I’m glad I took the time to jot these down, some real gold here.
Just came across an old draft in my writing folder called “3 thoughts”, and it has three bullet points. The first two are incomprehensible combinations of words (one is “Orthogonal Salmon Rushdie”).
The third is “Did I have a third thought?”
No. It can't even learn a mapping between a stimulus and 2 possible actions very well.
I tried to share this with the "Doom Neuron" community (a community around this project) and the mild pushback was immediately removed as "disrespectful".
There's good science here -- it's interesting we can culture these cells on a microelectrode array and get them to do at least some learning. But the hype on this is way out of line with what the research actually was!
Supposedly the network did do some learning in the doom situation -- if you freeze the RL agent, performance gets better. That's nice! But it's a far cry from saying the network is playing Doom. For all we know, the network is making the RL play worse!
The reinforcement learning algorithm output one of 8 stimulations to the cell network, and the network's activity controlled 7 possible game actions (move, turn, shoot). What is the network contributing here?
If the network could barely do anything in a simplified version of Pong, how could it play something as complicated as Doom? The trick is, it didn't! A reinforcement learning algorithm did, using the brain network as a kind of crappy dynamic controller
This was in a very easy version of Pong: no opposition, the ball moved slowly (slower than the paddle it seems). The simple mapping above, doable with a single conditional statement in any programming language, likely would have never let a ball through.
It barely learned to play! The controls would fail to land a single hit in 50-55% of rallies. So by chance, the paddle hits the ball nearly half the time. The brain network dropped this to 48%
You need statistics to tell the game is being played any better than chance.
The "Pong" it was learning to play was simplified, so stimulus told the network if the ball was up or down relative to the paddle. All the network had to do was learn a simple mapping between the stimulus and action (if up, move up, if down, move down)
This work builds on previous work having the same brain cell chip play Pong. That was published in a reputable journal. There's some real science here! But the science is along the lines of getting a small statistical effect of learning, not impressive Pong play
You might have seen the news headlines saying "human brain cells on a chip learned to play Doom".
It sounded really interesting, so I decided to look into it. The reality left me very disappointed
Tomorrow at 12 PM PST, I'll be chatting with @tommyblanchard.bsky.social on Substack livestream about extraordinary beliefs, cognitive anthropology, and the evolution of childhood. Tune in for a great convo :)
open.substack.com/live-stream/...
what the fuck is Buster's dad?
They say money can't buy happiness, but I'm not sure that's true since with enough cash you can hire a personal assistant whose full-time job is unsubscribing you from marketing email lists
There are only two types of days in the winter:
1. Days where the kids are home from daycare because it's a snow day
2. Days where the kids are home from daycare because they are sick
New post about animal consciousness and the evolution of minds (and a review of Peter Godfrey-Smith's book, Metazoa):
A 6-year-old saw me playing a silly game with my kid today and told me I'm like Bandit from Bluey
I've never had such high praise. I'll be riding this high for the rest of the week
The sci fi classic about having children under 5 that bring home a constant stream of daycare germs
Picking up my two-year old from daycare is great. When he sees me he runs to the door with his arms excitedly stretched out for a hug. I embrace him, then all of his friends run over, their arms outstretched as well, saying “Hug, hug”, like a dozen tiny affection-thirsty zombies
What to make of the neural activity observed before the moment people report their intention to act?
Nothing much, argues @tommyblanchard.bsky.social—naturally the intention to act precedes the action. This does not challenge ‘free will’, quite the opposite:
buff.ly/7x6N5fO
Normies are just weirdos you haven't gotten to know yet
This is the front cover of the new edition of @thelancet.com (thanks to @profstevegriffin.bsky.social for sharing) #HealthPolicy #Science 🧪🧵
New article about swarm intelligence and taking the neuron's point of view:
"Just as ant colonies regulate foraging without a leader and slime molds solve mazes without a nervous system, the human brain builds and adapts itself without a central control."
The Lego instructions for assembling “trash can beside trash bag”:
Heaven hath no joy like a 2-year-old with bubbles
