We estimate that ‘human-level hardware’— hardware able to perform as many computations per second as a human brain, at a similar cost to a human brain—has a 30% chance of having already occurred, a 45% third chance of occurring by 2040, and a 25% chance of occurring later. We are not confident about these estimates.
Say that computing hardware has reached ‘human-level’ performance when machines can perform as many computations as a human brain performs (under some natural interpretation of the brain as performing computations), at no greater cost than that of running a human brain.
We are interested in when hardware reaches this level of performance, because then if we have software that is also at least ‘human-level’,1 we will have ‘human-level’ AI overall: AI that can perform the tasks that a human brain performs, as efficiently as a human brain.
We may have human-level AI before having hardware and software that are both at least as good as the human brain—intuitively, if one of them is better than that, then the other may be worse. So the implications of human-level hardware alone are not straightforward. (We may also have disruptive change before we have human-level AI—the event of human-level AI is an upper bound for when some large changes in society can be expected to occur.)
It is unclear to us how important the contributions from hardware and software progress are, respectively, to overall AI progress. If hardware progress is much more important than software progress, then human-level hardware should approximately co-occur with human-level AI.
To roughly forecast when computing hardware will reach ‘human-level’, we can combine estimates of how much computing the human brain performs per dollar, current hardware performance per dollar, and the rate of improvement in hardware performance.
T = time until human-level hardware performance per dollar.
H = human-level hardware performance per dollar = 0.4-13*1010 FLOPS/$.2
C = current hardware performance per dollar = 0.3-30 *109 FLOPS/$3
R = 1 + growth rate of hardware performance per dollar = 1.16-1.784
Then we have:
T = logR(H/C)
=log1.16(0.4 x 1010/(30 x 109)) to log1.16(13 x 1010/(.3 x 109))
= -14 to 41 years
These are rough calculations, and the breadth of the intervals don’t necessarily mean a lot—the intervals were non-specific to begin with, and then we combined several of them.
If we do something similar (shown here), using more realistic distributions for each variable and calculating using the entire distributions rather than end points, we get -14 to 22 years using the narrower estimates for human-level hardware that we used above, or -31 to 99 years for a very wide set of estimates for human-level hardware. The chance that human-level hardware has already occurred is around 20-40%, according to these calculations.
Based on these calculations, we estimate a 30% chance we are already past human-level hardware (at human cost), a 45% chance it occurs by 2040, and a 25% chance it occurs later.5
These figures suggest that the period when we most expect human-level hardware has already begun, and we are a substantial part of the way through it. In the case that hardware progress matters a lot more than software progress, this means that we should expect to see human-level AI in the next several decades, or possibly in the past. This is some evidence against hardware progress being so important, but still overall makes human-level AI likely to be sooner than one might have thought without the evidence considered here.
- Here we say software is ‘human-level’ if it uses hardware as efficiently as the human brain to produce any behavior that a human can produce.
- We estimate that the human brain performs around 1—34 x 1016 FLOPS, based on using a communication benchmark (TEPS) as a proxy for computing. This range does not represent our uncertainty about this method’s reliability.
Humans earn very roughly $100/hour. This means that purchasing computing hardware that costs as much as a human per hour, and lasted for around three years (as computing hardware often does), would cost $2.6M upfront.
So we should consider hardware to be competitive with human brains when it performs somewhere between 0.4-13*1010 FLOPS/$.
- In 2017, cheap hardware appears to perform around 0.3-30 *109 FLOPS/$.
- The price of hardware appears to be declining at around an order of magnitude every 10-16 years. However in the longer term, the rate has been an order of magnitude every four years.
- This is based on weighting the narrower estimates for what constitutes human-level hardware at 60% and the broader ones at 40%.
It would be good to re-evaluate & update this blog based on the latest advances in hardware (gpu/tpu/asics etc) & software (RNN, tensor flow, autoML etc)
Or at least state when it was written?