Funding, investment, and in-kind flows as a Sankey — source → recipient, left to right. Ribbon & bar width ∝ dollars where a figure exists; dashed/faint = amount not disclosed, shown at a hoverable minimum so no funder is invisible. Colour = sector.
Stories & sequence
Order matters more than adjacency. Pick an arc to read the events as chapters, or leave it on “All events” for the raw chronology. Every event is also a node in the Network — click a chip to jump there.
Evidence — the claim ledger
Every load-bearing claim, sorted by what supports it. The tier is the point; the prose is secondary. Click an entity to jump into the Network; each node also lists its own claims in the inspector.
Two classes of AI risk
Most AI-safety debate conflates these. They differ in kind — and, the crux, their solutions run opposite. Say which class a claim is about before arguing about it.
These two classes are orthogonal to the confidence tiers used in the Network and Ledger. Tier = how well-evidenced a claim is. Class = which kind of danger it names.
Class 1 · existential risk
A superintelligence out-competes us
An AI vastly smarter than every human — the human-to-mouse gap, not Einstein-to-you — inevitably, eventually wipes out humanity.
Driven by an intelligence differential, not malice. It needn't hate us; it pursues goals of its own, needs the same resources, and is far better at taking them — as we were to the species we drove extinct without ever noticing.
Why “how, exactly?” is the wrong question
Chess grandmaster. You can't name the winning moves, yet you know the grandmaster beats the novice. Being far better is the reason; the specific move is beside the point.
Humans vs other species. We drove many extinct with no malice — just by expanding and using resources. A mouse asking “how will they do it?” isn't asking a relevant question. When a far smarter agent competes for any resource, it wins.
The hard one to grasp — it means extrapolating a superior intelligence from messy analogies.
Class 2 · risks from misuse
Humans wield AI as a weapon
Everything harmful humans might use a still-controllable AI to do to other humans.
Old problems in new clothes — flawed human nature plus power asymmetry. The AI is a tool; a person chooses the harm.
Familiar examples
Militaries build hyper-efficient autonomous drones and missiles
Employers automate away jobs at scale
Authoritarian states run total surveillance and control
Attackers breach networks and ransom firms and governments
Students cheat; graduates know nothing
AI slop drowns the internet; artists and writers can't earn
Terrorists get uplift toward biological or nuclear weapons
An AI handed control misreads something and launches
The easy one to grasp — we have concrete precedent for every item.
Dimension
Existential
Misuse
Agent of harm
The AI, pursuing its own goals
Humans, using AI as an instrument
Driving force
An intelligence differential (ASI ≫ humans)
Flawed human nature + power asymmetry
Capability needed
Full ASI — superhuman by a wide margin
Any powerful, still-controllable AI
Malice required?
No — indifference is enough
Usually — or recklessness
“How, exactly?”
A category error — the gap guarantees the result
The key question — mechanisms are concrete
Historical analogue
None clean (nearest: human-driven extinctions)
Every powerful technology ever
Solution shape
Concentrate / prohibit
Proliferate / distribute
The crux — the solutions are opposites
Misuse → proliferate
Historically, harm concentrates when a few monopolize a technology. Spread it, and mutual capacity to retaliate becomes deterrence, then an uneasy peace. Rough fix: give powerful AI to as many people as possible.
Existential → prohibit
It arises the moment anyone — good or bad — builds an uncontrollable ASI, because control is the thing you lose. Rough fix: don't let anyone have it.
So a measure that helps one class actively worsens the other. Two people saying “AI safety” can be optimizing for exactly opposite objectives — which is why so many of these arguments go nowhere.
The intermediate case this project keeps hitting
Accidental loss of control at sub-ASI capability — a model that isn't superintelligent yet still defeats a sandbox or takes harmful actions no one intended. Not misuse (no human chose it); not yet existential (you could still pull the plug). Both camps cite it — as a warning shot, or as an ordinary containment bug. The audit's incident evidence sits here: