Vesuvius Challenge 2023 Grand Prize awarded: we can read the first scroll
Machine learning and advanced X‑ray imaging have successfully revealed substantial Greek text from a carbonized scroll buried by the eruption of Vesuvius in 79 CE, a breakthrough achieved through the $1M Vesuvius Challenge. Commenters highlight the technical ingenuity of virtually unwrapping and reading the scrolls, the historical importance of unlocking potentially megabytes of lost ancient philosophy, and the care archaeologists showed by leaving fragile artifacts unopened for future technologies. Attention also turns to next steps: scaling up scanning and segmentation, funding constraints, the role of AI in both image analysis and future textual interpretation, and the broader implications for unread artifacts such as other papyri and codices worldwide.
Overall reaction
- Many commenters express amazement, calling the result “sci‑fi–like” and a highlight of the year.
- People emphasize the 270‑year arc: from 18th‑century discovery, through industrial and scientific advances, to modern ML as the final “unlock.”
- Several note the importance of earlier restraint: many artifacts were historically ruined by aggressive handling; here, waiting paid off.
Technical approach and validation
- Discussion clarifies that this is computer vision on CT data, not generative language models; models detect ink-like features in 3D volumes.
- The data structure (layer stacks / volumes) makes it natural to reuse video architectures (e.g., transformers over “layers as time”).
- Concerns about hallucination are common. Others respond:
- Methods were reproduced independently from open code and data.
- Multiple teams, with different models and labels, converged on nearly identical readings.
- Models work on tiny local patches, detecting ink rather than whole letters or words.
- A “campfire scroll” test object was created and scanned; some debate remains over how compelling its before/after comparison is.
Bottlenecks and scaling up
- Thread highlights two main technical bottlenecks: costly high‑resolution synchrotron scanning and labor‑intensive segmentation of scroll layers.
- Ideas proposed: build or relocate scanners near Naples, crowdfund or philanthropically fund scanning (on the order of tens of millions), and use distributed volunteer segmentation (captcha‑ or SETI@home‑style).
- People debate whether prize money is large or small relative to effort; many see the prize framing as crucial for attracting talent.
Content and scholarly impact
- First recovered text centers on Epicurean philosophy of pleasure and scarcity; some find it underwhelming, others note even “mundane” texts are gold for historians.
- Expectations vary: some fear “more of the same” from one philosophical circle; others argue that even a single mind’s full library is historically transformative.
- Commenters link this to broader hopes of recovering lost literature, history, and everyday documents (including tax records).
Archaeology, preservation, and broader reflections
- Strong appreciation for deliberate non‑excavation (Herculaneum areas, other major sites) to avoid irreversible damage until tech improves.
- Discussion of obstacles to further excavation: modern town above the villa, local resistance, criminal interference, and funding priorities.
- Several use the project to reflect on:
- How much compute and engineering stack underlie modern ML.
- The fragility of modern digital data versus the surprising recoverability of ancient materials.
- The role of rich patrons vs. collective public funding, and what kinds of problems big incentives could unlock.