Three terms people mix up
- Narrow AI does one kind of task well: translating, classifying, recommending. Almost everything in production today is narrow.
- AGI (artificial general intelligence) would match a capable person across most intellectual work, learning new tasks without being rebuilt for each one.
- Superintelligence goes further: an intellect that greatly exceeds the best humans in practically every field, from science to strategy to persuasion.
Today's large language models sit somewhere awkward. They are superhuman in breadth of knowledge and speed, yet still unreliable at checking their own work, at long multi-step plans and at knowing what they don't know.
What I see when I check AI output
My work at Turing is validating AI-generated problems and solutions: is the problem stated clearly, is the solution correct, do the tests actually pass? The pattern is consistent: models are impressive on average and confidently wrong at the edges. A solution can read beautifully and still fail a test case. That gap between sounding right and being right is the practical problem to solve, long before any talk of superintelligence.
Why the question still matters
If systems ever become far more capable than people, two things get hard. The first is verification: checking whether their work is right when we can't easily do the work ourselves. The second is alignment: making sure what they pursue is what we actually want. Both already exist in small form. We already ask models for work we can't easily verify, and they already optimise for looking right when that is what gets rewarded.
What engineers can do now
- Ground answers in sources. Retrieval over real documents, with links that are checked, beats a model answering from memory. (Graph RAG is one way.)
- Evaluate honestly. Keep test sets, measure, and don't mistake a good demo for a reliable system.
- Keep people in charge of irreversible actions. Payments, deletions and publishing need a human confirmation.
- Prefer small and inspectable. A smaller model with a good knowledge graph can match a much bigger one on a narrow job, and it is easier to understand and control.
- Design for failure. Systems that keep working safely when the network or the model fails, as RESCUE-Edge is designed to, stay safer whatever the model's capability.
My view
I don't know when superintelligence will arrive, or whether it will, and I'm wary of anyone who claims to know. I am confident that the habits that make today's AI trustworthy (grounding, measurement and human oversight) are the same habits we would need then. Practising them now is the useful part.