Manifesto
Nine commitments for building toward superintelligence.
What we think good work in this field looks like. Keep what holds up, argue with what does not.
Intelligence is the most general technology.
Every other tool solves one class of problem. A system that can learn, plan and invent solves the problem of solving problems. That is why the stakes are higher here than anywhere else in engineering.
Superintelligence is a direction before it is a date.
Forecasts disagree by decades. What can be observed is the direction of travel: systems that do more kinds of work, for longer, with less supervision. We plan for the direction and stay honest about the date.
Measure before you claim.
A capability that cannot be tested is a marketing line. We publish how a result was measured, on what data, and where it fails. Intelligence has no unit yet, so the method matters more than the number.
Capability and alignment ship together.
A system that is more able and less predictable is a regression. Safety work belongs in the same team, the same release and the same budget as the capability it constrains.
People keep the ability to say no.
Oversight only counts if it can change the outcome. Systems should accept correction, expose what they are doing, and stop when asked. This holds most of all when stopping is inconvenient.
Share knowledge openly, deploy carefully.
Understanding spreads through papers, evaluations and open tools. Deployment is different: it is staged, monitored and reversible. Openness about how things work is compatible with care about what gets released.
The benefits should be broad.
Minds that outperform ours in science and medicine could end scarcities that have lasted for all of history. A future where that value reaches only a few owners is a failed project, however impressive the technology.
Uncertainty is a reason to think harder.
Nobody knows how this goes. That argues for more scenarios, more red teams and more dissent inside the room. Confidence in either direction, doom or ease, has not been earned by anyone.
Build as if it matters, because it does.
The first systems smarter than their makers will be built once. The habits formed now, in small labs and small tools, are the habits that will be in place when it counts.