THE VIRAL TAIL
MACHINES HAVE JUST LEARNT HOW TO THINK
The dance between what we can imagine and what is possible is a delicate one. Optimism, no matter how latent in logic, often finds itself wrapped in the shameful garments of hype. Grand ideas appear absurd when measured against what we believe. This is because we comfort ourselves with the false dichotomy of being either the cautious skeptic or the gullible believer. In that false choice, we refuse to condemn ourselves as gullible, and select to be skeptical.
And you know, time is a cruel trickster: we see it in light of ourselves, but it doesn’t really care about us. We forget that reality is not beholden unto us, and we mistake the slowness in the unfurling of things for their implausibility. Periods of dormancy can stretch so long that even the most steadfast believers begin to doubt their own standing. Therefore, although skepticism is usually justified, it takes greater vision to see the inevitable, especially when it takes its time. And when the moment finally arrives, it unfolds with such sudden magnificence that it makes all previous hype seem like underestimations.
In 1950, Alan Turing put forth a remarkable prophesy. In his landmark paper Computing Machinery and Intelligence, he envisioned a future 50 years ahead, where machines could converse with humans in writing. It is here that he proposed what would become known as the Turing Test: a challenge for these artificial machine-minds to convince human judges of their humanity. To many, this projection was mere reverie, but to him, it was the culmination of the sound foundations that he had helped establish.
Fifteen years earlier, Turing’s groundbreaking paper, ‘On Computable Numbers, with an Application to the Decision Problem‘, had introduced what we now call the Turing Machine. This was an elegant mathematical abstraction that could simulate any algorithmic process. Working in parallel with Alonzo Church, Turing demonstrated that this machine could execute any computable algorithm. Unlike other thinkers of his time, Turing also chose to define intelligence purely in terms of verifiable performance. Instead of asking “Can machines think?”, he asked, “Can a machine’s behaviour be indistinguishable from that of a human?”. To him, intelligence was ultimately about producing the right responses, regardless of whether “true understanding” was present. This lead him to a profound insight: if human cognition could be distilled into algorithmic processes, then machines could replicate human intelligence.
Unfortunately, Turing’s insight was blunted by the technology of his time. The first generation of AI programs such as Logic Theorist (1956) and General Problem Solver (1959) were unremarkable and lacking in adaptability. They could prove some mathematical theorems, but nothing more. In the late 60s, SHRDLU was created. It was a remarkable demonstration of Artificial Intelligence, especially due to its ability to “understand” natural language. However, it too was completely incapable of mastering real-world tasks.
All successor programs that followed throughout the 70s suffered the same fate. As optimism waned, severe funding cuts in AI research followed, plaguing the path to Turing’s prophesy with the first AI winter. In the 80s, a brief period of spring came with the flourishing of a new generation of AI programs dubbed “expert systems”. These AIs demonstrated ability in specific tasks, but their limitations still could not be overlooked. Research funding remained scarce and a second AI winter followed. AIs remained primitive, struggling with tasks that children could master effortlessly, such as basic arithmetic, natural conversation and vision.
However, fate is a invisible force, seen only in hindsight. As Moore’s Law accelerated through the 90s, the first signs of the end of winter emerged. In 1997, IBM’s Deep Blue achieved what many had thought impossible: defeating the world chess champion. Garry Kimovich Kasparov, who was at the time the reigning World Chess Champion and who bore the nickname the monster with a thousand eyes who sees all , had triumphed over Deep Blue in their first encounter in 1996. Winning with a match score of 4-2, he dismissed the machine as having “the intelligence of an alarm clock”. However, the upgraded Deep Blue that he faced the following year was twice as fast and could evaluate hundreds of millions more positions per second. Deep Blue tasted revenge in their rematch, securing a narrow but famously controversial 3½–2½ victory. This milestone demonstrated that machines could surpass experts in structured domains, even though this was achieved through brute computation rather than true intelligence.
The early 2000s saw more advances in AI, such as in their ability to understand language. While still far from Turing’s vision of natural conversation, AI systems started demonstrating genuine comprehension of written text. By the 2010s, the AI spring had fully set upon the world with the emergence of AI chatbots. It is during this period that formal evaluations of AIs emerged to measure this progress accurately. And thus, from mastering reading comprehension (Stanford’s SQuAD benchmark in 2016), to elementary mathematics (GSM8K in 2021), to eventually achieving expert-level performance across 57 academic subjects in the Massive Multitask Language Understanding benchmark, AIs surged on.
By 2023, AI systems started achieving remarkable scores in graduate-level exams. At this time, I was also well into my graduate studies, reading my PhD in Genomics, while also collaborating with a close friend of mine on some AI projects of our own. In my private work, I used multiple chatbots to quickly write and review small bits of code, while also studying the behaviours of these chatbots with each new release. As they evolved, so did my use of them. I soon realized that these systems were getting good, seriously good. I could converse with them about my research as if I were consulting an equal. I sounded the alarm to my friends, but of course, affirmers were few, and skeptics many.
In the past 75 years of AI history, these skeptics have constantly drawn and redrawn the boundaries of machine intelligence, perpetually shifting the goalposts of what constituted “true” intelligence. First, they claimed machines could never master mathematics beyond mere calculation, then that they could never play chess at a grandmaster level, then that they could never understand natural language, nor create art, music, or poetry… One by one, these boundaries have fallen, validating Turing’s insight: if a task can be distilled into an algorithmic process, it can be performed by a machine. By the end of last year, AI tools had become medalists in the International Mathematical Olympiad and others had developed into agents capable of semi-autonomy. This forced the skeptics to retreat into what they considered their final, unassailable fortress. They held that despite these successes, some quintessentially human capabilities will still remain beyond the reach of machines. These, they said, were genuine creativity, insight, and the ability to generate novel ideas that could advance human knowledge. AIs will never generate hypotheses that were absent in their training data, and will forever lack the spark of genius. This was their final boundary, the exclusive province of human intelligence. Well, as of February 19, 2025, not any more. This fortress has been breached.
As scientists often do, Professor José R Penadés and his fellow colleagues in the Department of Infection Disease, Imperial College London, had been grappling quietly with a fascinating biological mystery for over a decade. Of interest was a type of genetic element that they knew to always be confined within their own bacterial species, but which kept being detected in other species. This genetic element is a chromosomal island that is usually activated by viruses when they infect these bacterial cells. The chromosomal islands also contain codes for capsids (protein shells). Hence, these genetic elements are called “capsid-forming phage-inducible chromosomal islands” (cf-PICIs). The mystery of finding a species-specific cf-PICI in a different species of bacteria is therefore like that of finding a unique archeological artefact far from its country of origin with no trace of how it got there. This is because although these genetic elements can package themselves into protein shells, these shells are inert, and cannot penetrate into other bacterial cells.
Through years of painstaking laboratory work, José and his team of researchers finally uncovered the answer: these genetic elements had evolved an ingenious method of transport. You see, tails are the primary structures that phages (viruses that infect bacteria), use for recognizing, binding to, and infecting bacteria. During an active viral infection, these viral tails are therefore strewn about in the environment. Therefore, after the genetic elements have packaged themselves into shells, they can co-opt these viral tails. The viral tails therefore give the shells the ability to detect and transfer their genetic elements into other bacterial species, especially species that those viral tails had already evolved to target. This groundbreaking discovery took countless experiments to validate and relied upon serendipitous insights by multiple experts.
Before this work was published, José presented an AI with the same mystery that had plagued his team for a decade. This AI, dubbed “AI co-scientist” had been designed by Google to think like a scientist. It relied on an innovative multi-agent architecture based on Gemini 2.0, allowing it to debate against itself, score and improve its own ideas iteratively and reason deeply for extensive periods of time. To prompt the AI, José offered it only a 1-pager containing basic background information about these genetic elements and the observation that they appeared in different bacterial species. Without access to any experimental data or specialized knowledge in the field, the “AI co-scientist” set upon the puzzle.
What followed was remarkable. Unlike traditional AI models that operate in seconds and minutes, this system engaged in prolonged hypothesis generation and refinement, mirroring the type of collaborative discourse that is often witnessed in academic and research institutions. After 2 days, the AI co-scientist independently proposed the exact solution that had taken the Professor and his entire team a decade to uncover. It also generated several plausible, well-reasoned and compelling alternatives to that solution, some which his research team is now actively investigating. For example, it proposed that these genetic elements might also use a form of bacterial “mating” as an additional transfer mechanism. This wasn’t just mere pattern recognition or regurgitation of training data, it was creative scientific thinking at the highest level, equivalent to that of experienced principal investigators. It was as if you described a complex crime scene to a detective, who then deduced the correct solution that had taken a team of investigators years to solve, and identified sources of additional evidence that the investigators had missed, all in an instant.
This breakthrough raises multiple profound questions about the future of scientific research, and even more consequentially, the future of AI. As AI systems become more competent at generating plausible hypotheses, how can we evaluate these ideas efficiently and at scale? We have created minds that outpace our careful methods of experimentation. In the time it takes us to verify one good hypothesis, machines can produce a billion others. The speed of ideation now far surpasses that of validation. This widening chasm between conception and confirmation may become the defining challenge of our era, the central bottleneck to human progress in this age of intelligences.
As these AI systems percolate into routine scientific work, we also face questions on how we attribute discovery. When an AI is used as an aid in generating hypotheses, who has the intellectual ownership of these ideas? Our current frameworks of attribution – built on centuries of human-to-human collaboration – did not anticipate the contributions of non-humans. Should we welcome these AIs as our intellectual partners, worthy of co-authorship and acknowledgments in our publications, rather than continuing to cite them amongst other tools that we use? Even more unsettling is the question of human agency in this new landscape. As AIs develop intuitions that once required decades of hands-on experience, what becomes of human expertise? What happens when machines can traverse in seconds what would have taken us years? When automated end-to-end pipelines of scientific discovery are developed, will human feedback be rendered unnecessary altogether, like the naive musings of a little child to its father as he is working? If we choose to embed ourselves within these pipelines, shall we not be the weakest link, if not the slowest? Or perhaps we shall be relegated to mere validators at the end of the pipeline and still be happy.
We need not rush to answer these questions. For now, let us enjoy the view as we stand at the moment Turing foresaw. Machines have finally joined us in cognition. The path ahead is now lit with the promise of discoveries we can barely imagine. Some of these discoveries will be made by us, but many more by the cognitors that we have created. Remember, no matter how impressive these AIs seem, this is as unremarkable as they will ever get.
V.
