The Internet Is Torturing a Fruit Fly... (opensource brain) – Transcript

00:00So, Google and Janelia just open-sourced a fly's brain. The complete wiring diagram of a real male fruit fly's brain and nerve cord, every neuron and every connection, for anyone to use. And naturally, the internet has done what the internet does. The fly's been beating Minecraft, playing Beat Saber, trading crypto, learning Python, and of course, playing Doom. Now, you might be wondering what the hell am I even talking about, open-sourcing a fly, but it's probably easier to just fly in and show you what I mean.

00:26>> >> Now, I want to start out by showing you what the internet did with this fly, and then I'll explain later. Just know that most of these demos do have an asterisk to them. It's obviously not as simple as downloading a sentient fly brain and just connecting them to things. With that in mind, though, my favorite demo was posted by Evan Sinclair Smith, where he showed the full connectome running inside Minecraft with simulated neural activity driving a fly around the Minecraft world, and even having interactions like player approach,

00:54attack attempts, and searching for food and light. I just think it's absolutely crazy that we have a fly in Minecraft that is essentially is using the mapping of a real fly's brain. My next favorite demo was seeing the fly play Beat Saber, absolutely nailing the track Among Us Trap Remix, which is a sentence I never thought I would say on this channel, but here we are.

01:14And naturally after this, when you get into video games, you also have to test it out on Doom. As Alex says here, each Doom frame stimulates sensory neurons, neural activity is mapped to the game controls, damage triggers stimulus to two dopamine cells as reinforcement, and this demo is actually being streamed on a site at the moment. It's had over 8,000 attempts, and I don't think it's actually come close to beating the game yet, but I'll definitely be keeping an eye on it. This same guy seems to be having a lot of fun creating what I'm going to dub as fly slop. Like he gave

01:41the fly $100 and let it trade Bitcoin, with the dopamine neurons stimulated on profit, and it actually beat a lot of day traders. >> Boom. Roasted. >> He also built a fly language model and connected three of them on a social media platform. Perhaps my favorite summary of all of this is Google, here's a complete map of the fruit fly brain.

01:58Internet, can it play Minecraft? Internet 2 hours later, can it launch a meme coin? People even started feeling bad for this fly, so they built it its own heaven where it just has an open world full of grass and fruit where it can just fly around and eat whatever it wants. Now, naturally, I had to give this a go myself. So, this is a circuit of 80 real cells from that map, and we'll get to what that map actually is in a second, but I've essentially pointed its eyes at this channel page, and every time the cursor lands on the subscribe button, we give it a reward signal which everyone on X is calling

02:26dopamine, and after about 200,000 attempts, the fly now has 100% success rate of subscribing to this channel, but I do hope it only takes you one try. So, all of those examples aside, what the hell is actually going on here? Well, what we're actually talking about is called male CNS version 1, and it's a connectome, a map of which neuron connects to which and how many synapses sit between them. Every dot you're looking at is a real cell body, over 166,000 of them. Blue is the optic lobes, gray is the central brain, gold

02:54is the sending neurons heading down to the body, and between them, around 125 million synaptic connections. Now, the way you make one of these is a little bit crazy. You take one fly and you slice it 8 nanometers thick. This is one real section through the brain, and there are 134,000 of these, and you image everyone with an electron microscope, and if you zoom in 100 times, that dark bar is a single synapse. On one side is a looming detector, so a cell in the eye that fires when something gets bigger fast,

03:22like a hand coming at the fly, and on the other side is a giant fiber, which is the neuron that actually makes the fly jump. That's the escape reflex. After this, Google's team then use something called a flood filling network, which starts at a single pixel and grows outwards to find every other pixel that belongs to the same neuron through every one of those sections.

03:40These are 80 of the real trace cells from that escape circuit, and the gold one is the giant fiber. That one neuron alone has nearly 25,000 synapses landing on it from over 3,000 other cells. The red and orange ones are the looming detectors with nearly 6 and 1/2 thousand synapses between them. All of this work then ends up as a giant table detailing who connects to who, how many synapses there are, and a predicted transmitter, which all together is 125 million rows.

04:05All of that data is now public. There's a web viewer called Neuroglancer, a query interface called Neuroprint with Python and R packages, and the raw image volumes and connectivity tables are all downloadable. But, as I mentioned with the demos earlier, there is a bit of an asterisk here. You obviously don't download a sentient fly brain. The map just tells you which neurons are wired to which. It doesn't tell you what happens when a signal goes down those wires. Think of it like having the wiring diagram for a house, but no idea what the switches actually do. In my subscribing fly example, I turn the

04:32screen into a few numbers based on where the button is. Those numbers then go into the fly's wiring and then sent through it, but because the map doesn't say how strong any connection is, I just made every single one of them the same. Then what comes out the other end just goes into a small translator that I made, and that essentially has 163 dials, and this translator then decides to move left, move right, or click. The training essentially just tries a load of random settings, sees if the cursor gets closer to the button, keeps the

04:58best ones, and then tries this over 200,000 times until it's 100% accurate. So, technically my fly wasn't a fly, and the same goes for every demo that you saw earlier. They all work the same way. Someone decides what the pixels mean, someone guesses how the cells behave, and a trained network on top does the actual playing. I mean, you probably already guessed that it wasn't a sentient fly. If we were actually at the point where you could download a sentient brain off the internet, I'd probably move to a farm and disconnect from the grid entirely. This doesn't mean that the map is useless though.

05:25Nico Christie switched off one type of neuron in the model, watched what happened to the courtship neurons, and got the same result from a real experiment on real flies from 2015. So, proper science is the real reason this data actually exists, not the fly slot that we have on Twitter, and we now actually have two complete fruit fly connectomes, one male, one female, and around 5% of that central brain is sex specific, and that's where courtship and aggression live. So, where the two differ, you can actually study that, and where they're the same, you can finally

05:52see how much wiring varies between two individuals of the same species, which no one has been able to do on this scale before. Janelia and Google are already working on a zebrafish brain, and then a mouse, and apparently the human brain with 86 billion neurons is still out of reach for now. What do you think of all of these crazy experiments, and have you tried any fly swap yourself? Let me know in the comments down below. Why not subscribe, and as always, see you in the next one. >>