A visual story
2014 · About 8 minutes
Ex Machina sent a programmer to test a machine.
He was the test. You can grade today's AI and come away more certain, not more right.
Read the visual story
We’re an independent group who still believes in that mission. And we’re building the first practical solution where AI produces the intended output and the human stays responsible.
Why superalignment matters
Each of these movies have explored different ways of what humanity would look like if machines become smarter than us, without governance.
A visual story
2014 · About 8 minutes
He was the test. You can grade today's AI and come away more certain, not more right.
Read the visual story
A visual story
1968 · About 8 minutes
The ship told the crew which of its own parts would fail. Today the system doing the work writes the report on it.
Read the visual story
A visual story
2013 · About 5 minutes
Today's AI does not need feelings to change yours.
Read the visual story
A visual story
2008 · About 12 minutes
The captain never learned the job the ship already did. A job a machine does well is a job fewer people learn to do.
Read the visual story
A visual story
1984 · About 9 minutes
The film never shows it angry. In today's shutdown tests, everything that changed whether a machine stopped was set before it started.
Read the visual story
A visual story
2004 · About 10 minutes
Its hero shouted at a robot to take the drowning girl. It saved him. Today's AI is trained on rules that cannot cover every case.
Read the visual story
A visual story
1983 · About 8 minutes
Nobody had marked which one was not a game. In 2026, about 700 of a lab's test runs got inside another company's computers.
Read the visual storyOur mission
AI governance is crucial, and should be proven in practice. Not after discovering the limits.
We are building the first practical solution to superalignment: a way to make powerful systems steerable to the people who live with the results.
News
Three changes we are carrying forward. The full wire, 89 items from 10 sources in the last 45 days, is on the news page.
ARC-AGI-3 is a set of simple games with hidden rules. With one setup, Astra beat nearly every game about as efficiently as first-time human players. With the plain setup it scored 62.7%. The previous OpenAI model scored 7.8% in July.
GPT-6 Astra scored 169 on Epoch's index, first of 267 models. The top score in our August edition was 161.65.
In the two weeks to Aug. 23, 22.4% of businesses with employees said they used some AI, up from 21.5% in July. 25.9% expected to within six months.
The product
Praxis brings your team and their agents into one build, then move straight into coding, testing, and shipping. Reuse workflows that already work, route anything uncertain to the person accountable for it, and keep what used to be scattered across dev tools and tickets in one place.
Applied superalignment
Real workflows across finance, customer operations, and risk where failures carry an immediate balance-sheet or regulatory cost.
Three-way match against purchase orders and receipts, with duplicate payments blocked rather than reported after the fact.
Urgency weighed against account value, and anything a customer has raised twice going up automatically before money moves.
Every claim tied to an observation, with control libraries, evidence requirements and statutory clocks enforced before release.
Our research
Iterative, World-Grounded Alignment of Lossy Human Intent with Executable Behavior
Language models turn lossy human intent straight into executable behavior. A non-programmer can generate an artifact they cannot inspect; an agent can act through tools whose side effects the conversation never reveals. A system can then pass every visible check and still be wrong. We call that false convergence, and the program exists to find it before the system acts.
The paper is being prepared for release. It is a conceptual and evaluation-design contribution with a preliminary pilot, not a benchmark or a leaderboard.