[bluealienpie]

Google summaries from websites mainly did this before “AI” came around. Now I just don’t know the source unless I dig myself. Would rather it went back to before, or at least they implemented the veracity checks better.

I thought it was interested that the movie looper had argidrones, but what inspired that wasn’t part of the training set or discussion online. I get remarkably bad answers and search results which seem misleading at best.

[metabagel]

> Now I just don’t know the source unless I dig myself.

Claude includes links to references.

[foresterre]

What's a problem though is that while it includes a reference, its distillation from the source, or worse, from the combination of sources, is often plain wrong. I check them regularly, and it made me very distrustful of Claude's capacity to faithfully summarise or explain from a source.

When asked to give specific links, it's usually even worse.

[40four]

I don’t disagree. Many times when I’ve checked the source links a language model provides, I get concerned. It’s definitely important to stay vigilant and exercise critical thinking, but the model’s ability to synthesize information and converse with it in natural language still blows my mind sometimes

[tmp10423288442]

So does ChatGPT. The hallucination problem is not completely solved, but much better than a few years ago, especially if you use reasoning mode, where it’s more likely to spontaneously do a web search

[lazide]

Those references often have no actual relation to the answer. A surprising amount of the time, if you check them they’ll prove it wrong.