In the last 10 years, the Artificial Intelligence industry has focused on scaling AI models, building bigger models that can process more data. To build these models, there’s also been a race to build massive data centres that offer compute for these models and which cost billions of dollars and consume as much energy as a city.
However, since late last year, a growing chorus of AI researchers in conferences and scrolling halls of X has been announcing that the current frontier AI models may be reaching their limits. The world needs new, innovative ways to build AI models.
Sara Hooker is one of these voices. In October 2025, she quit her job as VP of AI research at Cohere and started her new startup, Adaption Labs along with another Cohere veteran Sudip Roy. Hooker, who researched on Google’s AI models before she was at Cohere, is convinced that the future lies in AI systems that use less computing power, cost less to run and can adapt to the needs of users. In February this year, Adaption Labs raised $50 million to build thinking machines that adapt and continuously learn and are also energy-efficient.
Hooker was in India for the AI Impact Summit, and discussed this and other emerging ideas in AI with Hindustan Times. Edited excerpts:
Is an alternative to the ever-scaling AI models even possible?
AI progress in the last 10 years has been all about bigger and bigger models because it’s become a very predictable formula for success for all AI frontier labs (Open AI, Anthropic, Google, Meta, Amazon, etc). These AI models demand more compute footprint, not only in training but also when they serve an answer to users.
There’s two things that are changing and tech companies will need to adapt to it. One, business is recognising that it’s not sustainable to keep spending so much money on a simple query. The other is that the transformer architecture, on which these models are based, is now saturated. Frontier AI companies aren’t getting triple or quadruple performance gains by adding more compute. The next era of progress is going to come from creating models that interact and adapt to the world and use compute efficiently.
What you’re saying is quite controversial and also seems contradictory to others in the field. You wrote a paper on the slow death of scaling and why scaling compute is not the solution.
Logically, if the task is simple – which 90% of AI requests from users are – the AI model should use less compute. It should increase compute workload on difficult problems. This is not happening with current AI models as they are static, monolithic and optimised for the average. That’s the reason small AI models are now far outperforming larger AI models. The next phase of artificial intelligence would be a model that changes based on the incoming task and adapts in real time. This is like human intelligence. We adapt, we learn.
If transformers are saturated and adaption is the new way forward, why isn’t t Big Tech taking the new approach?
There is a massive inertia in people recognising a different path. In the last ten years, the drive for more compute has changed everything about our ecosystem. Companies have siloed pre-training and post-training teams. All of them worked on the assumption that if you build a model that covers everything and throw it into the world, you hope it works for all the tasks it is given.
Since the transformers approach is saturated, things have been changing. Different companies are trying different approaches. How do you adapt the model in real time? How do you spend different compute on different problems? Do we need new hardware design?
When we started Adaption Labs last year in October, people were very sceptical, but already I hear industry using the term ‘continuous learning’ more often, which basically implies an AI model adapting in real time to users. I think this is the new era and, in a year, it will become a dominant philosophy.
You’ve just raised $50 million as a seed fund for the idea of AI models adaption. How does that work?
We want to make the entire AI stack more flexible so it can adapt to any task like Pay-Doh. This includes the data, adaptive intelligence and the interface itself. We want to change how people are able to interact with the model and move away from the thumbs up, thumbs down approach to creating a feedback loop that’s more dynamic.
India’s your third AI Impact Summit after UK and France. What’s the reason you were at the summit?
The AI Impact Summits are critical to put out our point of view that this technology is global. Both my co-founder and I are immigrants to the US. I grew up in Africa. Most of my work to date has been about how we make models good at multilingual, at adapting and at being used around the world. When it comes to adaption, which is where we want to innovate, global perspectives are important for us. We want to give people the flexibility to own, work on, control and shape AI.
What’s promising with these summits is that people are building up ecosystems outside of the US. During the summit, these countries have announced government funding for AI research in their own countries and that’s important to build technical talent globally so people can have more ownership over their data. Our own team is set up remotely, all over the globe.
Sceptics of the India AI Impact Summit say that not much was achieved in Delhi last month. What’s your takeaway?
Currently, AI is used all over the world but is created in a few places and this summit and others like it have swung the pendulum to make it more global. It’s a big catalyst.
People who are rarely in the same room together, met each other, exchanged ideas. This creates a catalyst for technical talent that shapes innovation in different parts of the world. You should never underestimate catalysts. It’s what happens six months from now that’ll be more interesting. And that’s why I was here.
You’re talking globalisation in an increasingly geopolitically tense environment.
The two things can exist at the same time. There can be more mobility and discussion in the technical ecosystem, while geopolitical tensions exist. I want to focus on how we can support global talent because these technical interventions will last longer than current political tensions.
What’s your take on creating sovereign AI models?
National AI models are less important than the ecosystem they create. To build a model that is competitive, you have to create an intense talent density. Right now, there’s only a few hundred people in the world — maybe 700 in total — who know the full pipeline of how to pre-train, post-train and align AI models. They have a lot of resources available to them.
But there is no reason this remain that way. If the sovereign model becomes a national catalyst in developing an AI ecosystem, that’s very useful.
India’s strength is that its young people have huge ambitions. What India needs to do is drive these people to care deeply about technical excellence which is what you need for innovation.
If the vision you’re thinking of happens in the next five years, will the need for pesky prompt engineering go away?
The current models don’t work well as they’re optimised for the average. If you ship a massive static model to billions of people, the end user – be it you or an enterprise – needs to do acrobatics to get answers out of the model. That’s what we call prompt engineering today.
Massive complicated prompts are a user’s solution to a model that fundamentally fails them. When AI models adapt to our needs, we would still need to write the first prompt, but not give minute details.
We should be able to engage with it, highlight areas that we think need to change and the model will instantly shape its behaviour real-time. I think prompt engineering would be eliminated.
What about massive data centres? Will those also go away as models become more efficient?
Data centre need is driven by how many people are using AI. As we aim to give access of AI to all, we’re just going to need more data centres and more energy for it. We can make AI models as efficient as possible, but as there’s going to be massive increase in usage, I don’t see the need of data centres and energy going away. This is a policy conversation at a global and national level.
