Modern artificial intelligence systems have achieved something once considered impossible. They can write essays, generate software, translate languages and perform at levels that rival humans on increasingly complex tasks.
But Israeli AI researcher Gilad Levy believes the biggest challenge is not making models bigger. It is teaching them how to keep learning.
Gilad Levy
(Photo: Omer Hacohen)
Levy, who began studying computer science at 15 and completed his degree at 17, is now leading Manifold, an Israeli AI research lab focused on what it describes as a fundamental limitation of today’s artificial intelligence: once training ends, the model’s understanding of the world effectively stops evolving.
The company’s central idea is that intelligence is not only the ability to absorb enormous amounts of information, but also the ability to continuously adapt as the world changes.
That challenge has become more apparent as AI systems move beyond chatbots and into areas such as software development, cybersecurity, robotics and autonomous systems.
Current AI models are trained on massive datasets, often containing more information than any individual could process in a lifetime. But after that training process is completed, new experiences do not automatically become part of the model itself.
A person encountering a new object, learning a new word or discovering that a familiar assumption is wrong can immediately incorporate that experience into their understanding of the world.
AI systems work differently.
After training, new information can be provided through prompts, external databases, retrieval systems or previous conversations. But those inputs do not change the underlying model. Once the interaction ends, the model itself remains the same.
Developers have attempted to address this limitation through tools such as retrieval systems, memory features, fine-tuning and longer context windows. These methods can make AI systems more useful, but they do not fundamentally allow the model to learn continuously from experience.
Manifold was created to tackle that underlying problem.
The company is developing what it calls a neuroplastic AI foundation model, designed to adapt through interaction with the real world. The goal is to move beyond systems that are trained once and then deployed toward architectures that can continue updating their understanding without repeatedly rebuilding the model from scratch.
Levy’s interest in AI architecture began early. At 18, he co-developed and taught what the company describes as the world’s first academic course on Transformers, the technology architecture that became the foundation of modern generative AI, alongside professors Yonatan Belinkov and Mike Schuster.
At a time when many of his peers were still choosing their university fields, Levy was teaching advanced AI concepts to students with significantly more professional experience.
Now, he is revisiting the same technology from a different perspective.
Levy does not dismiss the achievements of Transformers. The architecture enabled AI systems to train at unprecedented scale and produced major breakthroughs across language, vision and other fields.
His argument is that scaling existing approaches may not solve the next fundamental challenge.
The future of AI, he believes, depends on systems that can adapt after deployment rather than simply becoming better at processing information from the past.
Building such foundational AI research in Israel presents its own challenge.
The country has built a strong technology ecosystem, particularly in cybersecurity, enterprise software and infrastructure. But foundational AI research requires longer timelines, significant investment and a willingness to pursue uncertain questions that may not immediately produce commercial products.
Levy argues that without companies willing to challenge the underlying architecture of AI, Israel risks becoming primarily a consumer of technology developed elsewhere rather than a place where the next generation of AI systems is created.
Manifold is attempting to prove that deep AI research can be built locally within a sustainable business model.
The company’s approach represents a different direction from much of Silicon Valley’s current AI race, which has focused heavily on larger models, more computing power and ever-expanding datasets.
The next stage of artificial intelligence, Levy argues, may depend less on how much information a model can absorb and more on whether it can continue learning once it enters the real world.
If successful, that shift could change not only how AI systems perform, but how researchers define intelligence itself.