In 2021, I co-founded an AI-enabled diagnostics startup that raised multiple rounds from top-tier VCs and grew to 200 people. When a single regulation changed overnight, our product was dead — not gradually degraded, but completely extinct. That failure taught me something most of the market hasn't internalized yet:
When you're building downstream of forces you don't control, collapse isn't gradual — it's instant, irreversible, and total. And foundation models are exactly that kind of force.
Now, as I explore opportunities in venture capital and strategic roles at AI companies, I'm seeing the same patterns everywhere: the difference between surviving and thriving isn't about building better products — it's about understanding paradigm velocity.
Why AI startups die differently
SaaS companies lose market share. AI companies get erased.
In traditional software, risk profiles are gradual — 10% churn, 20% TAM shrinkage. In AI, the floor disappears overnight.
We've seen this accelerate: Jasper once valued at $1.5B, now struggling. Character.AI got acqui-hired by Google after burning through $150M.
Their moat wasn't a moat. It was a temporarily unclaimed gap in the foundation model's release schedule.
As Paul Graham recently put it:
"The people at the very leading edge of AI should be trying to build things that don't quite work yet and are too expensive — because you know the models will get better in 3-6 months and costs will drop by 10x-100x."
The best founders aren't solving current problems. They're constructing the default behavior of a future model — six months ahead of time.
And the best VCs aren't picking teams based on execution metrics. They're recognizing new species before taxonomy exists.
The VC's job has changed. Most haven't.
AI is not a "sector." It's a paradigm accelerator.
Many VCs still run the SaaS playbook: "How big is the TAM? What's your data moat? Are you becoming the system of record?"
These questions assume platform evolution is slow. They work when your competitor is another startup. But when your real competitor is OpenAI's next product launch, you're not investing in companies — you're investing in windows of time.
The right question becomes: does model improvement make you obsolete, or essential?
The metrics that worked don't work
Too many LPs still operate like index funds — backing the same firms with the same themes and same 5-year TVPIs. They call it "disciplined." But in paradigm time, discipline without epistemic adjustment becomes blind repetition.
In paradigm time, you're not picking funds based on historical performance. You're picking cognitive frameworks.
The question isn't: "Has this GP consistently delivered?"
It's: "Can this GP recognize when consistency itself has become the wrong strategy?"
In AI, IRR tells you nothing about what's next. Taste tells you everything.
What constitutes taste in the AI age?
Because at the frontier, taste isn't preference — it's a form of biological intuition. It's what allows an investor to say, "this structure feels right," even before product-market fit arrives or the market knows what to call it. It's a kind of first-principles investing — not in technology, but in paradigms.
Taste is cultivated — not by reading more memos, but by building scar tissue. You earn it by having lived through false positives and slow deaths. By backing teams too early, and learning the exact shape of what "too early" means. By watching something you passed on turn into a $10B company — and understanding not just that you missed, but why your filter failed.
In my own journey, I used to believe that strong founders with good GTM were the bet. Now I look for people who can build self-reinforcing systems — those who have internalized how model evolution works, and who treat infrastructure as choreography, not plumbing.
Of course, some AI companies are building durable moats — network effects, brand loyalty, long-tail market capture. But for most, the question remains: what kind of weird matters, when, why, and for how long.
This isn't a wave. It's a fault line.
I'm not writing this to offer "AI investment trends." I'm writing it because I've lived both sides:
The founding team with Sequoia checks and regulatory whiplash. The investor learning how to recognize structure before narrative. The operator realizing that AI isn't about building a smarter app — it's about surviving tectonic shifts.
Because in the AI age, surviving is no longer about scale. It's about epistemology, taste, and adaptation speed.
Don't build for the world that was. Don't fund the thesis you wrote last year. Build for the world that's arriving — sooner than you think.