Post

Some heuristics that have helped me think about LLMs clearly: 1) The גּוֹלֶם (Golem) myth: words have a power of their own. Write them on a clay figurine and it will imitate a human in certain ways.

Darren Dahly@statsepi.bsky.social

Lots of academics seem confused about the difference between a heuristic and a model.

Ari B Friedman

@abfriedman.com

2) Plagiarism of all of the world's smart words written by smart humans, and it's remarkable how close anything you could write is to something that's been written before. 3) A model trying to predict the next word that has some remarkable emergent properties.

Ari B Friedman

@abfriedman.com

4) The cost of generating reasonable-sounding bullshit has gone down way faster than the cost of detecting it. 5) A useful thing that we will figure out more uses for, if we can find applications where false positives and false negatives are acceptable or mitigated.

2/10/26, 4:33 pm0 reposts0 likes

Ari B Friedman

@abfriedman.com

6) Just a predictive model that can interpolate well but can be quite sensitive to order and other minor details. Ask yourself, "Is the conversation I'm trying to predict likely to be anywhere in the training data?"

Ari B Friedman

@abfriedman.com

Apropos of sensitive to minor details, I'm playing with coding with one, and I had previously told it that the ZIP files were redundant with the CSV files in the data folder. Then I asked it to make a patient ZIP code variable and didn't capitalize ZIP. It next asked if it could unzip the files.