2026-09-06 12:01 UTC
@GreatBigTable@mastodon.social
This is simply not true in the general case*. The proof is simple and based in computer science and maths:
A great LLM to host locally is Qwen 3.8 27B. Quantized to 4 bit weights the size of the model is ~14GB on disk.
All of human knowledge cannot be compressed down to 14GB and then decompressed again (see Shannon's theorem). Thus LLMs do not work by storing training data.
*) When training has _failed_ and cause what's known as "overfitting" too much training data is stored close to verbatim in the model. This is unwanted (wastes model space) and are extreme outliers model creators work actively and successfully to make sure doesn't happen.
@eff@mastodon.social
Replies (1)
-
@GreatBigTable@mastodon.social 2026-09-07 21:57
@troed@swecyb.com @eff@mastodon.social I am talking about overfitting. That case has existed in models pushed to the public by frontier AI labs and is the basis of cases against them by orgs like the New York Times. Overfitting may not be the desired outcome, but it is inherent in the technology if not accounted for. The problem for the labs is if their weights make outputs too far from source text, they can get less useful results. If they adhere too close to the original source, it could lead to legal trouble.