Article: Neuralink and the Brain’s Magical Future (Wait Buy Why)
Tim Urban with a primer on brain physiology and brain-machine interfaces as context for the ‘what’ and ‘why’ of Elon Musk’s Neuralink
Prior to this article, I’d read some short pieces on Neuralink out of passive interest, but confess to have barely grasped the challenges and future products of the company. I’d wager I was in the majority there.
Enter stick-figure savant and pseudo-layman Tim Urban of Wait But Why.
‘Long form’ doesn’t totally capture the scale here. This 38,000+ word piece endeavors to give the reader enough context on brain physiology and brain-machine interfaces (aka the devices Neuralink will build) to get a handle on the roadmap and barriers for Neuralink. That roadmap is packed with candid exchanges between Elon, the Neuralink team and Tim that alone make this worth your time.
The structure of the piece exists in a series of chapters Tim describes as branches on a tree:
I’ll avoid summarizing the piece any further, and instead provide two main reasons you should read this as someone trying to gain fluency in machine learning.
1. Imitating the Brain
While not every innovation in neural network architecture is linked to a biological equivalent, many researchers will note a brain feature or function as their inspiration (i.e. the hippocampus as the brain’s “scratch board for memory”). It helps to not be totally in the dark about neuroscience.
Tim takes special care to reinforce that the brightest neurologists admit to knowing relatively little about brain function. That said, there is a lot to learn, and as usual, I found Tim effectively covered the basics with cheeky illustrations, tangential video links and more.
You’ll get a good sense for which parts of the brain govern which functions, how a neuron works, the scale of neurons (and synapses) in our brains and how our neural pathways change over time.
2. The Case for Democratizing AI
Once you get through Parts 2–5, you arrive at the ‘why’ of Neuralink, which centers on the future need to provide wide access and availability to AI.
As I read more about the risks of AGI and superintelligence, there is a clear opinion divide on how consolidated the world’s collective AI capabilities should be.
On one extreme, there is an ‘open source’ view that believes distributing capability among more users is our best option to mitigate the downside risk of malevolent use. This is Elon Musk’s view today and he speaks to that effect in the piece:
“Essentially, if everyone’s from planet Krypton, that’s great. But if only one of them is Superman and Superman also has the personality of Hitler, then we’ve got a problem.”
That’s a hilarious over-simplification of Elon’s view — thousands of words in Tim’s piece are dedicated to this — but you get the point. He also believes AI integration to brain function may be a path to solving value-alignment issues. If you’re looking for a good handle on the ‘democratize’ argument, this article is a good resource.
If you want a complementary overview, Open Philanthropy’s recent rationale of their $30M grant to OpenAI is also a worthwhile, concise read.
For the case against democratization, I’d recommend Slate Star Codex’s ‘Should AI Be Open’ and the accompany comments section:
All said, this is a complex subject fraught with predictions about the far future, so naturally there’s a significant debate. Even assuming malevolent use can be mitigated, issues like rapid takeoff from AGI to superintelligence and the accompanying control problems are yet unsolved.
Are brain-machine interfaces a panacea for value alignment, control problems and a host of other AI risk related issues? Not yet. But Elon believes firmly that it’s worth exploring. Think this exchange puts it best:
When I thought about all of this, one reservation I had was whether a whole-brain interface would be enough of a change to make integration likely. I brought this up with Elon, noting that there would still be a vast difference between our thinking speed and a computer’s thinking speed. He said:
“Yes, but increasing bandwidth by orders of magnitude would make it better. And it’s directionally correct. Does it solve all problems? No. But is it directionally correct? Yes. If you’re going to go in some direction, well, why would you go in any direction other than this?”
