AI Infrastructure for Investors – How to Start

Every call I get that question: “What about AI?”. Yes, what about it? If you have any summer interns they’ll probably make some proposals how to automatically source and select investment winners with AI.

With investors — LPs and GPs alike! — I rather like to talk about “AI Infrastructure”. Because most of them have experience in building data-driven organizations (remember “big data”?!).

Infrastructure: Data, AI, People, and Otherwise

Any investment of time and money into infrastructure (beyond incremental improvement and keeping the lights on) has to withstand 3 minimum requirements or tests:

1. Unfair advantage. Is the data you’re acting on, or the connections you make between different data points, proprietary or privileged? Will that create alpha or is it just “novel”?
2. Make decisions better/faster. But how much better/faster? And can you act on decisions? Will that create alpha or is it just “novel”? Could it lead to rash decisions?
3. Ethos. Does the infrastructure and its processes reflect your character, your cadence, your people? is that what your investors or stakeholders are underwriting? Can you communicate that?

#1 is particularly important for AI Infrastructure: If LLMs are constructing sentences based on most likely next tokens, then any competitive advantage based on generally available public information will quickly regress to the mean: all beta, zero alpha. Perhaps that might take a while, say 1-2 years, at most. Just like “Big Data”. And do you think the large banks and their Quants, or the very best VCs or private equity investors, are not looking at AI? Being the first in your company or the first regarding your LP base is not sufficient.

7 Steps For AI Infrastructure

My advice for a starting point is sheepishly simple. Sorry. Here are seven steps:

1. Pro or Enterprise: Subscribe to a pro or enterprise version that lets you build custom models and share models with employees and even external users.
2. Deep Research: Use deep research often.
3. Custom Models: Build several custom models for different subsegments or asset classes or conviction areas. They could also be competitive models for the same subsegment. But don’t divide your people, make it a healthy competition, not an ego trip.
4. Proprietary Data: Collect proprietary data that you feel is relevant to your problems or your LP’s challenges. Always annotate where you can – never just upload some document or conversation or data without context or key points or why you want to ingest this information. Clearly mark context from convictions from hunches (“I’m uploading X because of Y. I am convinced that A is not true, B is not a factor, and that C will happen. I have a hunch this could lead to P and Q. I wonder how this would affect our thesis in terms of D/E/F.”) Discuss your insights with the AI (which is another form of annotation and sentiment building)
5. Prompt Design: Create and iterate prompts with 3 purposes:
   • Discovery (brainstorming, loose association, sounding board)
   • Scenario Analysis
   • Counterfactuals, mostly near-world
6. Feedback Loops:
   • Assign probability values to scenarios and update and replace data often.
   • Correct wrong assumptions or misconceptions by your model right away.
   • Check both the LLM and your prompts (!) for risks regarding covert subversion, in-context scheming, goal nudging, alignment faking, sandbagging, strategic deception
   • present findings to stakeholders and gather feedback on storytelling, communication style, reasoning, character (not content!)
   • Ask your own model who else came to the same conclusion, who else is working on a similar area, who else is publishing similar content, who else is investing? It will help you estimate your actual alpha and find valuable future co-workers, advisors, multipliers, platforms.
   • Ask your own model who has an opposing view or approach: Against whom are you betting, and what are their arguments?
7. Sunset Models: Regardless of how much effort you spent or how big your prompts became: no ego, no sunken-cost bias! Focus on two models with the highest alpha potential given your risk target, cadence, character, people, operations. Select one true moonshot model, but not one that would be contrarian to your current fund product. Maintenance and management of these models take time, and I doubt you will have time and resources to build and maintain proprietary insights for more than 2-3 models. These models should already surface counterfactuals (per 5. above). Selecting a moonshot model that goes against your investment thesis or operational model, cadence, or character will be a waste of your resources and a distraction … and not something that your investors or stakeholders were underwriting in the first place!

Bonus: But Thor, what platform should I use?!

Like with publishing platforms or programming languages, the answer is the same: choose the platform where you (a) know someone friendly and helpful who will take your call when things get hard; or (b) have access to people who have done it before. Where can you get help the fastest when you need it?