An interesting email has reached me, noticing that work seems to have resumed at Google Books:
I don’t know if you’ve noticed this, but Google Books has been adding new material this year from a number of university libraries that have entered their partnership program. As could be expected, the [most likely] reason is that they are desperate to get more printed material to train their LLMs, and libraries are happy if big corporations put up the money and resources to scan their collections. Here’s a very recent article about digitization at Delaware University.
Lately I have seen many books I couldn’t find before with digitization date from around [last April], from Oxford and other places….
This seems entirely plausible, and also good news in a way.
It’s been known for a while that some AI companies are scanning books. An “AI” system seems to be basically a search engine with a chatbot on the front, which is why Google panicked when AI came along. It’s the same market. Naturally the search engine has to be populated somehow. They call it “training an LLM”, in order to obfusticate the violations of copyright, but in truth it seems to be something as simple as populating the database or “model” with the data to be searched. In this case, all the books in the world.
I’d rather have access to the books than whatever an AI engine spits out of them, tho.
AI is more complicated than a chatbot and a search engine, it’s just that this is what you are more likely to encounter. As someone in Remote Sensing, it is more complicated. AI is what we label complicated classification non-parametric classification methods. Parametric classification methods such as maximum likelihood and k-nearest neighbors have a simple relationship between the sample of each class (which is used in classification) and the classification product: this pixel gets classified as forest because the values of each of its bands is closest to what is labeled in the sample forest and not say urban. Main issue is when the classes are too human defined and not just spectral: agriculture can include row crops, vineyards, fruit trees which can be closer to lawns, shrubs, and forest trees than each other. Parametric classification fails that
For non-parametric methods such as gainful adversarial networks, deep learning etc, you need gigantic datasets of labeled samples. Using very complicated statistics, the language in LLM does not refer to actual human languages but an abstract statistical model of how human language work, allows each pixel to be more correctly classified based on the human derived classification category. AI models do work better than simple statistical models but it is an issue of resources. Do you want to spend several times the computation cost for something only 10% better? In other cases yes, in other no
For the kind of AI you are running into and the whole media discourse is about, the feedstock is human made texts, rather than images (or LiDAR etc) in Remote Sensing. Still, as someone who has been working with Copilot the last year, there are severe limitations and you need to be very careful what you ask and how you ask it, otherwise garbage in-garbage out. Go ask the AI models what are the Holy Mountains in Greek Orthodoxy, they will get Mt Athos, but not the other 4. They may answer Meteora, which however has never been a Holy Mountain. A feedstock of mostly English language literature means that it won’t get the right answer.
Useful thoughts – thank you. I get very reasonable results, not least because I rarely take it at face value.