Post #3769002
2026-07-12 20:47 UTC
Encoder-decoder language models and all sorts of stuff were used for translation and spellcheck, long before “LLM” was even in anyone’s vocabulary. Embeddings models were used in documentation searches, in IDEs, and other places. Whenever you used any search engine, pre Sam Altman, you were likely hitting text models too.
They worked alright.
It was not an issue. Because they are tools with a function.
I think people need to be careful of spilling quite reasonable hate of Tech Bro AI into the wider, older field of machine learning. In spite of the effort to conflate them, they aren’t the same thing.
Replies (1)
-
@brianpeiris@lemmy.ca 2026-07-13 05:26
I kinda agree with this, except the machine learning field should bear some responsibility for begetting LLMs. In particular, they got very used to the idea of scraping the internet for huge amounts of data needed for all types of models, and paid less and less attention to how much energy their training and inference was costing versus the value the models were providing. The seeds of the problems with LLMs existed before they landed on the scene.