Smart Architecture over RAW Compute: Deepseek breaks the approach “greater is better” to the development of AI

The AI ​​story reached a critical point of folding. Thehe Deepseek breakthrough -The up-to-date results without relying on the most advanced chips-what many of NEVERIPS had already announced in December: The future of AI is not related to throwing more problems with problems-the resumption of resumption is How these systems work with people and our environment.

As an educational in Stanford a computer scientist who witnessed both the promise and the dangers of the development of AI, I see this moment as even more transformative than the Chatgpt debut. We are entering what some call the “renaissance reasoning”. Openai’s O1Deepseek’s R1 and others pass along the scale of brute force to something smarter-and do it with unprecedented efficiency.

This change cannot be more timely. During his main note, a former chief scientist of Openai Iliya Sutskever declared This “pre -examination will be completed” because as the computing power is growing, we are limited by end internet data. Deepseek’s breakthrough confirms this perspective – China’s researchers have achieved comparable Openai O1 results with some of the costs, demonstrating that innovation, not just harsh computer power, is the way forward.

Advanced AI without mass pre -training

World models are activated to fill this gap. World Laboratory World Laboratory $ 230 million increase To build AI systems that understand reality as people parallel to the Deepseek approach, where their model R1 shows “AHA!” Moments-shaking to reassess problems, as people do. These systems, inspired by human cognitive processes, promise to transform everything-from environmental modeling to interaction between man.

We see early victories: Meta’s last update to their Smart glasses Ray-Ban Allows continuous, contextual conversations with AI assistants without alert words, along with real -time translation. This is not just an update of features-this is a visualization of how AI can improve human capabilities without requiring massive pre-trained models.

However, this evolution comes with nuanced challenges. Although Depepeek drastically reduced costs through innovative training techniques, this breakthrough of efficiency can lead to paradoxically lead to an increase in overall resource consumption – a phenomenon known as Jevons ParadoxWhere technological efficiency improvements often lead to an increase, not reduced resources use.

In the case of AI, more cheaper training can mean more models that are trained by more organizations, a potential increase in net energy consumption. But Deepseek’s innovation is different: demonstrating that the most modern performance is possible without avant-garde hardware, they not only make AI more efficient, they change radically how we approach the development of the model.

This switch to smart architecture over the harsh computing power can help us escape from the Jevons Paradox trap, since the focus moves from “How much calculation can we afford?” How intelligent can we design our systems? “As the professor notes notes From UCLA Guy Van Day Breck, “the total cost of the language model’s reasoning is certainly not diminishing.” The environmental impact of these systems remains significant, with the industry to more efficient solutions – the kind of innovation that Depepeek represents.

Prioritization of effective architectures

This change requires new approaches. Deepseek’s success confirms the fact that the future is not in the construction of bigger models -but for the construction of more intelligent, more efficient, which work in harmony with human intelligence and environmental restrictions.

Meta Yann Lecun’s chief scientist predicts future systems Spending days or weeks thinking through complex problems, like people. The Deepseek’s-R1 model with its ability to pause and review approaches is a step towards this vision. Although resource-intensive, this approach can lead to breakthroughs in climate change decisions, healthcare innovations. But like Carnegie Melon Amet journal Wise wisely, we must question anyone who claims where these technologies will lead us.

For enterprise leaders, this change presents a clear path forward. We need to prioritize effective architecture. One who can:

• Deploying chains by specialized AI agents, not single massive models.
• Invest in systems that optimize for both efficiency and environmental impact.
• Construction of infrastructure that maintains iterative development of the person in the outline.

Here is what I am excited by: Deepseek’s breakthrough proves that we go through the era of “Greater is better” and in something far more interesting. With the participation of their borders and innovative companies that find new ways to achieve more with less, there is this incredible space that opens up for creative solutions.

The intelligent chains of smaller, specialized agents are not just more efficient -they will help us solve problems in ways that we have never imagined. For starters and businesses wishing to think differently, this is our moment to have fun with AI again, to build something that actually makes sense for both humans and the planet.

Kiara Nirgin is a prize -winning technologist of Stanford, author of bestsellers and co -founder of ChimaS