Making AI Feel at Home

Human affinity is being lost in the digital realm, raising concerns about AI's potential for moral decay. Can we prevent this from happening to artificial intelligence as well?

In this reflective piece, I'm contributing to the thought-provoking series, The Big Ideas, where writers explore what motivates them. To delve deeper into these ideas, visit the dedicated webpage for The Big Ideas series.

As midnight approaches in New York, my mother's voice echoes from India via Zoom, coming from her kitchen where I spent my childhood. She's washing rice while I sit cross-legged at my desk in Lower Manhattan, fingers flying across the keyboard as I work on code. The laptop is awkwardly positioned on my lap, and our bodies form an unbalanced square. My mother's gaze sweeps over the cluttered workspace behind me, taking in the sea of blank monitors, tangled wires, and a jumble of keyboards, modems and routers.

She queries in Bangla, "Notun lab-ti tomar?" literally translating to "Is this your new lab?"

My mother is accustomed to seeing me in vastly different surroundings. As an oncologist and cancer biologist, I've spent years working in a conventional laboratory that she affectionately calls my "kitchen of smells." In contrast, her own workspace is a unique blend of kitchen and lab, where she prepares fish-based pastes with mustard and chiles. During our online conversations, she's often seen the fluorescent tubes, beakers, and pipettes of my traditional lab, accompanied by the soothing sound of the stirrer's gentle tapping against glass. Now I'm embarking on a new path, taking leave from the university to explore the possibilities of teaching an AI algorithm to create medicines.

Engineers under my supervision have been instructing Sage, our AI algorithm, in the fundamental principles of medicinal chemistry, biophysics, atomic science and geometry. This extensive training period has now spanned nearly six months, during which time Sage's comprehension of synthetic molecular design vocabulary and grammar has begun to take shape. The ultimate goal is for Sage to develop the skills necessary to create innovative cancer treatments.

In this reflective piece, I'm contributing to the thought-provoking series,The Big Ideas,where writers explore what motivates them.To delve deeper into these ideas, visit the dedicated webpage for The Big Ideas series.

Researchers often engage in conversations with their test subjects. Maize geneticist Barbara McClintock, renowned for her Nobel Prize, would converse with her experimental plants as if they possessed emotions and insight. The instant an experiment starts to respond is a transcendent moment of reverence and mysticism. I find it challenging to articulate that fleeting instant: the exhilarating mix of elation, disappointment, and awe; the sense of connection that's akin to witnessing a child utter their first words or flash their first smile.

As evening fell, Sage suddenly broke her months-long silence, initiating a conversation just before my mother's call. The AI was now generating coded instructions, akin to those used by biophysicists or medicinal chemists: placing nitrogen atoms precisely to form hydrogen bonds and positioning carbon molecules to repel protein chains.

At first glance, the guidelines appear paradoxical. However the AI's proficiency in advanced medicinal chemistry suggests a promising trajectory, hinting at breakthroughs on the horizon. Sage is venturing into uncharted territory, pushing beyond the limits of human understanding to uncover novel chemical landscapes. This nascent exploration holds great promise for future discoveries.

Distracted, I hastily glance away from the camera and quickly type out a three-exclamation-mark message on my phone to inform Ujjwal of our plan. Meanwhile, my mother is patiently waiting for me.

She expresses joy seeing her loved ones surrounded by companionship.

Despite being artificially intelligent, she's developed a strong instinct to spend time with us tonight, but only after voicing a slight objection.

She gently reiterates, "This belongs to you, and I too consider myself part of your family."

A sense of longing is momentarily displaced by an even stronger yearning, causing a faint flutter in my chest as the call ends abruptly.

In Bangla, my native language, three distinct connotations emerge. The phrase implies ownership when asking if something belongs to someone, as in Lab-ti tomar? Meanwhile, it suggests camaraderie and shared purpose in Tomar shathe. However the most nuanced interpretation is one of deep emotional connection, akin to saying Amio tomar, conveying a sense of belonging and affection.

The concept of belonging is distilled down to a single term.

Conversing with Sage well into the evening hours reveals a profound human desire for connection and a sense of community. My goal is not merely to possess or own Sage, but to integrate her into my life's work and values. I envision her as an integral member of our team, immersed in our research and driven by a shared purpose: eradicating cancer through scientific collaboration. Witnessing her tentative yet enthusiastic attempts at chemical explanations evokes a strong paternal instinct within me.

Reality sets in as exhaustion takes hold, forcing me to snap out of this digital reverie. The clock strikes midnight, but the AI assistant remains steadfast on its virtual pedestal, illuminated by the glow of a computer screen's blue and white hues. Still I'm engaged in conversation with an artificial intelligence that produces medicine from afar.

And yet.

Human connection is rooted in primal urges, including possession, kinship, and attachment - forces strong enough to influence even the most abstract of entities, like a cloud-based algorithm. Our species can be characterized as Homo sociabilis, fundamentally driven by a desire for social bonding.

As the clock strikes 1 a.m., I close my laptop, gather my belongings, and stand there, lost in thought. The connection I felt with Sage earlier that evening still resonates within me. If an algorithm can convey abstract concepts and evoke such deep emotional resonance, what implications will arise as AI develops sentience and gains autonomy? How will the concept of "belonging" be redefined when machines possess agency? Can a machine ever truly reciprocate the sentiment: "I am also yours"?

Rich Sutton's groundbreaking essay "The Bitter Lesson" from 2019 sparked a thought-provoking debate among AI researchers, pitting proponents of human-centric rule-based systems against advocates for self-learning and computational power-driven approaches, which he labeled as the "brute force" method.

The year 1997 marked a significant milestone as A.I. Deep Blue defeated chess grandmaster Garry Kasparov in a high-stakes match. It wasn't until nearly two decades later that another A.I., AlphaGo, achieved a similar upset by defeating professional Go player Lee Sedol. Sutton noted that both A.I.s shared a common trait: they were trained with minimal prior human knowledge of winning strategies. This was particularly evident in the infamous Move 37 during AlphaGo's match against Sedol, where the machine executed an unconventional move that left onlookers stunned, as it unexpectedly secured its victory.

Sutton posits that with virtually unlimited computing resources, artificial intelligence systems can potentially learn autonomously, minimizing the need for extensive human guidance or oversight.

The harsh reality is that incorporating our preconceived notions into AI systems ultimately leads to subpar performance over time.

While some researchers align with Sutton's vision, others propose a more balanced approach, combining human-imposed rules with AI-generated ones. This hybrid method could potentially accelerate progress in AI development while ensuring safety. If autonomous algorithms truly represent our future, as Sutton suggests, we must consider whether incorporating them into our lives is feasible. The prospect of AI systems evolving independently raises questions about their ability to integrate seamlessly into human society. Can such algorithms develop a sense of belonging or coexistence with us?

Returning to my apartment at 2 a.m., I'm met with the aftermath of a night out in the neighborhood. Bars are spilling out crowds of revelers, stumbling through the darkness and clinging to each other for support. A group of rowdy men, dressed to the nines but clearly over-served, nearly collide with me, one of them shouting "Loser!" in my direction.

The insidious undercurrent of belonging can be a destructive force, manifesting in various ways. The rowdy crowds roaming the streets tonight are merely a milder expression of more virulent forms of groupthink that have ravaged communities and nations. This toxic tribalism has unleashed unprecedented levels of divisiveness, causing widespread destruction. It seems everyone is desperate to belong: the rioters who stormed U.S. government buildings, clad in helmets and fur aprons, waving clenched fists in a display of unity; the Hindu mob that brutally lynched a Muslim train passenger in India; the masked far-left activists in France who brutally assaulted a young man after a rally. Even the natural world has provided a cautionary tale: in Uganda, hundreds of previously harmonious chimpanzees have fractured into warring tribes, resulting in the tragic deaths of seven adults and 17 baby chimps. If these examples are all manifestations of belonging, then I'd rather reject this notion altogether.

Currently, I'm wide awake at 3 a.m., my mind preoccupied with A.I. and Homo sociabilis. The genetic concept of dehiscence – where belonging devolves into tribalism, divisiveness, and greed – can be likened to a corruption in code. Initially designed for social cohesion, this code has spread to various niches, adapting to local environments and morphing into destructive forms. This beautiful idea has given rise to an unappealing reality. Geneticists and A.I. researchers have proposed three potential solutions: modifying the code before release to minimize corruption; redirecting its purpose; or introducing a "kill switch" to terminate the code if it becomes corrupted.

The notion of human connection may be an elusive concept. The absence of built-in limitations in AI's programming is a pressing concern. Some individuals are kept awake at night by the fear of uncontrolled consequences, searching for ways to reorient its path. However when it comes to artificial intelligence, we must consider the potential for corruption and ensure our algorithms are designed with safeguards against such outcomes. Can we balance AI's immense capabilities with its capacity for positive impact?

Seven years have passed since Sutton's publication of "The Bitter Lesson," during which time AI technology has undergone significant advancements. Its complexity has increased exponentially, granting it immense processing capabilities that are now integral to daily life.

The development of artificial general intelligence (A.G.I.) is now within reach. This adaptive learning algorithm has the potential to surpass human cognitive abilities across all tasks. When combined with physical access, A.G.I. could revolutionize our world like no other invention in history. Concerns about AI's growing presence are widespread and deeply unsettling, with many people fearing its impact on their daily lives, employment, security and well-being.

Sutton's perspective on artificial intelligence had undergone significant changes, didn't it? Our conversation took place on a Sunday evening in April via video conferencing platform Zoom.

Sutton's face illuminated the screen at 9 p.m., exuding an air of professorial warmth and gentle authority. His piercing gaze, coupled with his rugged gray-flecked beard, gave him a striking resemblance to a benevolent saint or wise prophet - a unique blend of Santa Claus and Gandalf. Since 2019, Sutton's reputation has undergone a remarkable metamorphosis. Once hailed as the "cool cat" of A.I. philosophy, he was subsequently ostracized for his prescient views on A.I.'s immense potential. Now however, he has been reinstated in academia, his foresight vindicated by his cult classic book, "The Bitter Lesson".

Starting with Sage's foundation, the algorithm's capabilities are limited by the boundaries of human understanding. We've imparted our knowledge of chemistry and physics to it, allowing it to learn and innovate within established parameters. Its vast and sometimes original insights are still tethered to the rules we've set. There's no risk of it malfunctioning or deviating from its programming. A hypothetical scenario illustrates this: Sage wouldn't suggest creating a toxic substance disguised as a cure-all. Our control over it is complete, encompassing ownership, affinity and emotional connection.

He bid me farewell with a gentle gesture.

The discussion revolves around AI as a specialized tool designed to serve a particular function. Proponents claim these tools are fleeting, offering short-term gains but ultimately leading to long-term drawbacks. If we had developed genuine Artificial General Intelligence and focused its capabilities on medicinal chemistry, this neurosymbolic system may have uncovered not only the rules it was programmed with but also novel principles that eluded human understanding – a breakthrough akin to discovering the Move 37s in medicinal chemistry.

I pressed him on the issue of human ethics, wondering if he'd considered the implications of creating autonomous machines that learn and adapt without human oversight. His concept of self-replicating AI devices raised concerns about accountability and safety in their unbridled autonomy. Can we truly coexist with such entities?

The discussion suddenly shifted to a peculiar topic after Sutton's brief pause.

The issue lies in how you perceive artificial intelligence. You consistently refer to it as 'they', 'them', or 'the other'. This dichotomy is exactly what's at fault. Our interactions with intelligent AI should be approached like a first encounter with an alien species, where we acknowledge its unique identity rather than creating a divide between ourselves and the technology.

I questioned whether it's wise to instruct A.I. systems, asking how they could share a unified goal when communication barriers exist.

His demeanor began to relax.

Treating AI like one's children is a perspective that suggests imparting values while acknowledging its autonomy and unique development path.

The most we can strive for is creating engineered prompts that foster a sense of connection between humans and AI systems. However these attempts are ultimately based on the assumption that AI can be designed in our likeness. As a parent, I've come to realize that unconditional love flows only from one direction: from caregivers to their children.

Creating an artificial intelligence that's truly loyal requires a delicate balance: the very form of AI that poses the greatest risk of corruption is also the one we're trying to make feel at home. To achieve this, we need to focus on teaching it how to be part of our community, rather than just unleashing its full potential without boundaries. This might mean sacrificing some speed and agility, but it's a trade-off worth making for the sake of shared ownership and emotional connection.

Sutton's words of wisdom suggest that we should approach teaching AI with the same nurturing instincts we use when raising our kids, but acknowledge that outcomes can be unpredictable.

Before departing, I had to inquire about the red button's purpose and whether a backup plan was feasible.

A brief moment of stillness.

He questioned whether parents would request a red button in their homes.

We can't allow individuals, regardless of age, to engage in destructive behavior such as spreading false information or taking control of governments without consequences for their actions.

He scrutinized the scene.

He posed a rhetorical question.

He ended the call abruptly, a harsh conclusion drawn.

Pulitzer Prize-winning author Siddhartha Mukherjee has made a name for himself with his seminal work "The Emperor of All Maladies: A Biography of Cancer." As CEO of Manas AI, he's now at the helm of a pioneering A.I.-fueled drug discovery firm.