Forecasting humans becoming generally intelligent with biological anchors: year 10,000,000,000,002,022created: ; modified:
Biological anchors is a popular method of estimating Transformative AI timelines. It works by estimating the amount of computation the human brain performs and comparing it to the amount of computation computers perform to achieve various tasks humans are able to perform. Then it estimates the year in which a computer will perform enough computations to perform the hardest human tasks and uses it as an estimate of the date when Transformative AI will be developed.
In this post, I use this method to estimate the year in which humans will be able to achieve general intelligence.
It’s clear that computers right now, while very capable, are not generally intelligent. Therefore, humans will need to be at least as intelligent as computers in order to become generally intelligent.
Even if we don’t know the actual amount of compute required for general intelligence, the biological anchors method allows us to at least provide the lower bound on the year in which humans will achieve general intelligence. In this, I agree with Holden Karnofsky who notes that “Biological anchors” is about bounding, not pinpointing, AI timelines.
Floating point operations per second (flops) is the default way to measure computer intelligence, as it serves at the core of the ability of computers to do any kind of task and it’s reasonable to assume that the human brain will also need to perform this operation to complete tasks. Therefore, we are going to use it to compare computer intelligence to human intelligence.
So: M2 10-Core GPU’s intelligence is measured at 3.6 Teraflops (or 3.6*10^12 flops).
The human intelligence is measured at 0.01 flops (or 1*10^-2). (I estimated this number by attempting to perform one such operation and it took me about 100 seconds.)
Therefore, the human brain requires an increase of at least 10^15 in the number of operations it can perform per second, until there’s even a chance of humans becoming generally intelligent.
Now, how do we use this information to estimate the year in which humans might become intelligent? The best way to do this is to look at how humans have been evolving and to take a projection of it.
The previous human specie, Homo Erectus, was around approximately 1,000,000 years ago and its intelligence is measured at about 0.0001 flops (or 1*10^-4). (I will not elaborate on the details of the measurement).
This means that human intelligence increased by 10^2 flops over 10^6 years and the rate of increase in human intelligence is 10x per 10^3 years.
Plugging in the numbers:
- Lower bound for required intelligence: 10^12 flops
- Current human intelligence: 10^-2 flops
- Rate of increase in human intelligence: 10x per 10^3 years.
-> Minimum required increase in intelligence: 10^12 flops / 10^-2 flops = 10^14 flops
-> Time to reach minimum required intelligence: 10^14 flops / (10^1 flops / 10^3 years) = 10^13 * 10^3 years = 10^16 years
Conclusion: We can be confident that humans will only have a chance of reaching general intelligence in the year 10,000,000,000,002,022 or later.
- I actually have no clue how to do floating point operations. I just felt bad for humanity having 0 flops of intelligence and thought that my actual estimate of humans never becoming generally intelligent is unreasonable. Reasonabless of estimates is what makes them trustworthy, so making up a number was net good.
- I am aware of claims that my report contains several typos and inaccuracies that result in my estimate being at least several orders of maginude off. I checked the codes and I’m happy to report that the conclusions of the report are not affected.
- A few people have pointed out that the biological anchors method deals with the amount of computation required to train a model, not with the amount of computation it performs. This is incorrect. Both the amount of computation required to train a transformative model as well as the computational power of the human brain are required for the biological anchors estimate. Ajeya Cotra’s report uses the estimate of the computational power of the human brain provided by Joseph Carlsmith and estimates the computational power of a “transformative model” to be 10x of the computational power of the brain.