Claude has a memory problem

 (^me working on my clickbait titles)

Claude (and ChatGPT) has gotten pretty amazing at conducting long-form tasks. But it still lacks a long-term, shared-across-the-entire-team memory, or what we used to call a "database." For example, I am doing research on X industry, asking Claude 10 questions in succession. Some of the links I click through and read (and don't like -- too low quality); some of content I disagree with and mentally ignore; some of it is repetitive or off-topic. 

Currently, there is no way to share the distillation of your thoughts. (Also - distilling your thoughts to just the essence is hard!) Claude (and other LLMs) don't make this easy. You can share a whole chat, but that is a huge pain to read through. You can Claude to give you a summary, but it sometimes doesn't capture things correctly. The only way -- at least, in my opinion -- is to transcribe the highlights by hand, and/or summarize it in your own words (i.e. old-fashioned reading and writing). I still subscribe to the belief that "writing is thinking," so the time taken to write is time well spent. 

In investment-land, institutional memory takes the form of memos (or investment decks). However, few places take the time to write up all the info on deals they did *not* participate in -- the effort is too high. The big opportunity, then, is lowering the activation energy it takes to turn research (i.e. reading an article, cruising through Google, skimming LinkedIn, interrogating Claude) into saved, institutional, and trustworthy memory. I think the workflow model that will work best long-term is human-as-curator (and AI as secretary and first-pass analyst). The big challenge is designing a workflow and data structure that fits, and the institutional buy-in to think this way.

Source / Pick / Win

 A little bit of introspection on being good at source / pick / win ... 

1. Sourcing is both "becoming commoditized" ... but also isn't. It *is* easier than ever to build AI tools to help list out all the companies in the known universe. It is *significantly* harder to get founders to respond to cold emails, without some sort of brand name or reputation (or in the worst case, a desperate need for cash). 

Nothing will beat having boots on the ground, spending quality time face-to-face with a founder, building rapport and telling your story. I've found a few start-ups that (a) I would've otherwise never found and (b) who otherwise wouldn't have responded to my emails. 

Sourcing is a learned skill, an extremely process-oriented step, and a very human process. It's one that I've been slow to uptake, but getting better at.

2. Picking is the due diligence. AI will be able to help accelerate diligence, but I believe good investment taste will be hard to find. The apprenticeship part of the business, one I continue to cut my teeth on (writing memos by hand, assisted with AI searches).

3. Winning is the very human part of it that will forever be human. Who do you want to be invested in you long-term? Who brings value outside of capital? Who do you trust? Who do you like? It seems crazy that you would take money from person A than person B just because you like being around person A and find person B annoying ... but that is a very real part of the process. 

This is the piece that spending 8+ years at Epic learning to build trust come in handy. 

VC Notes: Part ?

 Recently got looped into a conversation about a start-up in a little bit of trouble. Another fund (the lead) was spending a lot of time with it. Question came up: has the fund had any big exits yet? (Do they need this one to do well?)

Never thought much about this -- that a fund really really would like a few big early exits (because that helps raise the following funds, and a variety of other reasons). Incentives can get misaligned with start-up founders (or the success of the company). Weird behavioral quirk you don't think of if you're not at the table.

AI: a race to turn probabilistic into determinstic

A newer way that I've thought about the phase of AI and LLMs that we're in: it's a race to turn probabilistic/stochastic ("random") processes into deterministic ("not random") ones. 

• Easy example: adding two numbers together should use a calculator, not an LLM. Bad use case of AI.
• The company tracking suite I'm building out is built using the help of AI, but ultimately it's to build a strong database in the format I want, so that I can store companies, market maps, data, thoughts, etc. for the long-term. In my opinion, a good use of AI: turning a noisy, messy process (AI-driven coding) into something reliable (a deterministic database)
• LLM/agentic loops are all trying to build processes to reduce the error rate to close-enough-to-zero (e.g. double-checking/validation agents) or putting humans in the loop to bring it to "zero." The challenge is that you can reduce the probability of error close to zero, but can never quite hit it. I see this in the "document processing" realm -- we hope that the frontier models can read the PDF with 100% accuracy ... but for high stakes scenarios, would we trust it?