Very interesting. It would probably be more accurate to say that lower resource languages are a problem regardless of Latin or non Latin script.

And even more so for non Latin scripts.

Low resource languages are still a big problem for handwriting recognition. I work on Yiddish—all of the generative models currently return gibberish or pure hallucination when given a handwritten Yiddish letter. It’s a very interesting and exciting problem and one I think we can solve.

…And it turns out that a huge amount of what we seek from a human person can be simulated through this Frankensteinian reanimation of our collective dead letters. What a discovery!” www.newyorker.com/culture/the-...

Thank you, that’s very kind! I don’t have anything serious at the moment, mostly just playing around, but I will definitely let you know.

I guess I’ll have to try that, thank you! It’s always an effort vs results calculation of course. No GPU currently.

(I tried to put through a 340-page pdf of archive photos which I guess was a mistake). Is there an easy workaround for this?

I'm trying it out and so far it's great on crappy iPhone archive photos. But I'm getting a lot of "Expired ZeroGPU proxy token" error messages, and the model card does say "Despite its 1.7B parameter LLM foundation, dots.ocr is not yet optimized for high-throughput processing of large PDF volumes"

Interactive timeline: 150 years across 30+ predecessor orgs, with links to archival collections for further research: 150yearsofcare.org/150-years-of...

Oral histories: Video interviews with veteran staff, trustees and clients, exploring 40 years of change - AIDs epidemic, social justice movements, residential care, Medicaid evolution 150yearsofcare.org/voices-people/

Mental health pioneers: profiles of mid 20th-century leaders and their clinical contributions with rich multimedia archives 150yearsofcare.org/architects/

Case file deep dive: The remarkable story of Morris and Lena Adler through their National Desertion Bureau file 150yearsofcare.org/deserted-fam...

National Desertion Bureau Database: 19,696 case cards from 1911-1935, tracking down men who had abandoned their wives and children 150yearsofcare.org/ndb-database/

Interactive charity map: 1599 charity recipients and donors in 1870s New York City, revealing attitudes about “worthy” and “unworthy” poor in the 19th century 150yearsofcare.org/worthy-and-u... (Special thanks to @simonwillison.net - @datasette.io was crucial in creating the underlying data)

If you use a model dialogically, it’s effectively guiding you through the free-writing/revision cycle composition teachers recommend.

I get why this mode of use isn’t very visible. All the rhetoric around AI has been that it’s “generative” and does the work for you. Things like the “Dear Sydney” +