Why Companies
Still Hire Engineers
The layers of context AI cannot know. Lohit Talasila on the human judgment behind production software.
Intelligence for everyone
Steered by people
More powerful AI could help create a world of abundance. Our work is to make it steerable, aligned with human intent, and accountable to the people who live with its results.
Aligned
Read all storiesFrom the Superalignment manifesto
Every amplifier we built before this one amplified what a person could do. This one amplifies what a person can decide.
What AI can do. What the headlines leave out. What it means for you.
Why Companies
Practical Edge
Yes, Claude can build it. Your edge is knowing what good looks like.
Where human judgment moves when more execution is delegated.
Where AI stands
Capability
What can AI actually do?
Capability and reliability, with dated evidence and the limits of each test.
Delegated power
How much power have we handed over?
The tools, access, and authority that let AI affect the real world.
Real-world change
Where is AI changing everyday work?
Adoption and economic change, read alongside the capability headlines.
Protection
Do the safeguards hold?
What has been tested, what is claimed, and where protection remains uncertain.
Case studies from Praxis
Case study · Insurance
AI sales training for real conversations
NÜRNBERGER uses simulated customer conversations to help insurance agents practice and receive feedback before meeting clients.
Case study · Procurement
A clearer path to software renewals
VusionGroup combines procurement workflows, supplier comparisons, and expert negotiation to manage software renewals.
Case study · Healthcare
Less paperwork for pharmacy teams
Link Pharmacy automates prescription intake, faxes, and record updates, with pharmacists responsible for verification.
Research
Research program
Convergence Programming
Our research program for aligning human intent with executable behavior. The paper is being prepared for release.
The Library
Read the research behind the ideas.
Browse the Library, with source records and explanations that distinguish evidence from interpretation.
Research
Questions we are working on.
Explore the methods and open questions behind practical, measurable alignment.
Wiki
Start here
What is superalignment?
How do we check the work of a system that can do more than its supervisor?
Concepts
When passing tests is not enough
Understand false convergence: the signals look right while required behavior remains unchecked.
Concepts
Who gets to decide what AI does?
A practical introduction to control, authority, and the ability to stop a system.
Reading evidence
What does an evaluation tell you?
Learn to read a test result, its scope, and the questions it leaves open.
Research program
Convergence Programming
Our research program for aligning human intent with executable behavior.
The paper is being prepared for release.
News
Dated briefings on the changes worth carrying forward.
Briefing · Sep 7, 2026
A new AI model nearly maxed out one test. Europe delayed its rules.
The September briefing: what changed, what it means, and the primary sources.
Briefing · Aug 18, 2026
AI test scores rose fast. The wider economy did not.
The August briefing puts model progress beside adoption and real-world change.
The Wire · Automated leads
Follow the primary sources.
The Wire collects automated, unreviewed leads. Use it as a starting point for reading.
Culture
Familiar stories. A different way into the questions AI raises.
Culture · Visual story
Her: The wider frame
An intimate conversation becomes a question about the reach of an AI companion.
Culture · Visual story
Ex Machina: Who is testing whom?
A familiar story opens a question about observation, trust, and control.
Culture · Visual story
WarGames: When a game has consequences
What changes when a system can act beyond the screen?
Culture · Visual story
2001: One system, conflicting instructions
A story about the gap between a command and the world it must work in.
Get early access to Verity.
A dedicated Verity waitlist is being prepared. For now, email us to request early access.