Off-topic:BenFenner wrote: ↑2024-11-09, 10:28
ChatGPT said:
Pale Moon does not support ECMAScript features introduced after ES6 (ECMAScript 2015).
This is of course obnoxiously wrong, and if I were one of the devs who spent time implementing any ECMAScript features introduced after ES6 for Pale Moon I'd be furious.
I'm not furious because I know you can't use ChatGPT for any serious research.
The problem is that ChatGPT's model is trained on a lot of old data.
Once upon a time, that statement was true. But things change.
Which brings me to the glaring Achilles' heel of LLMs:
It is not possible to update existing models by updating their training or remove incorrect data once it has been incorporated. So the moment anything changes, which is the case here, the model's knowledge becomes incorrect, and the only way to correct it is to retrain the model from scratch, excluding every single instance of old data it was previously trained on.
In other words: LLMs can only be trained on a snapshot of data, and you can never learn them new things that it already has an "opinion" about. It cannot unlearn associations, unlike humans. One of the reasons LLMs are a pretty crude tool, even though they can help lower barriers for people.
This means that using an LLM for web search is pointless: web data changes constantly.
Using an LLM for academic research is pointless: the academia constantly discover new things or change established knowledge as new theories are tested.
Using an LLM for generic text is risky as well: if it has been trained on a lot of old data, you'll get issues with writing styles, incorrect meanings of words, etc.
... unless you are constantly re-training new models from scratch using new, vetted, accurately-tagged data (that one example of eating a rock from an Onion article being used for a serious answer, for example, was a good demonstration how unvetted data leads to poisoned models that don't understand separating truth from fiction)