, it is about reinventing how These systems work with humans and our environment.

As an computer scientist educated in Stanford who has witnessed the promise and dangers of AI development, I see this moment even more transformative than the Chatgpt debut. We are entering what some call a “rebirth of reasoning.” Openi’s O1, Deepseek’s R1 and others are going beyond brute force towards something smarter, and do it with unprecedented efficiency.

This change could not be more timely. During her key neurips note, former Operai chief scientist, Ilya Sutskever declared That “prevention of preado will end” because, although the calculation power grows, we are limited by finite internet data. The advance of Deepseek validates this perspective: China’s company researchers achieved a performance comparable to OPENAI O1 to a cost fraction, which shows that innovation, not only unprocessed computer power is the way to follow.

Ai advanced without massive pretraben

World models are intensifying to fill this void. I recommend Increase of $ 230 million To build AI systems that understand reality as humans, parallel to the Deepseek approach, where its R1 model exhibits “Ajá!” Moments: Stop to reevaluate problems as humans do. These systems, inspired by human cognitive processes, promise to transform everything from environmental modeling to human interaction-AI.

We are seeing the first victories: the recent target update to your Ray-Ban smart glasses It allows continuous and contextual conversations with Wake Wordless assistants, along with real -time translation. This is not just an update of characteristics: it is a preview of how AI can improve human capacities without requiring previously trained massive models.

However, this evolution comes with nuanced challenges. While Deepseek has drastically reduced costs through innovative training techniques, this progress progress could be made paradoxically to greater general consumption of resources, a phenomenon known as Jevons paradoxwhere technological efficiency improvements often result in an increase in the use of resources instead of decreasing.

In the case of AI, the cheapest training could mean more models trained by more organizations, potentially increasing net energy consumption. But Deepseek’s innovation is different: by demonstrating that avant -garde performance is possible without avant -garde hardware, not only make AI more efficient, but also fundamentally change the way we address the development of the model.

This change towards an intelligent architecture about unprocessed computer power could help us escape the trap of the Jevons paradox, as the approach moves from “How much computation can we pay?” To “How intelligent can we design our systems?” As Ucla Guy Van Broeck teacher points out, “the general cost of reasoning of the language model is certainly not decreasing.” The environmental impact of these systems remains substantial, pushing the industry towards more efficient solutions, exactly the type of innovation that Deepseek represents.

Prioritize efficient architectures

This change demands new approaches. The success of Deepseek validates the fact that the future is not about building larger models: it is about building the smartest and most efficient that work in harmony with human intelligence and environmental limitations.

The main scientist of Meta Yann Lecun Imagine future systems Spend days or weeks thinking of complex problems, just like humans do. The Deepseek’s-R1 model, with its ability to stop and reconsider approaches, represents a step towards this vision. While it is intensive in resources, this approach could generate advances in climate change solutions, medical care innovations and beyond. But as Carnegie Mellon Diary In warning wisely, we must question anyone who reclides certainty about where these technologies will guide us.

For business leaders, this change presents a clear path forward. We need to prioritize efficient architecture. One that can:

• Deployment chains of specialized agents instead of individual massive models.
• Invest on systems that optimize both performance and environmental impact.
• Build an infrastructure that supports the iterative, human development in the loop.

This is what excites me: Deepseek’s advance shows that we are going beyond the “biggest era is better” and something much more interesting. Since the presta reaches its limits and innovative companies that find new ways to achieve more with less, this incredible space for creative solutions opens.

Smart chains of smaller and specialized agents are not just more efficient: they will help us solve problems in a way we never imagine. For startups and companies willing to think differently, this is our time to have fun again, build something that really makes sense for people and the planet.

Kiara Nirghin is a Stanford technologist award, the best selling and co -founder of Chima.