Training an AI chat-bot for the TSE website has been a cool project for us, over the past 2-months. But today’s AIs struggle in gauging which content is most important to decision-makers. Hence, we went through the entire TSE research library, and simply added our own subjective “importance score” to every report and data-file. So you can now search the TSE website and rank the results by importance. These direct experiences with AI matter for some key debates in AI energy consumption.
This video is part of a new series, discussing actionable conclusions from TSE research. The most important debate in energy still feels like the upside-downside on electricity demand growth (aka load growth), in the AI era, as discussed in our electrical grounding note, and last month’s unpopular video.
Since this time, we have deepened our work into low-voltage instrumentation and controls, looking at flow-meters, Emerson case studies, Rockwell case studies and WiFi networking; and also enabling materials for the AI eco-system like Rare Earth magnets, part of a materials trinity with copper and lithium.
Trying to gauge the upside-downside in AI energy demand has taken us deeper into the topic of scaling laws in AI training: a key debate is whether better models require exponentially more compute; or conversely, new model architectures and software-driven advances.
What speaks to this debate is our own experience in building an AI chatbot for the TSE website. It was relatively easy to build a vector database and MCP with all of TSE’s content. But today’s LLMs cannot effectively gauge what matters to decision makers, simply from this vast amount of information.
In the end, we decided that what would be more useful than an AI chatbot, was simply to rank all of TSE’s content, according to a subjective “importance score”. So we built this into the TSE website. If you search the TSE website now, you can filter results according to “importance”. Or according to a “combined score” that captures relevance, recency and importance.
It is interesting that dumb linear programming can potentially be more helpful than training an AI chat-bot. Although we will continue exploring whether we can build an AI chatbot for the TSE website in the future, please do email in if we can help you using our old-fashioned human brains.