My attempt at a more accurate summary of the evidence from the article: some animal research labs focus on neuroscience, and some of those neuroscientists use machine learning. Sometimes, they get good enough at using machine learning that tech companies hire them. Some of the senior people -- say, the type that could lead a lab at a top university -- may eventually get paid over a million dollars per year.
My impression is that the article is trying to make a connection like "tech companies are trying to hack our brains...so they're getting in bidding wars over neuroscientists!" when the reality is more like "tech companies are always looking for people with machine learning expertise, and some neuroscientists fit that description, and in some cases their research is directly relevant".
I think the story is exacturated. There are some people who move from computational neuroscience to machine learning research, but it's not that common.
Put another way, PhD's who know how to work hard on projects that take years to complete are leaving academia for high paying jobs at big tech companies.
> To the relief of some ethicists, we're a long way from AGI, [...]
I often hear statements of this kind, but do they have basis? While there's not much to suggest that AGI will be developed soon, it doesn't seem sensible to me to say it's a long way away, as we still don't understand how difficult the problem is, nor do we have a clear path toward its development. It may be that it is achievable straightforwardly, with a breakthrough cognitive architecture.
Also, regarding ethical concerns, I don't think AGI is the problem, but rather the broader domain of super-intelligence. Plausibly, super-intelligence could result from the combination of the human mind with non-generally intelligent machinery, resulting in the same dangers.
I am convinced that we have the hardware already for AGI. Can an exaFLOPS scale Google datacenter with millions of TPUs really be less capable than the 20 watt ball of jelly we all carry around?
The barriers to AGI are in the algorithms. Nobody knows how long it will take to achieve unsupervised learning that is as general as the human brain. It might take 200 years of slow but steady progress or alternatively somebody could publish a breakthrough in a paper tomorrow and we could have parity by the end of the year.
Faithfully simulating the human brain requires much more processing power than presently available, so potentially yes. That said, we don't know how much of this complexity is actually vital to the functioning of intelligence and consciousness. Similarly, it would take a huge amount of processing to simulate a CPU at the electronic level, but our functional understanding allows us to build emulators at a high level of abstraction. The brain may be an immensely powerful computer running a simple program.
The brain is very complex, but it's not well understood how well that complexity contributes to its computation. Observable phenomena may contribute to the computational capacity of the brain, or may just introduce noise which impairs it.
Do you know of any tasks where the brain has demonstrated better performance than a computer, where the hardware is known to be limiting? I.e. where there can not be an unknown better program the computer could run.
Part of the complexity involves that it doesn't run purely on binary. It involves branches that have 4 to 6 nodes from a single input, with varying strength of signal (chemical) across a synapse, which could be a different chemical release. People are trying to simulate this with a simple on/off switch and things get very difficult, very quickly.
If you look at an integrated circuit like a CPU at the gate level, you will find complexity which goes beyond the logic being implemented -- the snaking physical routing of the traces; the width and layer of the traces; supply and ground lines; the orientation of the gates; the number of fins on a multi-gate finFET. Someone without a high-level understanding may wrongly assume that any of these things contribute to the logic.
Sometimes specific details are important, yet do not encode higher-level information. Supply lines must be able to carry enough current, so changing their size may break the logic. Clock distribution lines must propagate their signals at the correct rate, so changing their length may break the logic. Neither of these factors need be considered in a high-level emulation of a processor.
In the same way, complexity in the brain may be irrelevant to its function, redundant, or even problematic. As an example: the vascular anatomy of the brain is complex, but not generally considered to be instrumental to cognition; it's similar to the Vdd and ground, but shapes the neural matter around it. As we don't understand the mechanics of how intelligence arises from neurons, it would be wrong to assume all their properties are instrumental.
Given that an integrated circuit is an engineered system, and the brain is a evolved biological system, and evolution is in many ways more prone to unnecessary complexity because of local optima traps, I would speculate that much of the brain's complexity is likely not instrumental to intelligence (yet may be contributory).
My impression is that the article is trying to make a connection like "tech companies are trying to hack our brains...so they're getting in bidding wars over neuroscientists!" when the reality is more like "tech companies are always looking for people with machine learning expertise, and some neuroscientists fit that description, and in some cases their research is directly relevant".