Testing Deep Research: $UNH and the Medicare Crisis Hedge funds, overall, have been "slow adopters" of large language models. Early LLMs had poor numeracy, limited citations and a tendency for rampant hallucinations...not great attributes of a trustworthy tool for institutional equity analysis. Candidly, while I started a webinar series (The Cutting Edge) to explore AI tools in the investment process in the fall of 2024, up until about ~2 months ago the counsel I would give those who would ask is "pay attention, but no need for meaningful process re-engineering", with the view that the cognitive load & disruption to shifting an already established investment process to one that was AI-augmented was, on balance, not worth the hassle & the risk of erroneous analysis. And for all the talk about augmenting the investment process with AI, outside of few forward thinking "hackers" that pulled together some useful systems, most agreed with me that the tools were not ready for showtime. Over the last 1-2 months, that view has shifted. I see more and more areas of the research process where I believe these tools should be used now. A couple things clicked for me. First, "recursive prompting" has made it much easier to build useful prompts that generate useful responses. I personally use Claude Sonnet 4 to build prompts, and this allows me to speak like most finance people do (a few grunted sentences into the context window). It's so easy. And the output is so good. For example, Claude decided on its own to priorities these primary sources (agree). And built this "Special Instructions" section to avoid speculation & seek diverse sources. The dimensions that are generated from this recursive prompting structure allow me to ask a question like a normal human being, generate a ~6 page structured prompt, and get an extremely good result. I tested this out on a situation where I have some sense of what good looks like. From 2010-2021 I was a healthcare services analyst, and a damn good one. I initially hated covering healthcare services (the first investor conference I went to it felt like they were speaking Greek). But I soon learned that there was real alpha in understanding the intricacies of how these businesses work and in arbitrating the various debates that would arise (impact of minimum MLRs, exchange dumping, risk corridors, Stars scores, etc.). I loved it actually, it felt like putting together a complicated puzzle. The work was *incredibly labor-intensive*. But the half-live of investible insights in healthcare services is short. Recently, I will occasionally get a question on a name I used to know cold: United Healthcare ($UNH). In building Fundamental Edge (and raising 3 sons), I haven't had the time (or resources) to stay super current in the space. So I don't know the name so cold anymore. (it has been cut in half in recent months...a truly shocking outcome). I figured I would spend some time this Sunday evening running my recursive prompt playbook through four Deep Research tools (ChatGPT, Gemini, Claude & AlphaSense) to get "up to speed" on the $UNH situation. 1) I want to understand in depth what has happened & how we got here 2) I want to consider the reaction function menu of both companies & regulators 3) I want to quantify those scenarios, then compare that to what is baked into the stock, i.e. what is "UNH ex-MA" floor assuming 0% margin for the MA business. 4) I want to activate my antennae for what to look for from a catalyst path perspective..."if this happens, that is a sign CMS gets it and is supporting this industry" or "if this happens, MA is cooked and will see a prolonged retrenchment". Then, in the seat, I'm dog after a bone looking for those clues. To do this correctly, I'm updating all my HMO models, my HMO data sheet, catching up on the last 8 call & investor conference transcripts for the MA companies, then doing a deep (and super nerdy) reading stack of CMS reports, MedPac reports, Kaiser reports, etc. And speaking with a dozen+ experts, the key sell-side analysts, and representatives from the companies (multiple times). Out of all of this, a mosaic of insight starts to form, and that mosaic informs really one quantitative output: 2028 MA revenue & margin (and the trajectory to that point). This is what I think most silicon valley types don't get about stock research. Most stock situations can really be boiled down to 1 or 2 differentiated insights. Everything else in the process is an (important) wrapper around those insights. A standalone AI-summary of a 10-K or a mediocre AI-slop company primer doesn't really do anything useful. And accelerated pathway to deeper insight, with more conviction, on a differentiated view of the key driver of the business - now we are talking. AI can be marginally helpful if the tools can help build the wrapper with more speed & reliability, but they can be truly game-changing if they can both accelerate the speed and enhance the quality of insight & differentiation on the key drivers of the business. To me, this is what these tools are becoming. Don't take my word for it. Try this out on your own. 1) Identify an investment debate 2) Ask an LLM (I prefer Claude Sonnet 4) to create a detailed prompt from your plain-English question 3) Run it through 3-4 Deep research wrappers The reports I have sitting here are mind-blowing, to be honest (except Gemini. Gemini sucked, and Claude Deep Research was "meh"). But the combination of reading the ChatGPT report and AlphaSense report had me honestly feeling pretty up to speed (with the caveat that I have a decade+ of context to fall back on). It's not the ending point of the research process, by any means, but with these reports I feel ready to start getting on the phone and doing my field research. The scary part is ChatGPT 03 with Deep Research, in this instance, is not integrating any of my own research. If I could pipe in models, data sheets, mgmt meeting notes, sell-side research, expert network transcripts...it's scary how good this can become, and in short order. Enough to take me out of the skeptic camp, into the converted. I highly urge you to try it out for yourself (and let me know your thoughts).