Here's an idea that intentionally avoids the crowded "AI copilot," "agent," "workflow automation," and "foundation model" spaces.
Startup: Counterfactual Infrastructure
Tagline: The operating system for things that almost happened.
The observation
Every organization stores what happened.
Nobody stores the universe of things that nearly happened:
- the almost-signed contract
- the candidate who came in second
- the engineering design that lost by one vote
- the medicine that failed because dosage was slightly wrong
- the movie script that was rejected before the market shifted
- the scientific hypothesis abandoned too early
- the city plan never approved
These "dead branches" are treated as garbage.
An LLM changes that because it can reconstruct and reason over incomplete, discarded trajectories.
The product
Instead of becoming an assistant...
...it becomes a historian of unrealized futures.
It continuously watches organizations and builds a living graph of:
- abandoned ideas
- rejected decisions
- failed experiments
- forgotten conversations
- alternative plans
- unrealized partnerships
Years later it surfaces:
"This proposal failed in 2024 because GPU prices were high.
Every blocking assumption has now disappeared.
Probability of success today: 87%."
It doesn't search documents.
It resurrects alternate timelines.
Example
Imagine Airbus.
Millions of engineering decisions.
Thousands of abandoned aircraft concepts.
The AI notices:
A wing design rejected in 2016.
Reason:
Carbon manufacturing impossible.
Now:
- manufacturing exists
- regulations changed
- materials changed
- fuel prices changed
The AI says:
"This project is no longer impossible."
Nobody asked.
Nobody remembered.
Another example
A pharmaceutical company.
20 years of failed molecules.
Most failures are contextual.
The AI discovers:
Drug X failed because:
- patients weren't genetically segmented
Today:
New biomarkers exist.
The same molecule now has a viable patient population.
This is worth billions.
Even stranger
Governments.
Suppose every law proposal ever rejected is embedded.
The system notices:
This climate policy failed because batteries were expensive.
Current economics make it viable.
It recommends resurrecting legislation from 17 years ago.
The technical innovation
Current RAG systems answer:
"What do we know?"
This answers:
"What did we stop believing?"
The architecture is different.
Instead of embeddings over documents:
Build embeddings over:
- abandoned assumptions
- causal chains
- reasons for rejection
- dependency graphs
- uncertainty
Every rejected idea becomes a living object.
As the world changes, the object changes.
New primitive: Assumption Drift
Every decision depends on assumptions.
Example:
Launch product.
Because:
- market too small
- GPUs expensive
- regulation absent
- CEO against idea
Years later:
Market:
✓ changed
Regulation:
✓ changed
CEO:
gone
GPU:
100× cheaper
Decision should be recomputed.
Nobody does.
Counterfactual Infrastructure does.
Business model
Charge not for storage.
Charge for rediscovered value.
Examples:
"We found $18M of abandoned IP."
"We identified 7 patents worth filing."
"We found 32 projects whose blockers disappeared."
Revenue share.
Like venture capital for forgotten ideas.
Customers
Initially impossible to sell.
Eventually indispensable.
- Aerospace
- Pharma
- Semiconductor companies
- Law firms
- Universities
- National laboratories
- Governments
- Fortune 500 companies
Anywhere institutional memory exceeds human memory.
The moat
Every year the system gets stronger.
Not because the AI improves.
Because the archive grows.
Eventually it possesses the largest map of unrealized history ever assembled.
Competitors can't recreate 15 years of discarded internal reasoning.
Why LLMs uniquely enable this
Before modern LLMs:
Abandoned work was too unstructured.
Meeting notes.
Emails.
Whiteboards.
Slack.
Half-written documents.
Voice recordings.
Now all of that becomes machine-readable.
Not merely searchable—but transformable into structured causal narratives.
The really weird extension
Don't stop at one company.
Aggregate anonymized "dead ideas" across industries.
Example:
A mining company abandoned a sensor because batteries were inadequate.
A robotics company solved that battery problem years later.
Neither knows the other exists.
The AI creates a cross-industry resurrection.
Entire innovations emerge from stitching together failures that occurred in different domains.
The long-term vision
Imagine an internet where every abandoned idea has a persistent digital identity.
As technology, economics, regulation, science, and culture evolve, those ideas are continuously re-evaluated.
Innovation shifts from inventing entirely new concepts to identifying which previously impossible ideas have quietly become possible. The AI becomes a global engine for "temporal arbitrage," exploiting the gap between changing reality and outdated decisions.
It's an unusual direction because it treats history—not generation—as the primary raw material. Rather than producing more ideas, it aims to recover the immense value hidden in ideas that were discarded under conditions that no longer exist. If successful, it could reshape how organizations think about R&D, institutional memory, and strategic planning.
