AI Peers Inside a Burned Papyrus Without Touching It
A lump of carbonized papyrus in the Bodleian Library, once deemed unreadable, has become a symbol of AI’s new role in cultural heritage. Part of the Herculaneum scrolls buried in 79 AD, the artifact would likely crumble and lose its ink if anyone tried to unroll it. For decades, historians had no safe way to access its text. That changed when researchers used the Diamond Light Source, an X‑ray facility that maps an object’s internal structure without damage, to scan the sealed scroll. Deep learning algorithms then processed the 3D images, virtually reconstructing the layers and amplifying faint differences in density that reveal hidden ink. The result: legible Greek text inside a document that has remained sealed for over 2,000 years. This type of AI cultural heritage work transforms how we might approach other fragile artifacts long considered permanently inaccessible.
From Foundation Models to Trial Design: AI in Cancer Research
At the American Association for Cancer Research annual meeting, scientists described how AI is reshaping the entire discovery pipeline in oncology. Early machine learning tools focused on structured, tabular data. Today, foundation models trained on vast genomic and cellular datasets can recognize subtle patterns in single‑cell transcriptomics, DNA sequences, and protein structures. Bo Wang’s team, for example, built scGPT on a repository of more than 33 million cells to perform tasks like cell‑type annotation and gene network inference, and later X‑Cell, trained on nearly 26 million perturbed single‑cell transcriptomes, to predict how gene expression shifts under specific perturbations. These models support AI cancer research by generating hypotheses about drug targets and resistance mechanisms. Coupled with pathology image analysis and AI‑assisted clinical trial design, they promise more personalized treatment strategies. Yet researchers also warned that these tools must be carefully validated to avoid overfitting and clinically misleading predictions.
Deep Learning, Machine Vision and the Automated Lab
Behind both ancient scroll decoding and modern cancer labs is a common technology stack: deep learning and machine vision. Advances in deep learning machine vision allow systems to interpret complex visual data—microscope images, X‑ray scans, tissue slides—with a consistency humans cannot match over millions of samples. Market analysts project that deep learning in machine vision will expand rapidly over the next decade, reflecting demand for AI‑powered automation in research and industry. In the lab, this means robots that can visually inspect cell cultures, identify anomalies, and adapt experimental protocols in real time. In archaeology, similar techniques segment virtual cross‑sections of a scroll, track its layers, and highlight tiny variations that correspond to ink. As these tools become standard, AI scientific discovery becomes less about isolated algorithms and more about integrated pipelines that connect data acquisition, analysis, and robotic experimentation, tightening the loop between hypothesis and result.
Massive AI Funding and Lanier’s Reminder: It’s Still About Humans
The surge of UCL‑linked AI startups illustrates how investors are betting on AI‑driven discovery tools. Ineffable Intelligence, founded by David Silver, raised USD 1.1 billion (approx. RM5.06 billion) in seed funding, reportedly the largest seed round in Europe, with backers including Sequoia, Lightspeed, NVIDIA and others. Recursive Superintelligence, co‑founded by Tim Rocktäschel and Richard Socher, secured USD 500 million (approx. RM2.30 billion), aiming to build AI systems that continuously improve themselves and automate the frontier AI development pipeline. These UCL AI startups funding rounds signal strong confidence that AI can accelerate science itself. Yet Jaron Lanier urges a more grounded view: AI is not an alien intelligence, but a large‑scale collaboration distilled from human work. Language models, vision systems, and scientific foundation models all rest on human‑curated data, experimental design, and interpretive judgment. Recognizing this keeps credit, responsibility, and incentives anchored in the people behind the models.
Limits, Risks and the Next Frontiers of AI-Driven Discovery
Even as AI scientific discovery accelerates—from AI cancer research to AI cultural heritage—its limits are becoming clearer. Foundation models can overfit to particular datasets, learning quirks rather than general biological rules. Misinterpretation of AI outputs, especially when wrapped in persuasive visualizations or natural language, can mislead researchers and clinicians if results are not rigorously cross‑checked. That is why reproducibility, transparent methods, and traditional peer review remain essential safeguards. Many labs now pair AI predictions with independent validation experiments and maintain clear audit trails of data and model versions. Looking ahead, similar techniques are spreading into materials science, where models propose new alloys or battery chemistries, and climate modeling, where AI refines forecasts from noisy sensor data. The challenge is to build governance—data standards, bias checks, human‑in‑the‑loop review—into these systems from the start, so that AI remains a discovery engine rather than a black box.
