Just noticed this it the 140th LLM release I've written about in my blog!
@simonw
-
Kimi K-2.1: Moonshot’s Upgraded Open Weights Model
By
–
My notes on Kimi-K2-Instruct-0905, aka Kimi K-2.1 – an incremental improvement on Moonshot's previous trillion parameter open weights model, now with twice the context length (256k up from 128k)
-
Anthropic Case Sparks Book Scanning Projects Among AI Labs
By
–
I don't know if other AI labs are doing their own book scanning projects – I wouldn't be surprised to hear them start to do this now based on the outcome of the Anthropic case, though they may be waiting for more concrete legal precedence first
-
Headlines Over Details: How AI Stories Mislead the Public
By
–
It only "looks good" to people who read the headline without confirming the details of the story (Which is most people, so they probably benefit from that quite a bit)
-
Settlement Over Pirated Ebooks Highlights Digital Copyright Issues
By
–
They are just buying used copies now – this settlement was over ebooks they pirated back in 2021
-
Anthropic Hires Google Books Project Partnership Former Head
By
–
Anthropic actually hired the "former head of partnerships for Google's book-scanning project" for this
-
Lawsuit Over Illegal Pirated Content Downloads Prior 2024
By
–
That's exactly what they are doing now – the lawsuit was over their activities prior to 2024 where they were found to have downloaded illegal pirated copies
-
AI Training on Books: Legal and Economic Implications
By
–
They'll only lay you that $3,000 if they are caught pirating your book – if they buy your physical book (or a second hand copy) they can train on it for much less money than that
-
Authors’ Copyright Concerns in AI Training Data Usage
By
–
Doesn't help the millions of authors who's books were used for training data without being one of the 500,000 that were illegally downloaded
-
Document value in model training data
By
–
One thing I've never understood is how much value an individual document has to training a model – are there single books that, if included in the training data, would
materially improve the resulting model?