When society first develops any AI which matches human-level performance in any particular domain, there will probably exist an improved AI which would be many orders of magnitude faster than a human in that domain (using the same hardware as the original AI). Moreover, the design of this AI is not likely to require radical new ideas or fundamental changes in our approach to AI, but can instead be arrived at by continuing incremental improvements to a human-equivalent AI; if an improving understanding of the human brain is driving progress in AI, such an understanding can probably continue to drive progress well past the human level.

Typically the mere possibility of improvement is not very important. However, it may be a relevant input into understanding the extent to which artificial AI researchers could develop more powerful AI without needing to wait for progress in other domains.

It seems likely that the first broadly human-level AI will be built using substantially more resources than are effectively used by the human brain, or at the very least nearly as many resources as are effectively used by the human brain. Existing computer power⁺ appears to come close to the available computing power in the human brain⁺, computing power is falling in cost rapidly⁺, and broadly human-level AI does not seem likely to arrive in the very near future⁺.

It seems quite likely that there exist AI's which make much more effective use of computational resources than the human brain. The brain is designed under biological limitations, and it seems that human-level AI's could make significantly more effective use of their resources; at the same time, it seems unlikely that the design of the brain makes optimal use of available resources for economically relevant tasks.

The brain is designed under a number of limitations that may or may not apply to artificial intelligences: neurons perform a relatively narrow range of computational and storage functions and can only be adapted to other functions with some expense, formation of long-term memories is relatively slow for biological reasons, humans are not disposed to devote all of their energies to economically relevant tasks, and scaling up human brains further would result in substantial biological difficulties (for example complicating childbirth).

It appears that general intelligence has been increasing over evolutionary timescales; humans evolved relatively recently, and it seems likely that continuing natural selection for intelligence would lead to substantially improved general intelligence. This suggests that evolution 

It is conceivable that much of humans' apparent capability comes from meeting the minimal level of intelligence necessary to facilitate cultural accumulation, but it seems quite unlikely that this is responsible for all of humans' apparent intelligence or that there are no further improvements to be made.

It seems likely that genetic variations in human intelligence are driven either by deleterious mutations or design tradeoffs⁺. In either case, machine intelligences could be designed without concern for other design constraints imposed on humans, and could entirely avoid design problems associated with mutation load. This could plausibly lead to benefits which are relatively large compared to existing variations in human intelligence or in other relevant cognitive capacities.

Information learned or skills gained by one machine intelligence can potentially be shared by many machine intelligences, which may result in large economies of scale for communities of machine intelligences. The most effective machine intelligences for a task can be duplicated, training programs can be more exhaustively tested, moments of peak productivity can be stored and reused, and other similar advantages can be reaped.

We know that it is possible to build machines which perform symbolic manipulations many orders of magnitude faster than humans, and can effectively store and recall orders of magnitude more symbolic information. This suggests that at least some human capabilities can be seriously improved, and this seems to suggest that it would be possible to improve human cognition more broadly (if only by improving those aspects of cognition which we know can be improved). Some of these benefits may be achieved by brain-computer interfaces prior to the arrival of AI, but it seems unlikely that all of them could.