“Enterprises need to move to adaptive, energy-efficient AI models”

However, since late last year, a growing number of AI researchers have announced in unison at conferences and rolling halls at X that current cutting-edge AI models may have reached their limits. The world needs new, innovative ways to build artificial intelligence models.

Sarah Hook is one of them. In October 2025, she resigned as Cohere’s vice president of artificial intelligence research to launch her new startup, Adaption Labs, with Sudip Roy, another Cohere veteran. Hooker, who studied Google’s AI models before joining Cohere, firmly believes that future AI systems will use less computing power, have lower running costs and be able to adapt to user needs. In February of this year, Adaption Labs raised $50 million to build thinking machines that adapt, continuously learn, and are energy-efficient.

Hook was in India for the AI ​​Impact Summit and discussed this and other emerging ideas in the AI ​​field with Hindustan Times. Edited excerpts:

Is it possible to replace the ever-expanding AI models?

The advances in AI over the past 10 years have been all about larger and larger models, as it has become a very predictable formula for success in all AI frontier labs (Open AI, Human, Google, Meta, Amazon, etc.). These AI models require more computing resources, not only during training but also when providing answers to users.

Two things are changing, and tech companies need to adapt. First, businesses recognize that continually spending large amounts of money on a simple query is unsustainable. Another is that the Transformer architecture on which these models are based is now saturated. Cutting-edge AI companies aren’t getting triple or quadruple performance gains by adding more compute. The next era of progress will come from creating models that can interact, adapt to the world, and use computing efficiently.

What you are saying is quite controversial and seems to be contradicted by others in the field. You wrote a paper about the slow death of scaling and why scaling computing is not the solution.

Logically, if the task is simple (90% of AI requests from users are simple), then the AI ​​model should use less computation. It should increase the computational effort of difficult problems. This doesn’t happen with current AI models because they are static, singular, and optimized for averages. This is why small AI models now far outperform large AI models. The next stage of artificial intelligence will be a model that changes and adapts in real time based on incoming tasks. It’s like human wisdom. We adapt and we learn.

If the transformer is saturated and adaptation is the new way forward, why aren’t big tech companies taking a new approach?

There is huge inertia in people’s understanding of different paths. Over the past decade, the pursuit of more computing has changed everything about our ecosystem. The company has siled pre-training and post-training teams. All of them are based on the assumption that if you build an all-encompassing model and throw it out into the real world, you want it to be able to do all the tasks it’s given.

Since the Transformer method has become saturated, things are changing all the time. Different companies are trying different approaches. How to adjust the model in real time? How to use different calculations on different problems? Do we need new hardware designs?

When we launched the Adaptation Lab last October, people were very skeptical, but I’ve heard the term “continuous learning” used more frequently in the industry, which basically means AI models that adapt to the user in real time. I think this is a new era and in a year it will become a dominant philosophy.

You just raised $50 million in seed funding for your idea of ​​AI model adaptation. How does this work?

We want to make the entire AI stack more flexible so that it can adapt to any task like Pay-Doh. This includes data, adaptive intelligence and the interface itself. We want to change the way people interact with models, moving away from the thumbs-up and thumbs-down approach and creating a more dynamic feedback loop.

India is the third AI Impact Summit after the UK and France. What brought you to the top?

The AI ​​Impact Summit is critical to expressing our view that this technology is global. My co-founder and I are immigrants to the United States. I grew up in Africa. Most of my work so far has been about making models multilingual, adaptable, and usable around the world. When it comes to adaptation, which is where we want to innovate, a global perspective is important to us. We want to give people the flexibility to own, use, control and shape AI.

The promise of these summits is that people are building ecosystems outside of the United States. During the summit, these countries announced government funding for artificial intelligence research in their countries, which is important for cultivating global technical talent so that people can take more ownership of their data. Our own teams are assembled remotely and globally.

Skeptics of the India AI Impact Summit say the summit in Delhi last month did not achieve much. What’s your takeaway?

Currently, AI is used around the world but created in few places, but this summit and others like it have made AI more global. This is a big catalyst.

People who are rarely in the same room together meet and exchange ideas. This creates a catalyst for technical talent that shapes innovation around the world. You should never underestimate a catalyst. What happens six months later will be even more interesting. That’s why I’m here.

You’re talking about globalization in an environment of increasing geopolitical tensions.

These two things can exist simultaneously. While geopolitical tensions exist, there can be more fluidity and discussion in the technology ecosystem. I want to focus on how we can support global talent because these technological interventions will outlast the current political tensions.

What are your thoughts on creating sovereign AI models?

National AI models are less important than the ecosystems they create. In order to build a competitive model, you must create dense talent density. Currently, there are only a few hundred people in the world (perhaps 700 in total) who understand the complete process of how to pre-train, post-train, and tune an AI model. They have many resources available.

But there is no reason to continue this way. It would be very useful if the sovereign model became a national catalyst for developing the AI ​​ecosystem.

India’s advantage is that young people have high aspirations. What India needs to do is get these people to care deeply about technological excellence, which is what innovation requires.

If the vision you’re considering is realized in the next five years, will the need for pesky just-in-time engineering disappear?

Current models don’t work well because they are optimized for the average. If you serve a large static model to billions of people, the end user (whether it’s you or the business) needs to do something to get answers from the model. This is what we call just-in-time engineering today.

A large number of complex prompts are the user’s solution to the model, fundamentally setting them up to fail. When the AI ​​model is adapted to our needs, we still need to write the first hint, but we don’t need to provide details.

We should be able to jump in, highlight the areas we think need to change, and the model will immediately shape its behavior in real time. I think just-in-time engineering is going to be phased out.

What about massive data centers? Will these disappear as models become more efficient?

Data center demand depends on the number of people using AI. As we aim to make AI accessible to everyone, we will need more data centers and more energy. We can make AI models as efficient as possible, but as usage increases significantly, I don’t think the need for data centers and energy will go away. This is a policy conversation at global and national levels.