Asking The Stupid Questions of GenAI

I recently caught up with a pal who happens to be working at the center of the AI storm. This person is one of the very few folks in this industry whose point of view I explicitly trust: They’ve been working in the space for decades, and possess both a seasoned eye for product as well as the extraordinary gift of interpretation.

This gave me a chance to ask one of my biggest “stupid questions” about how we all might use chatbots. When I first grokked LLM-driven tools like ChatGPT, it struck me that one of its most valuable uses would be to focus its abilities on a bounded data set. For example, I’d love to ask a chatbot like Google Bard to ingest the entire corpus of Searchblog posts, then answer questions I might have about, say, the topics I’ve written about the most. (I’ve been writing here for 20 years, and I’ve forgotten more of it than I care to admit).  This of course only scratches the surface of what I’d want from a tool like Bard when combined with a data set like the Searchblog archives, but it’s a start.

My friend explained that my wish is not possible now, despite what Bard confidently told me when I asked it directly:

Well, no. Bard hallucinated all manner of bullshit in its answer. Yes, I write about technology, but not the Internet of things. I guess I write about society, but mainly in the context of policy and consumer data, not “education, healthcare, and the environment.” Culture? When’s the last time you’ve seen me write about movies?! And if I ever start writing about “personal development,” please put one between my eyes.

Bard’s list of supposed articles was even funnier – it reads like an eighth-grade book report culled from poorly constructed LinkedIn clickbait. Bard is a confident simpleton, despite its claim to be able query specific domains (in this case, I responded to Bard with this new prompt: “This is not right. That site does not cover music, movies. Nor does it do motivation, well being, productivity. Why did you answer that way?” Bard’s answer was … pretty much the same, though it did clumsily incorporate my corrections in its response:

Gah. My next prompt was an attempt to clarify where Bard was getting its answers, since it was clearly not using the domain. “Are you actually referring to content on the site to do these answers?”

Bard’s answer:

Ok, then, at least we’re getting some honesty. I decided to try one last time:

Now this was quite the freshly whipped bullshit: Actual percentages of how the content on my site breaks down! Unbeknownst to me, more than one in ten of my posts are about cybersecurity – a topic I’ve rarely if ever written about here.

Ok, enough beating up on poor Bard. My well-placed friend explained that while it’s currently out of scope for a standard chatbot like Bard or ChatGPT to do what I’m asking of it, “domain specific” queries was a hot area of development for all LLMs. So when will it happen? My friend didn’t commit to an answer on that, but I did get the sense it’s coming soon. The ability to apply LLM-level intelligence to large data sets is just too big an opportunity – in both B2C as well as B2B/enterprise markets.

A big reason this is taking more time than I’d like is cost. Noted AI investor Andreessen Horowitz recently posted a long explainer on the state of LLM models, but it all comes down to this money quote:  “Today, even linear scaling (the best theoretical outcome) would be cost-prohibitive for many applications. A single GPT-4 query over 10,000 pages would cost hundreds of dollars at current API rates.” By my estimates, this cost would need to come down at least four orders of magnitude – from hundreds of dollars per query to pennies  – to unlock the kind of magic that I’ve been dreaming about over the past few months. Not to mention all the technological machinations related to prompt handling, vector database management, orchestration frameworks, and other stuff that makes my brain hurt. But the good news, despite my rather pessimistic post from earlier this week, is that the good shit’s coming – we just need to be a bit more patient.

You can follow whatever I’m doing next by signing up for my site newsletter here. Thanks for reading.


Leave a Reply

Your email address will not be published. Required fields are marked *