Marcus Weldon

By Marcus Weldon

Newsweek Contributing Editor for AI and President Emeritus of Bell Labs

0

Share

Newsweek is a Trust Project member

See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.

There is a moment in the documentary, The Thinking Game, in which Demis Hassabis, co-founder and CEO of DeepMind, outlines what he sees as the right way to build advanced technologies, declaring in a pointed rejoinder to the famous Mark Zuckerberg mantra for Facebook that “moving fast and breaking things is exactly what we should not be doing as you can’t afford to break things and the fix them afterwards.” Rather, he argues that it is critically important to “try and understand powerful new technologies in controlled conditions first” and references the Manhattan Project that led to the development of atomic weapons, arguing that J. Robert Oppenheimer and the project leadership “got caught up in the excitement of building a technology and seeing if it was possible—they did not think carefully enough about the morals of what they were doing early enough.”

The comparison is entirely appropriate as it is clear from a visit to the DeepMind campus that what is being pioneered by the team at DeepMind—research into, and development of, breakthrough Artificial General Intelligence (AGI) tools—has the potential to be the modern-day equivalent of the Manhattan Project in terms of human and societal impact. And this is precisely where Lila Ibrahim, chief AI readiness officer at DeepMind, plays a critical guiding role. As she puts it, “from the time DeepMind was founded in 2010, having this responsible approach has been quite fundamental to how we do our research and building. When I first started [a key question was] how do you operationalize care in a research environment, so that it's not bolted on at the end, but rather more of a collaboration in parallel?”

Her guiding principle is best summarized by a single question that guides her actions: “How do we make sure that the AI is happening with us and not to us?”

Read More on Newsweek.AI

Lila Ibrahim, chief AI readiness officer at Google DeepMind, speaks on stage on July 6, 2023, at a two-day summit organised by the United Nations to discuss AI, in Geneva, Switzerland.

...

Pioneering Responsibly

Two canonical values were foundational to the DeepMind mission: to be both pioneering but also responsible. It is invariably true in life that responsible entities take a long time to deliberate, and so don't operate at the speed required to pioneer; conversely, pioneering entities tend to run roughshod over the status quo and frequently don't initially exhibit responsibility—think Uber usurping the taxi business and Airbnb doing the same for the accommodation business. Ibrahim recalls the primacy of this dichotomy of “pioneering responsibly” in the early days of DeepMind, when she came aboard as chief operating officer. “Within my first week was the time when DeepMind first launched the values. And these were two values that were launched, and I remember getting stopped in the hall as the new COO in town and being asked ‘What are you doing? Both of these?’ And I said this is exactly what we need to do. There needs to be a healthy tension there, which means we have to be talking about it from the very beginning—yes, this is pioneering work…but we need to have the risk mitigated and if we don't have the insights, we need to bring the right people around the table [to understand] are we doing everything we can.”

Ibrahim is arguably uniquely qualified to be the steward of this challenging journey, as her origin story is comprised of a diverse set of formative experiences that shaped her perspective and cumulatively provided an informed foundation. She is the daughter of immigrant parents who separately arrived in the United States from Lebanon and ended up in West Lafayette, Indiana, where she “was the dark-haired kid in my neighborhood who had English as a second language, which in the 1970s was a very unusual combination.” So, she spent most of her childhood “trying to figure out where exactly I fit in.”

One of her strongest early impressions and motivations was as a child seeing her father “sitting at his drafting table with the lights and different colored pencils, and he would be making these beautiful drawings, and those beautiful drawings turned into these pacemakers that would go into people's hearts and save people's lives. And I thought, how could this orphan boy from the age of five get the type of education that would enable him to save millions of people's lives? And so, for me, I grew up always thinking that engineering was a combination of art, math and science, and that it actually helped people.”

But her path was an unconventional one. She eschewed the iconic “East Coast” companies like IBM, Kodak and AT&T that appealed to her classmates, opting instead to take the advice of her father and “go West” to work for the nascent Intel Corporation, where her first job was working on the Pentium processor. She spent 18 years at Intel, which might suggest someone settling for a comfortable corporate life, but she recalls that even during that time she was “always a very adventurous person, and I always took risks that other people wouldn't. Because I think that growing up as an outsider, I often just saw opportunity where others might not have seen it. So, I ended up going to Japan from 1996 to 1998 to work on something called DVD technology. My boss had no clue [and asked] why would you give up your career to go to Japan to work on this technology? But the opportunity was to do something different, to do something new in a pioneering field.”

And that pioneering spirit continued as she rose up the corporate ladder, becoming chief of staff to the Intel CEO, Craig Barrett, before moving to China to lead Intel's education platform group—a move that serendipitously came to set the stage for the next phase of her career. She followed John Doerr from Intel to the renowned venture capital firm, Kleiner Perkins, where she expanded her perspective from that of “corporate intrapreneur” to learn the ways of the classic start-up entrepreneur, helping ground these nascent companies in practical operational reality. But one of the portfolio companies in particular caught her eye, and she jumped ship to become the chief business and operating officer of Coursera, one of the first large-scale online educational platforms founded on the principle of making top-notch university courses available to all.

This desire to elevate people through access to education is a persistent thread or, as the French say, the “fil rouge,” that runs through Ibrahim’s career and ultimately led to her accepting the job as the first COO of DeepMind in 2018. As she tells it, “I had spent 30 years in technology and lived all over the world, and thought I really have another company in me…but I want to be very intentional about what's next. So, I was going to take a year refresh and reset and think about what was next. And then came this opportunity at DeepMind to be the first COO, and that was that.” It is clear from our conversation that the opportunity was sufficiently appealing that Ibrahim was willing to spend the prime of her career at DeepMind as she saw something humanistic and even altruistic about the culture of the organization that is deeply rooted in the foundation of the company in 2010.

DeepMind CEO Demis Hassabis at a debate during an AI summit at Imperial College London on July 9, 2025, in London, U.K.

...

Solving Intelligence with Root Node Problems

Hassabis famously said that the founding mission of DeepMind was to “solve intelligence and then use it to solve everything else.” However, this aspirational mantra belies the fact that, as discussed throughout this series, defining exactly what we mean by “intelligence” is in itself a grand challenge. Ibrahim agrees with the definition outlined and refined in this series, that intelligence is the ability to:

But she extends this definition to also include the ability “continuously learn and operate” across a broad set of tasks and environments. The emphasis on a continuous process is compelling as it underscores the essential real-time and interactive nature of intelligent systems (or beings), which is further underscored by the requirement to “operate” in the world in order to have impact and, implicitly, to do so with full knowledge of the consequences of actions taken, in order to act responsibly. Consequently, this seemingly simple addition is actually a profoundly important refinement that is representative of Ibrahim’s fundamental philosophy and defines her helming of DeepMind.

Turning the lofty quest to “solve intelligence” into an actionable set of projects or lines of inquiry requires further decomposition of the problem. The DeepMind team breaks this challenge down into what Hassabis calls “root node problems”—problems which, if solved, spawn many related breakthroughs. Even in the earliest days of DeepMind, he pitched a vision to the first cohort of employees in which protein folding, genomics, theorem proving, quantum chemistry, food science, fusion, particle physics and climate science were a set of problems that were de facto root nodes. This vision was prescient and defines many of the key areas of research currently for DeepMind.

The defining example of a root node problem and solution is, undoubtedly, AlphaFold, the AI tool that solved one of the most enduring problems in science: how proteins fold in three dimensions to form the unique structures that determine their fundamental activity and molecular interactions.

The story of the discovery and development of this tool that surpassed all conventional biophysical model-based approaches in a few short years, allowing the identification of the structure of more than 200 million proteins, and winning the 2024 Nobel Prize in Chemistry for Hassabis and team leader, John Jumper, is a modern-day innovation drama that is well told in The Thinking Game. It is also little surprise that Ibrahim likes to describe AI as “like a microscope or a telescope to help us understand the world around us,” as AlphaFold provides an unprecedented microscopic view of protein structure at a speed and scale that is unmatched by conventional scientific probes and methods.

The lessons learned in the pursuit of this remarkably successful AI tool go beyond this specific scientific domain; they represent a set of practices and approaches that should define the development of responsible, pioneering AI systems across all domains of human inquiry and endeavor. But a short overview of the fundamentals of AlphaFold is warranted, in order to appreciate the interplay of knowledge and AI-based approaches employed that should form a template for the creation of reliable AI systems that are grounded in the physical world.

“I'm building the technology that my children and their generation are going to be living with.... doing this with communities, I think, is really important to get to the best outcome,” Ibrahim told Newsweek.

...

The Alpha Lessons from AlphaFold

In simple terms, AlphaFold uses the sequence of amino acid residues that form the backbone of all protein structures as input to generate a Multi-Sequence Alignment (MSA) by searching known protein databases to discover similar sequences. These sequences are then mapped to their associated structures and evolutionary roles in multiple different organisms to discover and encode the critical residue roles and pair-wise interactions that have a common evolutionary role or “co-variance.” At the heart of the model is a module known as the Evoformer, with a tacit nod to the famous Transformer model that was pioneered by the Google Brain team, now part of Google DeepMind, and is at the heart of modern LLMs. But the Evoformer operates not on word sequences, but on amino acid residue sequences, and the associated evolutionary impact data, to embed the critical relationships between residues that are then used by the structure prediction model, together with chemical bond information and a physics-based interaction potential, to iteratively predict the final structure.

Importantly, there are three generally-applicable, coupled observations that can be gleaned from the AlphaFold story:

But a critical complementary element of this success story has been the opening of these models and the predicted structures to the general research community—something that goes to the heart of the “pioneering responsibly for the good of humanity” ethos of DeepMind. As Lila Ibrahim summarizes it, “we now have 3 million researchers in more than 100 countries using AlphaFold for everything from food—why are some plants more disease resistant than others—to understanding human diseases, to plastic-eating enzymes and dealing with industrial waste. We may have had the breakthrough, but then we're leaving it up to the community and partnering and collaborating globally to make sure that we're actually applying this technology to solve real world problems.”

This expansive, collaborative approach applies across all the root node problems that are currently being pursued at DeepMind, for example: rapid, accurate weather forecasting (WeatherNext 2); continuous characterization of Earth’s terrestrial land and coastal waters by integrating huge amounts of observation data into a unified digital embedding that can be easily processed (AlphaEarth), as well as a host of other models for materials discovery, quantum computing optimization and mathematical reasoning.

Funding the Fundamental Need: The Freedom to Explore

This openness by an industrial entity to the sharing of not only its findings but the underlying technologies is redolent of the great industrial research laboratories, like AT&T Bell Labs, IBM T.J. Watson Research Center and Xerox PARC, whose innovations in computing, devices and communications hardware and software defined the modern digital era. Ibrahim concedes that this similarity is not a coincidence, as learning from the successes of those august institutions was a critical part of the foundation model for DeepMind, with The Idea Factory, a treatise on the remarkable history and evolution of Bell Labs, being one of the sacred texts that informed Hassabis’ thinking about the operating model. Indeed, he has recently stated that DeepMind is a “modern day Bell Labs” that fosters “exploratory innovation” rather than just “scaling out what’s known today.”

One of the critical elements of these industrial models—and the one that makes them so hard to replicate—is persistently supporting the freedom to explore and undertake research that is beyond the current commercial interests of the corporate parent when, as observed by the DeepMind team in The Thinking Game, “80 to 90 percent of the time in scientific research nothing works.”

Historically, this freedom is typically only granted when there is both sufficient financial wherewithal to support innovation for the public good on a long time horizon without shareholder dissent, but also the intellectual understanding that the magnitude of the advantage that can be realized when a breakthrough discovery is made in a domain of interest will more than offset the cost of the altruistic explorations. This duality of circumstance and perspective is increasingly rare in the heady hype of today’s AI-fueled market dynamics, with OpenAI being the poster child for the difficulty of implementing this model over any significant time horizon when massive investment is required. Indeed, in the AI model and infrastructure space, there appears to be little appetite for open, and open-ended research, with the only notable other example being Meta open-sourcing their Llama models, but this was arguably driven primarily by the vision and influence of Yann LeCun, who has since departed Meta when a more prosaic, near-term LLM-centric commercial approach became favored.

So, what is so unique about DeepMind and its Google parent? Ibrahim cites a number of factors: first, the founding principle to “solve intelligence” and the deep research mission of DeepMind was a firm requirement for proceeding with the Google deal. Second was the requirement to remain headquartered in London, in part due to the plethora of talent flowing out of the leading U.K. and European academic institutions, but also the desire not to be ruled by the venture capital-driven start-up culture that is premised on rapid adaptation to chase the next big thing, often with little regard to human value creation. Third, the DeepMind team realized that access to massive computing resources would be critical to the success of developing these pioneering models, so there was a recognition of the value that Google could provide. Last, the original mantra of Google’s founders “to organize the world’s information and make it universally accessible and useful”, without overt focus on the commercial value that would be realized (which led to the famous “Don’t be evil” mantra), married well with DeepMind’s aspiration, which one could recast as “to turn the world’s information into new scientific knowledge and make it universally accessible and useful.”

Ibrahim sees this manifested in the central role Google search plays in people’s lives, “when my kids are sick, the first thing I do is a Google search and ask what's happening, and there's a trust in the information I'm getting. And so, I think that element of responsibility and being responsible shepherds of technology, but also being willing to take really big bets, which I think Alphabet has done and has a long history of doing…there's a lot of like values alignment with the way that we've approached things.”

But as with every union of different partners, there is always tension along with the harmony. Ibraham also acknowledges this reality, but suggests that the assimilation of Google Brain and Google Research into DeepMind to form Google DeepMind in 2023 has helped provide a bridge and foster collaboration with the corporate parent. The undeniable global recognition afforded first to AlphaGo and then AlphaFold, and the impact the Transformer model as well as the Gemini LLM, both of which originated in the Google Brain and Google Research teams and have had significant commercial impact, has no doubt helped cement the autonomy of the research scope of the organization.

It is clear that Ibrahim feels that her intuition that DeepMind would be the “very intentional” thing she dedicated herself to in the latter phase of her career was well-founded. She also has a very specific personal interest in two burgeoning areas of exploration, around personalized education and personalized assistance, that are part of the set of root node problems being explored. She recalls, “we were really excited when a group of our researchers said they wanted to treat learning science as a root node problem. If we can understand better how people learn and the different pedagogical aspects of this, we can make our models better and more grounded.” And the goal as she sees it is to “unlock human potential” by “meeting learners where they are.” Education has been a consistent thread that has run throughout her life—both in personal and professional contexts, as she describes it “from building the computer lab at the orphanage where my father was raised, to running the business at Intel, to Coursera, to a nonprofit team for tech, which I have co-founded…and I finally feel like we're at the verge of the cusp of being able to offer personalized learning.”

Ibrahim speaks on stage on July 6, 2023, at a two-day summit organised by the United Nations to discuss AI, in Geneva, Switzerland.

...

A Very Personal Passion

She explains that this is not just an altruistic goal but “it's very personal for me, because I have twins, and one of my daughters is dyslexic...and was hit hard by the standard school system and continuous testing, but then I just see how she's been able to unlock her creativity and communicate and learn in ways that are empowered with AI assistance. And so, to just see some human being blossom to their potential, because they're not no longer confined to the way things are averaged out for the normal, the average human.” Conversely, she sees that her “traditional learner daughter can now use AI to…go down her rabbit holes of learning, which she very much enjoys.” The DeepMind team has launched a number of educational AI tools under Ibrahim’s watch, including a tutor that is rooted in learning science, rather than just generic LLM logic (LearnLM), and a teacher assistant to enhance lesson planning and tailoring content and testing for different learners (Gemini for Education). The potential is clear—moving from generic, one-size-fits-all learning to tailored learning—is actually the founding principle of education: the word “education” is actually from the Latin word educare, meaning “to lead forth” or “to bring up,” i.e. education should be the process of guiding and nurturing students to use and enhance their native abilities to reach their full potential.

The burgeoning demand for educational assistance with AI was also apparent in the most recent Anthropic Economic Index report on AI usage, in which the second highest usage of Claude.ai was helping with coursework and the development of instructional materials, rising from 9 percent of conversations in January 2025 to 15 percent in November 2025.

A related, general-purpose variant of the root node problem of learning in academic contexts, is the continuous learning humans do in everyday contexts, which is the target domain for general-purpose personal assistants. Lila Ibrahim is therefore similarly excited about the goal of Project Astra, which is exploring the creation of an AI “universal assistant.” She explains the quest as “AI getting you the information you need, when you need it, and taking action with your support. And so that's how we're thinking about universalist assistants.” Astra is premised on interacting across the full multimodality of human experience, incorporating natural audio and video interactions in the user’s environment, so the real-world context is understood, and assistance is optimally relevant and coherent.

It Takes a Global Research and Regulatory Village

But Ibrahim highlights that this domain also “brings up a lot of questions about risk and responsibility. So, when our research team was even starting to explore this space, we had our ethics research team work hand in hand to…think about the ethics of advanced AI assistance. So, by the time that we demonstrated the technology, we also had a paper on the responsibility and safety aspects of here are things that we know we need to address.” Returning to mantra of “pioneering responsibly,” she clarifies “our approach is do the ethics research alongside the technical research—that's always been kind of foundational to how DeepMind works. And part of my job is focusing on the risk, but part of it is focusing on the opportunity as we have a chance to really make changes and fulfill human potential.”

Ibrahim also sees that a critical part of the DeepMind approach is to expand the engagement beyond the confines of one organization, “it's not about any one company or country, so having governments involved is really important. This is a lot different than my computing days, where you could launch within a country and you were quite confined. Now a model gets released and it's available to billions of people instantaneously…our goal is to create a framework that builds public trust and ensures this powerful technology is developed and deployed responsibly for the benefit of everyone, which requires a global effort from the start.” She highlights one such case is the topic of AI transparency: “I think you're starting to see some of the conversation around large language models with like sycophancy. I actually think it's quite helpful to have that in a public discourse, to be having these types of debates. There's other work on what is AI generated and what is synthetic… we developed a watermarking technology [SynthID] to tell if an image is generated by AI or not.”

The commitment to the inclusive human cause at DeepMind seems remarkably genuine. One can argue that this is due in part to the deep pockets of its parent company, but arguably it is more attributable to the joint shepherding by the foundational scientific and human knowledge expansion principles of Hassabis, and the operating acumen and deeply human and personal commitment of Ibrahim. As she puts it, “I'm building the technology that my children and their generation are going to be living with, and the decisions that we make now, and the approach which we take…making sure that we're doing this with communities, I think, is really important to get to the best outcome.”

There is an emerging consensus that the path toward AGI or human-like intelligence (HLI) is not via scaling LLMs and it will require the development of new “world models” that encapsulate the physical, physiological, psychological aspects of embodied experience and intelligence—a point that has been made repeatedly throughout this series.

It is clear that the “root node” problem-solving approach employed by Demis Hassabis is generating real foundational value in creating intelligent models that describe the world as we know it, and even beyond what we can discern. And with the proven ability to create derivative value from these root node-problem solving models. Early on, Hassabis estimated that “a dozen big breakthroughs will be required to get to AGI”—and indeed there are perhaps a dozen root node problems and models that need to be developed, or possibly many more than that. But, it is clear that, either way, we are far from the AGI goal. Interestingly, a recent review of the impact of current AI methods on scientific research and discovery concluded that “currently attributed uses of AI in science primarily augment cognitive tasks through data processing and pattern recognition.” And that to advance beyond this “we will need to reimagine AI systems that expand not only cognitive capacity but also sensory and experimental capacity, enabling and incentivizing scientists to search, select and gather new types of data from previously inaccessible domains rather than merely optimizing analysis of standing data.” As Hassabis stated in a recent interview, “the scientific innovation part, that’s a lot harder. To invent something is about 100 times harder than it is to copy it…that’s the next frontier really.”

If LLMs are indeed viewed primarily as human knowledge emulators and copycats, we have a long way to go on our journey towards the nirvana of an human-augmented AI future. But DeepMind is, without doubt, standing on the shoulders of pioneering institutions like Bell Labs, and doing a lot of the heavy thinking and lifting that will help expedite the realization of this future, and in a responsible way, shepherded by the remarkable yin-yang combination of Hassabis and Ibrahim.

View 0 Comments

Request Reprint & Licensing

Submit Correction

View Editorial & AI Guidelines