Measuring intelligence against everything a person can do seems a natural reference point. Yet no particular person can do everything, and humanity distributes capabilities across specialists, instruments and institutions. Choosing that benchmark requires deciding whether the comparison concerns an individual, many experts or an organisation. The definition of the destination already contains a theory of which capabilities matter and how they should be combined.
Goldfeder, Wyder, LeCun and Shwartz Ziv propose Superhuman Adaptable Intelligence: learning to exceed human capabilities through specialisation while also covering tasks beyond us. Their abstract challenges the usefulness of familiar AGI definitions. It offers a conceptual direction for AI's future; defining that direction does not demonstrate that a particular architecture has fulfilled all its requirements. Source: arXiv
Specialising without becoming trapped
The word adaptable matters when considering the proposal. A program exceptionally good at one fixed problem can be useful without learning to solve another. In a hypothetical example, optimising delivery routes is not equivalent to reorganising a service when constraints change, new needs appear or a resource becomes unavailable. The relevant capability includes recognising what has changed and which parts of the previous solution no longer work.
Evaluating that adaptation requires observing situations sufficiently different from those used to prepare the system. Resources and assistance matter too. If every variation requires a human team to rebuild the procedure, the service may work well, but the adaptation belongs to the whole arrangement. Explaining that dependency allows comparisons without attributing to the model capabilities produced by an organisation the buyer cannot see.
The proposal also opens a discussion about choosing important tasks. What a company considers valuable may differ from the needs of a small community or a person in an unusual situation. A system may surpass specialists in profitable activities while leaving other needs unattended. Calling it superhuman describes comparative performance; it does not establish that the chosen repertoire represents every legitimate priority. Evaluation needs to make that selection visible before treating the result as a universal achievement.
Surpassing a capability does not supply a mandate
Solving a task better provides a reason to consider an answer but does not automatically confer power to determine the procedure's purposes. A model might find a highly effective resource allocation under a given objective. If the dispute concerns who should take priority, that objective needs justification and affected people need an opportunity to challenge it.
Specialisation does not remove responsibility for connections between systems either. Imagine every component of a service performing its own function brilliantly while the categories used by one differ from those used by the next. Local quality does not guarantee that the sequence preserves a request or that anyone can correct it. Deployment would need to test information transfers and preserve an institution able to answer for the complete outcome.
As recommended reading, the preprint's abstract helps reconsider an overly comfortable destination: resembling us. Its limitation is that changing vocabulary still needs to produce evaluation criteria distinguishing adaptation, specialisation and external dependency. The fruitful question is not simply whether the machine surpasses us, but in which activity, under what conditions and in support of which decision. An extraordinary capability may deserve investment without settling who is authorised to use it over other people's lives.