A new Google model is nearly perfect on automated handwriting recognition
A Substack article claims an unreleased Google model is nearly perfect at reading historical handwritten documents and even infers subtle accounting details, prompting both excitement from historians and skepticism from technologists. Commenters contrast this apparent leap in handwriting OCR and visual reasoning with more familiar LLM strengths and weaknesses, debating whether such feats represent genuine “understanding” or just sophisticated pattern-matching on well-covered training data. Many foresee major benefits for archival research and transcription, while others criticize the article’s hyperbolic tone and caution against over-interpreting isolated, non-reproducible examples as evidence of emergent abstract reasoning.
Historical & practical use cases
- Several commenters are excited about strong handwriting recognition, especially for:
- 16th–18th century archival material (Conquistador accounts, colonial Spanish files, ledgers, local town records).
- Genealogy, Renaissance Neo-Latin texts, family diaries, and children’s handwriting.
- People describe current LLMs (Gemini 2.5 Pro/Flash, Claude, o3) already being very useful for:
- Transcribing handwritten notes and food logs with few errors.
- Searching, summarizing, and translating scanned historical documents.
- Acting as research assistants via custom tooling and agents.
Skepticism about OS clones and “wild capabilities”
- Many doubt claims that the model “codes full Windows/Apple OSes, 3D software, emulators” from one prompt:
- Most likely outputs are web-based UI clones (HTML/CSS/JS) that resemble OS desktops, not kernels.
- With abundant open-source OSes and emulators on GitHub, such results may be remixing or near-copying, not deep novelty.
- Some see this as classic social-media hype and suspect astroturfing and engagement farming around new model launches.
Novelty, reasoning, and “stochastic parrots”
- Long debate over whether LLMs:
- Only interpolate from training data vs. genuinely extrapolate and create novel solutions.
- Are “just next-token predictors” vs. systems that necessarily build internal world models to predict well.
- Examples used on the “they reason” side:
- Math Olympiad-style problem solving.
- Material-physics intuitions (“can X cut through Y?”).
- Multi-document code or research synthesis.
- Critics respond that:
- Impressive feats often align with dense training coverage (e.g., NES emulators, sugar loaves, ledgers).
- There are no clear signs yet of breakthroughs comparable to relativity or the transistor.
Handwriting example and trust issues
- The sugar-loaf ledger case that impressed the author is heavily debated:
- Alternatives: the model may have simply seen the space (“14 5”), recognized period notation, or drawn on prior examples of typical loaf weights.
- Regardless, it violated the explicit instruction to transcribe “exactly as written,” which some see as a reliability red flag.
- Historians worry about:
- Being subtly biased by AI “guesses” in ambiguous passages.
- Using models on primary sources without strong provenance and error-characterization.
Concerns about hype, regressions, and evaluation
- Many find the article hyperbolic, with marketing-style language about “emergent abstract reasoning.”
- Several users report:
- Earlier Gemini 2.5 Pro previews feeling stronger than later releases, possibly due to cost optimizations.
- Models that once worked well for research later hallucinating sources or references.
- There’s interest in standardized handwriting benchmarks; some are surprised none are widely cited.