Rudyard Kipling's "If" has been on my mind: "If you can keep your head when all about you..." The companies I follow are being disrupted by AI at speed. I'm trying to avoid both the paralysis of p(doom) and the hype cycles. Here is where I've landed, and where I haven't.
Obvious problems, unclear moats
Intelligence is being democratized globally. Users can switch applications at negligible cost. New products can take share quickly from a standing start. Despite the scale of the spending so far, the moats are not obvious.
The problems are. Data quality and integrity. Security. Affordability. Energy reliability. Businesses solving those have demand that depends far less on which model wins, and that is where my search for companies solving real, recurring problems already points.
In some ways that is the easy part. The harder debate is how and where AI gets used inside organizations, including my own.
Two rules
To keep things simple, I've always held two rules on AI:
First: if a tool gives you far more than you pay for it, you are probably what’s being sold.
Second: never give a system information you wouldn't want on the front page of the New York Times. It was the most practical advice I received from my first supervisory analyst on the sell side.
Cymer's AI policies are more detailed than this, but they rest on those two principles.
The harness
This year, much of my time has gone into building what AI people now call a harness: a research taxonomy and process designed to stay flexible and provider agnostic as the technology changes. The work isn’t finished. LLMs are built into the initial research stage to surface ideas, an orchestration that once would have required several analysts.
The quality of data going in matters as much as checking the quality of what comes out, just as it does with human only analysis. The downside is still underwritten first. Outputs are still checked against source. I’m less concerned with the model running under it. I’ve tested most of the major ones, and they will keep changing. The process shouldn’t have to.
Toward a private server
The next step is running the same process on open weight models on a private server. That would address the main concerns around security, data safety, intellectual property, and dependence on third party vendors' decisions. It isn't necessary today. But the direction of travel is clear, and being ready and able to make the move matters more than making it first.
What I still can't resolve
I have theories on how this plays out, not firm evidence. Three questions stay open for me:
If switching costs stay near zero, does value accrue to the model layer at all, or to whoever owns the data, the distribution, or the power?
How much of today's enterprise adoption holds up once security and data leakage risks are fully weighed?
How far can the more leveraged second tier keep funding itself if credit keeps tightening?
I'll keep working on all three in public as the evidence comes in. Until then: keep calm and carry on.
Musical coda
For informational purposes only. Not an offer to sell or a solicitation to buy any security, or investment advice.