Live Demo of Deep Background for Claude
I finally do a decent video, and maybe also get closer to the core value
I hate the term fact-checking. It’s a term I have to use often for my work, because if you start ranting about a contextualization engine people start giving you the “OK, grandpa, let’s get you to bed” look.
But one nice thing about doing all this as a hobby right now is I’m realizing I get to use the terms I want. So at the risk of the go-to-bed-grandpa look, I’m going to talk about what excites me about this tool most: it bootstraps you into a discourse.
Is that the two sentence pitch I’ll end up with? Probably not. But if you’re interested in seeing a very different way that students might use LLMs you’re going to really enjoy the video below, I promise.
Oh, right - I’m calling SIFT Toolbox Deep Background now, across both the GPT and the open source prompt.
If you are not paying for ChatGPT or Claude, do not use this Deep Background prompt. This prompt is for the paid version of Claude Sonnet, and while a careful user can use a free version of ChatGPT to get some rounds with this with o3, in practice people using free versions often end up using 4o, which will hallucinate. I strongly recommend the paid version of Claude Sonnet 4 for this prompt.
HI Mike, I am really enjoying your ongoing thinking and development of this tool. I can see how it surfaces and makes visible what we try and teach students. One possible enhancement I can see is to have the tool surface the assumptions that underpin the different standpoints within the discourse you are exploring
I appreciate you sharing (and developing) this tool. I was very excited to use it after watching your video, but immediately ran into some issues. I did try to verify that I was using o3, so that hopefully wasn't the issue. Not sure if I just had particularly bad luck, but a) initial response conflated the specific subset I was asking about with tangentially-related groups (leading to statements that are simply not backed up by any research), b) context report included sources invented of whole cloth, and c) source table had author names and publication years not match the actual article linked to.
I really like the idea and how it can model the research process, but it runs into the same issues I've had with my forays into genAI so far: the likelihood of inaccuracies or straight-up hallucinations is so high that I don't see it saving time for experienced researchers and can see it leading novice learners astray.