Building a marketing chatbot pt. 3
Why voice control turned out to be a dead end, and the security reason this chatbot stays private rather than going on the open web.
Ben Rees - 16 March 2025
A short post, and really a video.
I have abandoned attempts to control the whole thing by voice as it just seems an enormous amount of work the minimal return. It's cool, but not practically helpful.
Instead and I think more interesting, is going into the world fine-tuning more. If I just wanted general information about marketing, I can do a Google search. The value comes from producing something based on my specific views. As a simple example, I attach much more credence to expertise to search engine optimization. This isn't the view held by everybody in marketing :).
The other interesting point is about security. I have thought about putting this chatbot on the web for anybody to use. But I would need to do an audit of every single piece of information in the underlying database. So for now this is something that I use in general conversation as an aide.
Anyway, I will post more videos as it starts giving me more and more insight.
Your scaling story is your most AI-retrievable content and the most interesting
Your B2B scaling story is your most differentiated content asset. Here's why it stays invisible to AI systems and what to change.
My AI visibility went from zero to 26% in six weeks. The one category still at zero is the one this post is about.
Six weeks after publishing a zero-appearance rate across every AI platform, most categories moved. The one still at zero is AI Search Visibility itself.
Running AI locally doesn't change what you know. It changes what you can keep private.
Local models protect your data but can't change what external AI systems believe about your brand. That's still a retrieval and training problem.
I ran a local AI model through 179 rounds of self-refine. It collapsed. One sampler setting fixed it.
An unguarded local model collapsed into repetition after 179 rounds of self-refine. Adding one sampler setting and rerunning the identical task produced 201 clean cycles instead.