We should all continue to have humility around the gap between our abilities and the models. Rate of progress is blistering. Brett is higher signal on this topic than most.
Analysts who haven’t had a good call in 12 months nor update their models will say AI can’t do the job and Brett is building literal systems that do all this in rapid loops.
In April of this year, I attempted to build out AI-native coverage on my old healthcare coverage (153 healthcare stocks).
Some things worked nicely, but I hit four walls that shut down progress:
1) AI Excel wasn't good enough (the biggest issue)
2) The token bill for the workflows I was building added up FAST and I think I would have ended up well into the six-figures to do it right (a little steep for a pedagogical curiosity, lol)
3) I still had this nagging feeling that the outputs were AI slop and not incisive enough to really drive insight & conviction sufficient to construct a 15x25 paper portfolio
4) It was really hard to gather the right context without doing some heavy duty data engineering (the MCP stack was where it was today, and, again, thanks to my friends at CarbonArc, Daloopa, BigData & Canary for their pedagogical generosity i.e. giving me free pipes to explore).
Fast forward five months, I think this is possible. Or at least *much closer*.
Just today, I've updated 3 dozen models (thanks to my friends at Daloopa for raising my data call ceiling!) with a full validation loop embedded into the Skill (the ability of Opus 5.5 to adhere to a detailed Skill in the token inefficient Excel wrapper is pretty remarkable).
I was the guy ripping apart the AI slop tweets 24 months ago saying "AI can do anything a hedge fund analyst can do", but we are getting much closer to this becoming a reality. Simultaneously a little scary and a lot exciting.