I don’t have any knowledge. My guess is that they are decoding visual text from a non human brain. Which should be very doable. Probably even easy. How to cheat if the device is not great? Make the dictionary of sentences small. Make presentation slow.

https://twitter.com/neuralink/status/1595643786872225794

but it may also be spoken text

AI and natural intelligence (NI) appear so different.

A group of us asked “why?”

The answer: we have to go back to the future!

arxiv.org/abs/2201.07372

A thread.

1/6 NIs (e.g., animals) appear to learn for the future while AIs usually learn for the past (i.e., classical empirical risk minimization).

AIs predict for new experiences only if they arise from the same distribution as old experiences.

2/6

I think that the brain almost certainly approximates gradient descent. And here is why: Any learning episode only appears to change the brain a tiny bit. 1/5 Given that the brain appears to be quite noisy and somewhat linear, this means that we can almost certainly locally approximate the task loss L linearly around the starting parameters W_0 as L(W_0+ΔW)≈L(W_0 )+ΔW ∇L(W) 2/5

People sometimes wonder why i am so worried about popular ideas in neuroscience being misleading. Well, I joined my undergrad and later PhD lab to show that the brain was doing binding by synchrony. My thesis proposal with @konigpeter was truly ambitious, using natural scenes and in vivo recordings to truly convincingly show that the brain does binding by synchrony.

Science is about overcoming failures. List: (1) once I got to be professor, my first 23 grant proposals we're rejected. I am very comfortable now. (2) I was mediocre at school (~B) and particularly bad at stats. Much of my lab's work is close to stats now. (3) my first paper went through >100 iterations. I can now finalize one in a day. (4) I never got accepted into a summer school. But Cosmo with @GunnarBlohm and @Schrater_Lab is very popular.