I think Deep Research shows that these things can be very good at precision and recall of information if you give them access to the right tools... but that's not enough, because of source quality. A model that has great precision and recall but uses flawed reports from Statista and Statcounter is still going to give you bad information.
Deep Research doesn’t give the numbers that are in statcounter and statista. It’s choosing the wrong sources, but it’s also failing to represent them accurately.
Wow, that's really surprising. My experience with much simpler RAG workflows is that once you stick a number in the context the LLMs can reliably parrot that number back out again later on.
Presumably Deep Research has a bunch of weird multi-LLM-agent things going on, maybe there's something about their architecture that makes it more likely for mistakes like that to creep in?
Have a look at the previous essay. I couldn't get ChatGPT 4o to give me a number in a PDF correctly even when I gave it the PDF, the page number, and the row and column.
ChatGPT treats a PDF upload as a data extraction problem, where it first pulls out all of the embedded textual content on the PDF and feeds that into the model.
This fails for PDFs that contain images of scanned documents, since ChatGPT isn't tapping its vision abilities to extract that information.
Claude (and Gemini) both apply their vision capabilities to PDF content, so they can "see" the data.
So my hunch is that ChatGPT couldn't extract useful information from the PDF you provided and instead fell back on whatever was in its training data, effectively hallucinating a response and pretending it came from the document.
That's a huge failure on OpenAI's behalf, but it's not illustrative of models being unable to interpret documents: it's illustrative of OpenAI's ChatGPT PDF feature being unable to extract non-textual image content (and then hallucinating on top of that inability).
Interesting, thanks.
I think the higher level problem is that 1: I have no way to know this failure mode when using the product and 2: I don't really know if I can rely on Claude to get this right every single time either, or what else it would fail at instead.
Yeah, completely understand that. I talked about this problem on stage as an illustration of how infuriatingly difficult these tools are to use because of the vast number of weird undocumented edge cases like this.
This is an unfortunate example though because it undermines one of the few ways in which I've grown to genuinely trust these models: I'm confident that if the model is top tier it will reliably answer questions about information I've directly fed into the context.
[... unless it's GPT-4o and the content was scanned images bundled in a PDF!]
It's also why I really care that I can control the context and see what's in it - systems that hide the context from me (most RAG systems, search assistants etc) leave me unable to confidently tell what's been fed in, which makes them even harder for me to trust.
