Delphi's research reports are legendary in crypto circles. When they publish analysis on a new token mechanism or DeFi protocol, founders take notes, VCs adjust their theses, and traders reposition their portfolios. Their work has influenced billions in capital allocation across Web3.
But here's the thing: being the gold standard for institutional research created an unexpected problem. The same depth and rigor that made their analysis invaluable also made it intimidating. A typical Delphi report might reference dozen other reports, technical concepts that required background knowledge, and market dynamics that assumed familiarity with crypto's evolution.
"We had this incredible body of work, but we kept hearing that people struggled to navigate it," explains Carter Lundy, SVP Operations at Delphi Digital. "Someone might land on a report about MEV and get lost because they didn't fully understand the underlying concepts. We were leaving value on the table."
The obvious solution seemed to be an AI assistant. Something that could explain concepts on demand, summarize lengthy analyses, and guide readers through Delphi's extensive research library. It was 2023, ChatGPT was taking the world by storm, and the path forward seemed clear.
The Failed First Attempt
Delphi's initial exploration into AI assistants revealed just how challenging the problem really was. The team integrated a cutting-edge language model into their platform and started testing. The results were concerning. The AI would confidently explain concepts incorrectly or invent token metrics that sounded plausible but were completely fabricated. Sometimes it would even misrepresent Delphi's own published positions.
"We couldn't ship something that might give wrong information with our brand attached to it," Lundy recalls. "Our credibility is everything."
Even when they tried using the most advanced models available, the economics didn't work. Each complex query about tokenomics or DeFi mechanics could cost several dollars to process. For a platform with thousands of daily users, the math simply didn't add up.
After months of frustration, they killed the project. The AI assistant would have to wait for better technology.