I can understand thinking the graphic isn't clear but the blog itself makes no claims that this graphic refers to any particular user requirements or workflow plans. The text closest to the image seems to indicate the intent is:
The corpus of data and content on our public platform is constantly growing, evolving, and improving from the contributions and feedback from our users.
Or, from the "Future of Community" blog:
The knowledge store supports user and data paths to access existing knowledge on demand while accelerating the creation and validation of new knowledge. The Knowledge Store is continuously growing, improving, and evolving at the speed of technical innovation, reinforcing user and data loops and solving for “LLM brain drain” and “Answers are not knowledge" problems.
As such, it seems the graphic's purpose is to show the "data loop" that the "knowledge" follows.
- The content is created,
- added to the model/database,
- used to provide search results.
- If the search results can't answer the question, a new question is asked, leading back to 1.
I just don't see anything hiding in the bushes here to be wary of, in regards to requiring use of an LLM before asking. In fact, the only aspect that's not in the current model (and the one that's existed since the beginning) is the specific "Summarized results with attribution" instead of the more general "search results".
Ramblings on search and LLMs
The rest of this is my thoughts on the subject more broadly spawned by the framing of the question.
It's OK to be distrustful and concerned about things - that's kinda the state we're all in but there's nothing in this graphic that concerns me. In fact, it looks like a process that I'd like to see as how users should engage with the sites. In fact, I'd go as far to say that this is how the SE ecosystem currently works and it's being adjusted to include the capabilities of improved search tools.
The reality is, many questions asked - particularly on SO, are duplicates or should be closed for some other reason and a big part of that is that the question asking process doesn't include an (adequate) search feature. While the expectation has been that users should do research first, many askers don't, or if they do, they can't find the answer they need. I've often struggled (as a non-dev) to even know what terms I need to use to describe what I'm trying to do, so it's understandable that someone might struggle to find the answer they need.
Additionally, I've stated many times that the current UI makes it so easy to ask a new question, there's not much incentive to do some searching first or to understand the scope and expectations of the site. There need to be more speed bumps - for at least some users - that walk them through the optimal preparatory process manually. While Staging Ground and the Ask Wizard kinda touch on some of this, they still overly focus on "quality", with the end goal being getting a new question posted rather than finding existing answers.
I know many sites have sought out solutions for large volumes of close-worthy questions being asked. Addressing this was one of the things I worked on most consistently, whether it was three vote closure or encouraging customization of the Ask page. When I was a mod on Interpersonal Skills, I remember discussing having a "review" phase for all newly-asked questions because so many of the questions were low quality or out of scope but got quick answers instead of closure. It's a real issue that communities here on SE want help addressing.
I don't think that LLMs can't be part of that process, and they may be quite good at some aspects of preventing problematic questions or helping guide users trying to find an answer. Ihave a hard time believing that any process would involve required use of an LLM for all users, if only because the company has - up to now, at least - been so gunshy about building anything that makes it less likely a question will be asked.
Aside: Regaining lost site usage
While the site has seen a reduction in use over the last decade, there was a big drop around the same time that ChatGPT was released in fall of 2022. That's because, as much as LLM can hallucinate or give bad (if technically accurate) information, they can do a pretty reasonable job of answering simple questions. While I've not tried it, I'm guessing that it'd be quite able to tell me how to create a table in HTML or escape VIM.
Fewer people are using the site on a daily basis. As a business, SO needs to address that before the site collapses entirely, particularly as LLMs are expanding what they can accurately answer over time. This is why the company is putting effort into isolating the content (to the degree they can) and building in LLM tools on site or requiring partner AI companies to link to SO content. If AI tools don't have constant access to enough high-quality, current data, they lose the ability to provide answers to questions and users will seek out a better LLM.
SO wants to be that better LLM so they can bring back the people who might have once visited the site from search results but were lost to LLMs who could answer many of their questions. That's where the blue circle in the graphic comes in, particularly the two blue dots. The newly-asked questions and their answers become the seeds for ensuring the data the LLM is using is high-quality and current. This also benefits their Teams product but a bit out of scope here.
Building search tools that fit SE
The biggest question is "What do the improved search tools look like?" Search can be a lot of things - only some of them LLMs - and, as I mentioned in the last section, LLMs are actually perfectly capable of answering many questions - often the questions that are extremely commonly asked and have been answered many times over - and often still are because the current search tools suck.
The difficulty is ensuring that the information you're looking for is something the LLM is capable of answering well - which may not be the case - and ensuring the search/prompt includes enough information to be narrowed down to a particular issue. I was talking to a researcher who works at one of the major AI projects a month ago and we actually discussed the issues LLMs have, including how they struggle to say "I don't have enough data to answer this" or rate how sure they are of their answers, as this is often what leads to the hallucinations so many have experienced.
In my perfect world (which may not be possible), some things I'd want to see in an LLM search tool on SE include:
- quoting/citing excerpts of answers in addition to (or instead of) summarizing them.
- ensuring they have enough information before attempting to answer the question (avoid making assumptions)
- indicating clearly how sure they are of the answer
- falling back to "dumb" search (semantic/keyword) in cases of low confidence or if user requests it.
Asking questions via LLM
Assuming the LLM guides the user to clarify the question sufficiently, the info the LLM search tool has gained may be useful to create a new question if there isn't one that exists already. But, as we've already seen, there are many pitfalls here that need to be avoided.
While I can't know what SO is planning, if they're smart, they'll focus on improving the question wizard so that it "learns" how to review a question and catch common issues or make recommendations to the asker to ensure questions meet site guidelines, possibly using training data from Staging Ground. It should be able to
- identify questions that belong on other SE sites (including meta),
- determine if the question is in the correct language,
- prevent common issues like images of code,
- review tags to determine if they're used correctly, draw attention to tag warnings/blocks
- review the title,
- fix formatting, grammar, spelling, etc.
This won't catch everything (and there's likely stuff I forgot) but it's basic stuff that doesn't need to get added to the plates of reviewers - and a lot of it is universal to the network (and probably Teams, too).
Done well (big caveat, of course), there's so much room for improvement in this ecosystem that even small changes can have huge positive impacts. There's no reason to leap to the interpretation that LLMs will be required or even core to the process.
