I Put a Fly’s Brain in Minecraft. It Learned to Farm. — Transcript

Full transcript

1. 0:00In October of 2024,
2. 0:01scientists published a complete wiring diagram of a fruit fly's brain.
3. 0:05Every neuron, every connection, free to download.
4. 0:09So I downloaded it.
5. 0:10And then I took a piece of it, I put it inside of Minecraft,
6. 0:13and I spent four days teaching it to farm wheat.
7. 0:16This video is that experiment,
8. 0:18and I'm going to explain all of it from the beginning.
9. 0:20What a brain actually is made of, how you map one,
10. 0:23which piece I took, and what that piece does inside of a real fly,
11. 0:26how it runs on my computer, how it plays Minecraft,
12. 0:29how you teach it anything at all.
13. 0:30Then 15 lessons from put a seed into the ground to being
14. 0:34dropped in a field with nothing but a bucket of water.
15. 0:36I have been staring at Minecraft in wheat for four days,
16. 0:40so let's finally get into this.
17. 0:46So you start at the bottom. A brain.
18. 0:48Yours or fruit fly's is made of one kind of part, the neuron.
19. 0:52A neuron is a cell shaped like a wire.
20. 0:54At one end, it has branches that collect signals.
21. 0:56At the other end, it has a long cable that sends a signal out.
22. 0:59And it only does one thing.
23. 1:00It listens to the neurons connected to it,
24. 1:02and if enough of them are shouting at once,
25. 1:04it fires, and passes a signal down to its own cable to the next neurons in line.
26. 1:08The point where one neuron hands a signal to the next
27. 1:10is called a synapse, and there are two kinds.
28. 1:13An excitatory synapse says fire more.
29. 1:15An inhibitory synapse says fire less.
30. 1:17Every neuron is constantly adding up its fire
31. 1:19mores and its fire lesses and deciding. That's it. That's the whole machine.
32. 1:24Seeing, remembering, wanting a sandwich.
33. 1:26It's all that, repeated a lot of times. How many times?
34. 1:30Well, you have 86 billion neurons.
35. 1:32A fruit fly has about 140 ,000, which is still a lot,
36. 1:36but it's few enough that you could, in theory, write down every single one.
37. 1:40And that list has a name, a connectome.
38. 1:42The complete wiring diagram of a brain.
39. 1:44Every neuron, every synapse between them. And here's how you make one.
40. 1:48A fly's brain is the size of a poppy seed.
41. 1:50Researchers took one from an adult female fly and cut it into 7 ,050 slices.
42. 1:56Each slice is over 1 ,000 times thinner than a human hair.
43. 1:59And every single one of them went under an electron microscope,
44. 2:02which gave them about 21 million images that look like this, gray static.
45. 2:07Every blob in that static is a cross section of a neuron.
46. 2:10To get the shape of one neuron, you find its blob in the next
47. 2:12slice and the next and follow it through thousands of images.
48. 2:16Then you do that for every neuron in the brain.
49. 2:18And AI did that first pass and got a lot of it wrong.
50. 2:21So a project called FlyWire, led out of Princeton, put the whole thing
51. 2:25online and hundreds of scientists and volunteers spent years fixing it by hand.
52. 2:29The result? 139 ,255 neurons, about 50 million synapses,
53. 2:35researchers from 127 institutions published in Nature.
54. 2:39And what you actually download is this table.
55. 2:42This neuron connects to that neuron with this many synapses.
56. 2:46More synapses means a stronger connection.
57. 2:48They even predicted for each neuron,
58. 2:50whether it's the fire more kind or the fire less kind.
59. 2:53One thing to be clear on though, this is a wiring diagram, not a recording.
60. 2:57It tells you what's connected to what.
61. 2:59It doesn't tell you what the fly was thinking.
62. 3:01It's the circuit board, but it's not the software.
63. 3:03But a circuit board is something a computer can run.
64. 3:06So if you're anything like me, you've probably seen a few videos
65. 3:09recently of people playing with this project.
66. 3:11I was absolutely fascinated, you know,
67. 3:14not with hurting the whole it or torturing it thing,
68. 3:17but more teaching it and seeing, you know,
69. 3:20how an actual mapped brain can like learn and teach.
70. 3:23And that's kind of what fascinated me.
71. 3:24And for some reason, the first thing that I actually thought of was
72. 3:27shoving it in Minecraft and seeing what happens.
73. 3:30But for the first asterisk,
74. 3:32I'm not going to be running the whole brain for this experiment.
75. 3:35Most of it does things Minecraft doesn't need like flying.
76. 3:38I took 1536 neurons and the 42
77. 3:42,921 connections between them, about 1% of the brain.
78. 3:46But every one of those connections is real.
79. 3:48It exists inside of the actual fly and at that actual strength.
80. 3:51I didn't add any and I didn't rewire any.
81. 3:53And it's not a random 1%.
82. 3:55It's a one complete pathway front to back.
83. 3:58It's the fly's sense of smell.
84. 3:59And I picked it for two reasons.
85. 4:01It's the most studied learning circuit in the entire fly.
86. 4:04And it's shaped like a decision.
87. 4:05Something comes in and at the other end of the fly either wants it or it doesn't.
88. 4:09It has four stages.
89. 4:10And I'm going to walk through all four because once you see what each one
90. 4:13does in a real fly, the rest of this video will actually make sense to you.
91. 4:16A fly smells with its antennal.
92. 4:18They're covered in receptor neurons, about 1300 on each one.
93. 4:22And each receptor only reacts to certain chemicals inside of the air.
94. 4:25So a smell to a fly isn't one signal. It's a pattern.
95. 4:29A rotting banana sets off one combination of receptors.
96. 4:31Vinegar sets off a different combination.
97. 4:33The smell is which ones are firing like a chord on a piano.
98. 4:37So my slice has 256 of these.
99. 4:39And it's the kind that the data set labels as responding to any food smells.
100. 4:43So this is going to be our front door.
101. 4:44And this is where anything that the brain is going
102. 4:46to need to know about has to come in first.
103. 4:49Those receptors all send their cables to the same place.
104. 4:52A relay station at the front of the brain called the antennal lobe. And it's tidy.
105. 4:56Every receptor of the same type plugs into the same little ball of wiring.
106. 5:00There are about 50 of these balls.
107. 5:01So you can picture the relay station as a panel of 50
108. 5:04lights and every smell lights up its own combination.
109. 5:07Two kinds of neurons work here.
110. 5:09Local neurons run sideways between the lights and they're
111. 5:12inhibitory, the fire less kind. They're a volume control.
112. 5:15If a smell is overwhelming, they'll just turn everything
113. 5:17down so it doesn't just max out every single light. And they sharpen things.
114. 5:20Strong lights push down their weak neighbors.
115. 5:22Messy signal in, clean signal out.
116. 5:24Then projection neurons are the output cables.
117. 5:27They carry the cleaned up pattern deeper into the
118. 5:29brain and they send it to two places at once.
119. 5:31I'll come back to why there are two.
120. 5:33I've got 128 local neurons and 256 projection neurons.
121. 5:37So the first destination is the part that matters the most.
122. 5:40It's called the mushroom body because in 1850, a French biologist
123. 5:43looked at an insect brain
124. 5:45under a microscope and thought it looked like a mushroom. That's the whole reason.
125. 5:48It had a stalk and a cap.
126. 5:49It is the fly's learning center.
127. 5:51If you damage it, a fly can still smell perfectly well.
128. 5:53It just can't learn anything about what it smells anymore.
129. 5:56The mushroom body is made up of neurons called Kenyon cells,
130. 5:59named after the man who first described
131. 6:01them in 1896. A real fly has about 2 ,000 on each side of its brain.
132. 6:06I've got 512. Here's what a Kenyon cell does.
133. 6:09Each one listens to about seven projection neurons,
134. 6:12picked more or less at random, and it's stubborn.
135. 6:14It only fires if several of its seven are active at the same time.
136. 6:17One is not enough. So each Kenyon cell is a detector for one specific combination.
137. 6:22These three lights together, and every Kenyon
138. 6:24cell is watching for a different combination.
139. 6:26The result is that any given smell only switches on about 5% of the Kenyon
140. 6:30cells, and different smells switches on a different 5%.
141. 6:34Every smell gets its own sparse little pattern, its own barcode.
142. 6:37And that's pretty cool if you think about it.
143. 6:39Two smells that are nearly identical at the receptors,
144. 6:41say ripe banana and slightly overripe banana,
145. 6:44light up almost the same lights on the panel.
146. 6:46Hard to tell apart, but run them through the Kenyon cells,
147. 6:49and they come out as two clearly different barcodes,
148. 6:51which means the fly can attach a memory
149. 6:53to one of them without messing up the other.
150. 6:55So how does a fly actually learn?
151. 6:57The Kenyon cells all connect to a small number of output neurons.
152. 7:00Those output neurons push the fly towards something, or away from it.
153. 7:04Now, say the fly smells a particular smell and then finds sugar.
154. 7:07A third set of neurons fires, and they release dopamine.
155. 7:10Same channel as in your brain.
156. 7:12The dopamine lands on the synapses between the Kenyon cells that are
157. 7:15active right now, that smells barcode, and the output neurons.
158. 7:18And it changes their strength. That's it. That's a memory.
159. 7:21Next time that barcode shows up, those synapses are different.
160. 7:25So the output neurons respond differently, and the fly goes toward the smell.
161. 7:28A fly's memory is a handful of synapses getting a bit stronger or a bit weaker.
162. 7:32And the second destination for those projection neurons?
163. 7:35It's called the lateral horn, and it's the instinct route.
164. 7:38Hardwired reactions a fly is born with, no learning needed.
165. 7:41The mushroom body is what the fly hasn't learned.
166. 7:43The lateral horn is what it already knew.
167. 7:45For my outputs, I took neurons from both routes.
168. 7:48The ones the dataset labels as approach,
169. 7:5096 from the mushroom body's outputs, and 288 from the lateral horn.
170. 7:54So that's my 1536. Smell comes in through the receptors,
171. 7:58gets cleaned up at the relay station,
172. 7:59gets turned into a barcode by the Kenyon cells, and what comes out?
173. 8:03The far end is one answer.
174. 8:05Do I want to go toward this or not?
175. 8:07Which, if you think about it, is all you
176. 8:09need to choose your next move in Minecraft.
177. 8:11I just have to make Minecraft smell like something, to the fly.
178. 8:14And here's how you actually simulate that.
179. 8:16In my version, each neuron holds one number.
180. 8:18How active it is right now.
181. 8:20Every step, each neuron looks at every neuron connected to it,
182. 8:23takes their activity,
183. 8:24multiplies it by the strength of the connection, and adds it up.
184. 8:27Fire more connections, push its number up.
185. 8:29Fire less connections, push it down.
186. 8:3135 ,339 of mine push up, 7 ,582 push down.
187. 8:36That split comes from the real fly.
188. 8:38I run that four times.
189. 8:39Four steps is how long it takes a signal to get from the receptors
190. 8:42at the front to the outputs at the back, and then I read the outputs.
191. 8:46And I should be honest about how simple that is.
192. 8:48A real neuron fires in sharp electrical spikes,
193. 8:50runs dozens of chemicals, and has timing down to the millisecond.
194. 8:54Mine is one number, so the wiring is real, but the neurons are more of a cartoon.
195. 8:58It runs on my graphics card because a graphics card is built to do
196. 9:01thousands of small multiplications at the same time, and that's all this is.
197. 9:0542 ,921 of them per step.
198. 9:07One full pass through the brain takes about six milliseconds.
199. 9:11The fly is not the slow part of this project, Minecraft is.
200. 9:14The brain needs a body, so I built a Minecraft mod,
201. 9:17which is just extra code that loads inside of the game.
202. 9:19Mine's called FlyBridge, and it does two jobs.
203. 9:22Job one is it tells the brain what's happening.
204. 9:24The mod runs a tiny server
205. 9:26inside of Minecraft that only my own computer can reach.
206. 9:29The brain's code asks it a question,
207. 9:31what's going on, and it answers with two things.
208. 9:33A screenshot shrunk to 96 by 54 pixels, which looks like this.
209. 9:38And a list of facts, straight from the game.
210. 9:40How many seeds I have, whether I'm holding a hoe,
211. 9:42for each tile nearby, is it dirt,
212. 9:44is it farmland, is it something growing, how ripe is it, is it wet, how far away?
213. 9:48That's the second asterisk, and it's the biggest one.
214. 9:51The fly does not play from pixels alone.
215. 9:53It gets facts a human player never sees.
216. 9:55With 1536 neurons and only that blurry picture,
217. 10:00this video would be four years long and end pretty badly.
218. 10:03And its second job. The mod controls the character with real inputs.
219. 10:07It presses the actual keys, walk, look, left click,
220. 10:10right click, no teleporting, no spawning items.
221. 10:13It's a normal survival world,
222. 10:15and don't get me wrong, this is controlled by the brain.
223. 10:17It's more of a way for the brain to control the character.
224. 10:21Think of it as, you know, fly a friendly controller for Minecraft.
225. 10:25So how do you get till that tile into a smell receptor?
226. 10:29A fly has no wiring for that, so I added two small pieces that are not fly.
227. 10:34At the front, a translator.
228. 10:35It takes the facts and the picture and converts them into 256 numbers.
229. 10:39How active each of those 256 smell receptors should be.
230. 10:43It's an artificial nose.
231. 10:44It turns a Minecraft situation into a smell pattern.
232. 10:47And at the back, a reader.
233. 10:49It looks at the approach neurons and boils them down to one score.
234. 10:52Both start out at random, and both have to be learned.
235. 10:55And everything in between is the fly.
236. 10:57So here's one decision, start to finish.
237. 10:59This whole loop is the whole video.
238. 11:01One, the game state becomes a list of options.
239. 11:04Each option is a job and a target.
240. 11:06Till this tile, plant that one, go to the chest, wait.
241. 11:09Two, each option goes through their translator
242. 11:11and into the receptors as its own smell.
243. 11:13Three, the signal crosses the fly circuit.
244. 11:16Four steps, out comes one number.
245. 11:18How much do I want this?
246. 11:20Four, do that for every option. Highest number wins. No randomness.
247. 11:24Five, and this is the third asterisk, the brain picks what and where.
248. 11:28It doesn't walk. Once it picks plant on that tile,
249. 11:32a script I wrote handles the walking over, aiming, and clicking.
250. 11:35The fly is the manager, it makes every decision,
251. 11:37and it has never once touched the mouse.
252. 11:39If it picks something impossible, like planting with no seeds,
253. 11:42the game rejects it, nothing happens, and it has to choose again.
254. 11:45Remember that, because it comes up constantly during the learning process.
255. 11:48Right now, every score is garbage.
256. 11:50A fly's smell circuit has no opinions about wheat, and it has to learn them.
257. 11:53Fifteen lessons, one brain carried through all of them. Here's lesson one.
258. 12:00Here's what you're looking at, because it's on screen all video.
259. 12:03Left, Minecraft, live.
260. 12:05Top right is the brain.
261. 12:06Those are my real 1536 neurons,
262. 12:08drawn where they sit in the actual fly, letting up with each decision.
263. 12:13Brighter doesn't mean smarter.
264. 12:14Below that, the fly's view,
265. 12:16what the mod is reporting with an outline around whatever the brain just picked.
266. 12:19Then four lines, one for each stage you just learned.
267. 12:22Receptors, relay, Kenyon cells, and output. And a fly at a desk.
268. 12:26He does nothing.
269. 12:27I made him on day one when I still had the energy for this project.
270. 12:30The label in the corner tells you the phase,
271. 12:32and every lesson goes through the same five.
272. 12:34First is baseline.
273. 12:35Before I teach it anything, I let it try.
274. 12:37Four empty plots, 64 seeds, with no training.
275. 12:41It chooses to wait, and then it chooses to wait again. 12 times.
276. 12:45Its scores are random, and wait happened to come out on top.
277. 12:48Nothing changed, so wait came out on top again,
278. 12:50and I capped it at 12. Second is the teacher demo.
279. 12:53The teacher isn't me, it's a plain boring script that already
280. 12:56knows how to plant because I wrote the steps out.
281. 12:58It plays the lesson perfectly, and the system records pairs.
282. 13:01Here was the situation. Here's what the teacher picked. 16 of them.
283. 13:05This way of teaching has a name. Imitation learning.
284. 13:08You don't reward it, you don't punish it,
285. 13:10you show it someone doing it right and train it to make the same choices.
286. 13:14Third is teacher validation.
287. 13:16The teacher does one more run, and those examples get locked away.
288. 13:19The brain never trains on them.
289. 13:20They're the exam, and I'll show you why in a second.
290. 13:22The fourth stage is learning from examples.
291. 13:25The game sits idle, and this happens on the graphics card,
292. 13:28and here's what is actually happening.
293. 13:30The system is full of adjustable numbers. Thousands of them.
294. 13:33I show the brain one recorded situation and ask, what would you pick?
295. 13:36It picks wrong. Then some maths work out for every single adjustable number.
296. 13:40If I made you slightly bigger,
297. 13:41would the teacher's choice have scored higher or lower?
298. 13:44And then it nudges the number a tiny bit in the helpful direction.
299. 13:47Next example, nudge.
300. 13:49Next example, nudge.
301. 13:5016 examples, 90 times through the pile, about 1400 rounds of nudging. That's it.
302. 13:56That is how basically all modern AI learns,
303. 13:58and all the chatbots that you use included.
304. 14:00Guess, check against the right answer,
305. 14:02nudge every number slightly, and repeat a ridiculous number of times.
306. 14:05Mine just has a fly in the middle.
307. 14:07So which numbers are adjustable? Three groups.
308. 14:10The translator, which is the artificial nose, the reader,
309. 14:13which is on the output, and the fly connections themselves.
310. 14:16Each of the 42 ,921 real synapses has a volume
311. 14:19setting that training can turn down to zero or up to double.
312. 14:23But there are two things training can never do.
313. 14:25It can't add a connection that isn't in the fly, and it can't flip one.
314. 14:29A fire less synapse stays fire less forever.
315. 14:32Whatever this brain learns, it has to learn inside the fly's wiring.
316. 14:35And to be straight with you, this is not how a real fly learns.
317. 14:38A real fly uses dopamine in the mushroom body.
318. 14:40Mine uses calculus from outside.
319. 14:42Real wiring, artificial learning. Now, the exam.
320. 14:45If you study the same 16 examples 90 times,
321. 14:48you can score perfectly by memorizing them, without understanding anything.
322. 14:51So after every time through the pile, I test the brain on the locked away
323. 14:54examples that it's never seen, and I keep
324. 14:56whichever version scored the best on those.
325. 14:58Here, that was around 10 of 90. Fifth is the frozen check.
326. 15:02Frozen means learning is switched off,
327. 15:04the numbers can't change, the teacher is gone,
328. 15:06new layouts it hasn't seen, whatever it does now is what it actually learned.
329. 15:10One, two, three, four.
330. 15:12Four seeds, ten seconds, then twice more. Three out of three.
331. 15:16That's the method for every lesson here.
332. 15:18Try it untrained, watch a teacher, train, freeze, test.
333. 15:221536 fly neurons now know one thing, and it's where seeds go.
334. 15:26And here's lesson two.
335. 15:30The ground is plain dirt, you have to hoe it into farmland first, and then plant.
336. 15:34Here's the untrained attempt, don't blink, that was it.
337. 15:37Slowed down, it picks a plant, on dirt, rejected,
338. 15:40plant, rejected, four times in under a second.
339. 15:43That's how the brain fails, and you'll see it in nearly every lesson.
340. 15:47It does not fail by doing nothing, it does the last thing that worked,
341. 15:50with total confidence, somewhere that it makes no sense to do that.
342. 15:54And you can see why. It scores the options.
343. 15:56Plant wins, because planting is the only thing it's ever been taught.
344. 15:59The game says no, nothing in the world changes, so it scores the same options,
345. 16:03gets the same numbers, and plant wins again.
346. 16:06It carries nothing over between decisions,
347. 16:08every neuron starts from zero each time,
348. 16:10every decision is the first decision of its life.
349. 16:13Teaching it lesson two has a trap, and it's a famous one.
350. 16:16Planting and tilling are stored in the same adjustable
351. 16:17numbers, so if I train on only tilling,
352. 16:20those nudges will happily wreck whatever made it get good at planting.
353. 16:22It's called catastrophic forgetting.
354. 16:24Teach it lesson two, and lesson one falls out of its head. Two fixes.
355. 16:28I mix the old planting examples back into the pile
356. 16:30so it has to keep getting those ones right,
357. 16:33and I keep a copy of the brain from before the lesson
358. 16:35and penalize the new one for drifting away from the old one's answers.
359. 16:38Frozen check, till, plant, till, plant, three out of three.
360. 16:44Lesson three is the first one that actually needs judgment.
361. 16:47For wheat plants, one of them is fully grown.
362. 16:50Harvest that one, and you leave the rest alone.
363. 16:52In Minecraft, wheat has eight growth stages,
364. 16:54and yes, the brain gets that number from the mod.
365. 16:57It is not squinting at pixels to work out if the wheat looks yellow enough.
366. 17:00Untrained, it tries to plant seeds on top of plants that already exist.
367. 17:03Same failure. The last thing that worked.
368. 17:06Trained, look at the map.
369. 17:0743%, 57, 71, 100. It outlines the 100. And there's no safety net here.
370. 17:14The script that does the walking doesn't check the brain's work.
371. 17:16If it had picked the 71, the script would have walked over and destroyed it.
372. 17:19Three out of three passed.
373. 17:24Lesson four joins it all up.
374. 17:25Harvest the ripe wheat, pick up what drops, replant the empty tile.
375. 17:29The untrained attempt is interesting this time.
376. 17:31It harvests one, then it replants it.
377. 17:33No one actually taught it that, and I was
378. 17:35pretty actually surprised that it did that.
379. 17:37Harvesting from lesson three and planting from lesson
380. 17:39one just happened to line up.
381. 17:40But then it tries to plant on wheat that's still standing and gets stuck.
382. 17:44So it half already knew it.
383. 17:46Trained, the full chain.
384. 17:47And I want to point out what this isn't.
385. 17:49It isn't a recording being played back.
386. 17:51Wheat drops land in random places, so one run took 11 decisions and the next took
387. 17:5410. After every single action,
388. 17:56it takes a fresh look at the world and asks, what now?
389. 17:59Three out of three, it can farm a tiny plot, but now it gets harder.
390. 18:03And that brings us to lesson five.
391. 18:07A proper three by three garden bed.
392. 18:0918 decisions to finish it.
393. 18:11Now, completely untrained, it just does it.
394. 18:13Till, plant, till, plant, six out of nine plots,
395. 18:17zero mistakes, before it hits the 12 decision cap.
396. 18:19The skills from the tiny plots carried over to a bigger one.
397. 18:22So how do I teach it the rest without damaging what it already has?
398. 18:25This is where I changed how it learns for the rest of the video.
399. 18:28From here on, everything it learned so far gets locked. Read only.
400. 18:32The fly connections, the translator, the old reader, all of it.
401. 18:35Then, for each new lesson, I add one small
402. 18:37new reader on the output, called a head.
403. 18:40It looks at the exact same 1536 neurons.
404. 18:43It just forms its own opinion about what their activity means.
405. 18:46Training only touches the new head.
406. 18:48And afterwards, the computer checks that every locked
407. 18:50number is identical to before, down to the last digit.
408. 18:53So it's one fly brain growing a team of specialists that all read from it.
409. 18:57It physically cannot forget lesson one because lesson one is now locked.
410. 19:01I know it matters, because the first time I filmed this lesson,
411. 19:03I forgot to switch that system on.
412. 19:05It trained the old way and passed two checks out of three.
413. 19:07So I threw that take away, I went back to the old save
414. 19:09brain from the end of lesson four, and I started again.
415. 19:12That's the rule for this whole project.
416. 19:14A failed take will keep nothing.
417. 19:16Nine of nine, three out of three passed.
418. 19:20Lesson six is where I take away its stuff.
419. 19:22The hoe and seeds are in a chest now.
420. 19:24Untrained, it tries to till the dirt with its bare hands,
421. 19:27which gets rejected, obviously, four times. Dirt means till.
422. 19:30It has always meant till.
423. 19:32The fact that it isn't holding a hoe is right there in its list of facts.
424. 19:35Hoe, zero. But up to now, that number has never once mattered.
425. 19:39It always was holding a hoe, so it never learned to look for it.
426. 19:42So that's what this lesson really teaches,
427. 19:44not how to use a chest, that your inventory is part of the situation.
428. 19:48And to be clear, who does what?
429. 19:49The brain makes one decision, go restock, walking to the chest,
430. 19:52opening it, clicking the items out, that's the script.
431. 19:55Trained, straight to the chest once, then it three out of three.
432. 20:01Lesson seven, water.
433. 20:03In Minecraft, farmland needs water within four blocks.
434. 20:06Without it, the soil dries out and eventually turns back into plain dirt.
435. 20:09So here's an empty hole, and here's a full bucket.
436. 20:12Untrained, it stands next to the hole, holding the bucket,
437. 20:15plants all four seeds in bone dry soil, and then gets stuck.
438. 20:19Placed water is just a job that has never been taught the right answer before.
439. 20:23This needed 24 teacher demos, the most of any lesson,
440. 20:26because placing the water is only one decision out of six in each run.
441. 20:30And this is where the exam really earns its keep.
442. 20:32It went through 90 rounds of training.
443. 20:34The version I kept was round 17. Look at the two lines.
444. 20:37Its score on the examples it was studying keeps going up the whole time.
445. 20:41Its score on the hidden exam peaks early and then gets worse.
446. 20:44That gap is the brain giving up on learning
447. 20:45the idea and memorizing the answers instead.
448. 20:48There's a name for it, overfitting.
449. 20:50If I'd kept the final version,
450. 20:51I'd have kept a worse brain that looked better on paper.
451. 20:54Trained, water first, then plant, and in two of the runs,
452. 20:57a dry tile had already turned back into dirt before it got there.
453. 21:00Nobody demonstrated that exact situation.
454. 21:02It just re -tilled it and carried on.
455. 21:04Because tilling dirt is something it's known since lesson two.
456. 21:07Three out of three passed.
457. 21:11Lesson eight, a seven by seven field with water in the middle.
458. 21:1548 plots already planted.
459. 21:16The job, keep it running.
460. 21:18Harvest what ripens, pick it up, replant it, and store wheat in a chest.
461. 21:22Do the whole field twice, store at least 96 wheat, break zero young plants.
462. 21:26Asterisk, wheat takes forever to grow.
463. 21:28So from here on, crops grow at 20 times normal speed.
464. 21:31Only the crops, the brain,
465. 21:32and the character are moving at normal speed still.
466. 21:35The teacher demo for this is seven and a half minutes and 360 decisions.
467. 21:39I'm playing it fast because I watched at a normal
468. 21:41speed and I would not wish that on you or my worst enemy.
469. 21:44After training, it agreed with the hidden exam 88% of the time.
470. 21:47Best score yet on the biggest task yet. Frozen check.
471. 21:54That was it. One second.
472. 21:56It walked up to a field of ripe wheat and chose
473. 21:58store wheat in the chest with no wheat.
474. 22:00Rejected four times. 88%.
475. 22:03Here's what happened, and it's the most useful
476. 22:05thing I learned in this whole project.
477. 22:07Look at what those 360 teacher decisions are made of.
478. 22:09About 100 harvests, about 100 pickings, and about 100 replants.
479. 22:13And storing, about 15. You only
480. 22:15store when your pockets are full so it's more rare.
481. 22:17That means the hidden exam only had three storage
482. 22:20questions in it and the brain got all three wrong. Zero of three.
483. 22:23But it nearly got everything else right so the average still came out at 88%.
484. 22:27The score buried the one thing it couldn't do and the one
485. 22:30thing it couldn't do happened to be the first decision of the run. Second problem.
486. 22:34I'd been teaching it. Here's the situation.
487. 22:36Here's the one thing the teacher picked. Copy that.
488. 22:39But that says nothing about all the other options.
489. 22:41It never learned that storing with empty
490. 22:43hands isn't just not what the teacher did. It's impossible. So two fixes.
491. 22:48One, every job gets an equal amount of time in training
492. 22:50no matter how rare it was in the demo.
493. 22:52Two, I stop labeling just the right answer.
494. 22:55For every situation, every option gets a label. This one's productive.
495. 22:58This one's impossible.
496. 22:59This one would damage a crop.
497. 23:01It learns what not to do and why.
498. 23:03I also made extra practice material by taking
499. 23:05real examples and changing them slightly.
500. 23:07A different camera angle, a different look, and that's called augmentation.
501. 23:11It's generated, not played.
502. 23:12There was no new teacher demo and I didn't have another 10 minutes of wheat in me.
503. 23:16And to be fair, this chapter is three recordings joined together.
504. 23:19The teaching, that repair, and the retest.
505. 23:22Every other chapter but one is a single continuous take. Frozen check.
506. 23:26Harvest, pickup, replant, and there, the chest with wheat.
507. 23:3010 minutes, 325 decisions, 99 wheat stored,
508. 23:34all 48 plots twice, rejected zero, broken zero.
509. 23:38Remember the empty chest because it's going to come back.
510. 23:42Lesson nine is the other half.
511. 23:44Lesson eight was looking after a field that already existed.
512. 23:47This is building one.
513. 23:48Empty pockets, a chest of supplies, dry ground, and 98 decisions.
514. 23:53It took three takes and the second failure taught me something.
515. 23:56It got its gear, placed the water, and then tilled nine
516. 23:58tiles in a row without planting any of them, and then stalled. Here's why.
517. 24:02The teacher always goes till, plant, till, plant.
518. 24:05So the teacher is never in a situation
519. 24:07with lots of tilled soil and nothing planted,
520. 24:09which means the brain has never seen that situation, not once.
521. 24:13And the moment it wandered into it, it had nothing to go on.
522. 24:15That's the big weakness of imitation learning.
523. 24:17The teacher only ever shows you the right path.
524. 24:20Step off of it, even slightly, and you're somewhere with zero examples.
525. 24:23In AI, this has a name. Distribution shift.
526. 24:26The situations you meet aren't the situations that you study. The fix?
527. 24:30I generate worksheets. Take real recorded moments and change the facts.
528. 24:33What if half the field were tilled and none were planted?
529. 24:36What if you only had four seeds left? Not gameplay.
530. 24:39Practice questions for places the teacher never goes.
531. 24:4248 of 48. 98 decisions, zero rejected.
532. 24:45Exactly the same count as the teacher.
533. 24:47Not one wasted move.
534. 24:49So it can farm, but I've been handing it everything.
535. 24:51Tools, seeds, a chest, a hole I dug for it.
536. 24:54For the finale, it starts with nothing,
537. 24:56so it has to learn where those things actually come from.
538. 24:58So here's the shopping list.
539. 25:00Four logs make sixteen planks, four planks for a crafting table,
540. 25:03two for sticks, two for planks, and two sticks for a hoe, eight for a chest.
541. 25:08That's sixteen exactly. Nothing spare.
542. 25:10Plus twenty -four seeds, which come from punching grass. Five lessons.
543. 25:17Lesson ten. Wood.
544. 25:18Untrained, it breaks a log. Good.
545. 25:20The log drops on the ground and it never even picks it up.
546. 25:23It chooses mine that block again on the spot where the log was, which is now air.
547. 25:28Three more times. It is punching the memory of a tree.
548. 25:31Breaking something and owning it are two separate decisions. It had only made one.
549. 25:35This head learned from one demo.
550. 25:37Eight examples, two seconds of training, mine,
551. 25:39collect, mine, collect, four logs, two out of two.
552. 25:44This brings us to lesson eleven. Seeds.
553. 25:46And this is real Minecraft randomness, because when you break grass,
554. 25:49there's only a small chance it drops a seed. Most drop nothing.
555. 25:52The teacher needed a hundred and thirty -nine grass
556. 25:54plants to get twenty -four seeds.
557. 25:56Frozen check.
558. 25:57It starts well, it clears the grass around it, thirteen seeds, and then it stops.
559. 26:01There are two hundred and eighty grass plants still
560. 26:03standing, it's just not going to them.
561. 26:05It keeps choosing the bare ground right next to it where the grass used to be.
562. 26:09Four hundred and three rejected choices in a row. Here's the cause.
563. 26:13In the teacher's demo, the closest option was almost always real grass.
564. 26:16So this is close, and this is grass looked like the exact same thing.
565. 26:20Every single time.
566. 26:21And when two clues always agree, the brain learns whichever is easier.
567. 26:24It learned close. It took a shortcut, and the shortcut worked
568. 26:28right up until everything close was gone,
569. 26:30and it was standing in a bald patch that the teacher had never stood in itself.
570. 26:33The fix. Worksheets with nearby grass deleted, and a new penalty in training
571. 26:37for any time bare ground scores higher than real grass.
572. 26:40I refilmed just the retest, so this chapter has two recordings,
573. 26:44and I put the failed run back in there so you could see it.
574. 26:46Frozen check, it clears a patch, and it leaves.
575. 26:49Two hundred and twenty -five grass, twenty -four seeds, with zero rejected.
576. 26:53It passed.
577. 26:56Lesson twelve. Crafting.
578. 26:58I hand it four logs.
579. 26:59Untrained, it crafts planks.
580. 27:01All four logs into sixteen planks. Correct.
581. 27:03Then it picks craft planks again, with no logs. Four times.
582. 27:07It found something that worked, and it had no concept of done.
583. 27:11After training, it gave us planks, sticks, the table, placed it,
584. 27:15ten planks left over, which is exactly what the hoe and chest will need.
585. 27:18The brain picks which recipe to make, and where to put the table.
586. 27:21Dragging items into the right slots is the script. It's a fly.
587. 27:24I'm not teaching it to use a mouse.
588. 27:28Lesson thirteen. The hoe and the chest. Untrained, it waits.
589. 27:32We've come full circle.
590. 27:33Trained, open the table, hoe, chest, close, place the chest. Five decisions.
591. 27:39One detail though. On the hidden exam, this one only scored eighty percent,
592. 27:43and then it passed both live runs perfectly.
593. 27:45Back in lesson eight, eighty -eight percent died in one second.
594. 27:48So that's what these exam scores really are.
595. 27:51They're a smoke alarm, but they're not a verdict.
596. 27:52They can warn you, they can't tell you that it actually works.
597. 27:56The only test that counts is the frozen one, in the side of the game.
598. 28:02Lesson fourteen. Until now, I've always dug the waterhole for it. Not anymore.
599. 28:06Here's a patch of ground, part soil, part stone, pick a spot, dig,
600. 28:09fill it, and build twenty -four plots that are all in reach of the water.
601. 28:13Untrained, in one point six seconds it chose place
602. 28:16water twelve times, which, you know, there is no hole.
603. 28:19The game's message, word for word,
604. 28:21fill requires a supported contained hole and a filled bucket.
605. 28:24It had half of those things. Trained, dig, water, then twenty -four plots,
606. 28:28working outward, never past the water's reach, twice. That's everything.
607. 28:32Wood, seeds, table, tools, site, farm, and upkeep.
608. 28:36Fourteen lessons, each passed on their own,
609. 28:39and each passed in its own tidy arena with a reset before every attempt.
610. 28:42Now it has to do all of it, in a row, in one world, with no resets.
611. 28:49A courtyard, two trees, a few hundred
612. 28:52grass plants, a cottage, which is decoration,
613. 28:55the fly gets one item, a bucket of water,
614. 28:57because I could not be asked to teach it smelting, I was already suffering.
615. 29:01From that, four logs, twenty -four seeds, table,
616. 29:04hoe, chest, dig a hole, fill it, twenty -four plots.
617. 29:08Wait for the wheat to grow, harvest and replant every plot,
618. 29:11and get twenty -four wheat into the chest.
619. 29:13Learning is switched off for the entire run.
620. 29:15Last asterisk, and it's an important one.
621. 29:18The brain cannot choose the order of the stages. I do.
622. 29:21A few lines of code say you're on wood until you have four logs.
623. 29:24Now you're on seeds. Each stage hands over to the head that learned it.
624. 29:28Inside each stage, every single decision is the fly's,
625. 29:31but the plan to build a farm is mine.
626. 29:33Fly's don't have five -year plans.
627. 29:35It took five attempts, and every failure broke somewhere new. Attempt one.
628. 29:39Wood was fine, seeds were fine, and then it
629. 29:41tries to place a crafting table it hasn't made yet. Twelve times.
630. 29:45In the crafting lesson, the spot for the table was always a few steps away.
631. 29:48Here in the courtyard, however, it was much further.
632. 29:50Further than any distance this brain had ever been shown.
633. 29:53And when you hand a system like this a number outside of everything
634. 29:55it's trained on, its output isn't slightly off. It's meaningless.
635. 29:59It knew crafting, but it didn't know crafting from far away. Attempt two.
636. 30:03Dead in thirty -two seconds.
637. 30:04It broke the fourth log, grabbed for it while it was still
638. 30:06falling through the air, missed, and the controls locked up.
639. 30:09That was my script.
640. 30:11I made it wait for items to land. Attempt number three.
641. 30:14It crafts a table, places it, then crafts a second table, which costs four planks,
642. 30:18sixteen planks, nothing spare, and now it can't afford the chest.
643. 30:22It had never learned that a table you've already got is a table you don't need.
644. 30:25It had to pass 256 simulated crafting
645. 30:27runs before I let it back inside the game again. Attempt four. This one hurt.
646. 30:32Wood, seeds, one table, hoe, chest, digs the hole,
647. 30:36fills it, twenty -four perfect plots, eight minutes,
648. 30:39zero mistakes, everything that had ever gone wrong was fixed,
649. 30:43and then it walks over to the chest and tries to store wheat.
650. 30:46It planted that wheat ten seconds ago, rejected, twelve times, and it was over.
651. 30:50It's the empty chest from lesson eight, back from the dead. And it makes sense.
652. 30:54In lesson eight, the field it learned on was
653. 30:56full of ripe wheat from the first second.
654. 30:57It had never in its life seen a field where
655. 30:59everything was young and there was nothing to do.
656. 31:01The right answer was the one job it basically had never been shown. To wait.
657. 31:05And I hadn't tested that either. I tested every skill.
658. 31:08I hadn't tested the gaps in between them. And that one's on me.
659. 31:11So worksheets for the whole life on a farm. Just planted. Wait.
660. 31:15Something's ripe. Harvest. Pockets empty. Do not go to the chest.
661. 31:18Sixty -four simulated farm cycles later, I let it back in.
662. 31:22And that brings us to attempt number five. Frozen. Wood.
663. 31:25One, two, three, and the fourth log, the one that killed attempt number two.
664. 31:29It waits for it to land. Seeds. And the game was stingy.
665. 31:33Six seeds from the first eighty -eight grass.
666. 31:35Nine and a half minutes to find twenty -four. I've sped it up.
667. 31:38I had to watch it at normal speed at one in the morning. But look at the map.
668. 31:41It clears a patch and it moves on. Every single time.
669. 31:44Crafting, where attempts one and three died.
670. 31:46Planks, sticks, table.
671. 31:48The long walk across the courtyard and it places it. One table. Just the one. Hoe.
672. 31:53Chest. Sixteen planks in. Zero left. Dig. Water.
673. 31:57Twenty -four plots.
674. 31:59And now it's standing in front of a field of seedlings.
675. 32:01Next to a chest with nothing to do.
676. 32:03Exactly where attempt number four ended.
677. 32:06It chooses wait.
678. 32:08Wait.
679. 32:09Wait.
680. 32:10Wait.
681. 32:11Twenty seconds of a fly brain doing nothing.
682. 32:14And it's the best thing that it did all week. First one's ripe. Harvest. Pickup.
683. 32:19Replant.
684. 32:20Now the chest with wheat. Ten. Twenty. Twenty -four. And then it was done.
685. 32:25Eighteen and a half minutes.
686. 32:27Three hundred and eighty -seven decisions. Rejected. Zero. Damaged. Zero.
687. 32:32And every number in its brain identical to when it started.
688. 32:35It didn't learn anything during the run. It already knew.
689. 32:38It started with a bucket.
690. 32:40So can you teach a fly to play Minecraft?
691. 32:42What's real? The wiring.
692. 32:43One thousand five hundred and thirty -six neurons and forty
693. 32:46-two thousand nine hundred and twenty -one connections from a real
694. 32:49fly's sense of smell never rewired with all three
695. 32:52hundred and eighty -seven of those decisions passing through it.
696. 32:54What isn't the fly was walking and clicking as I've told you.
697. 32:57The facts from the mod, the translator, and the reader on either end.
698. 33:01The order of stages, the way it learns,
699. 33:02and it's one percent of a brain with cartoon neurons.
700. 33:05It took five attempts, and one success doesn't tell you how reliable it is.
701. 33:09I also never tested whether a random tangle of wiring the same size would
702. 33:12have done just as well, but that's another video and a different week.
703. 33:15I have got to go to bed.
704. 33:17But here's what I got out of it.
705. 33:19Every time it failed, it failed the same way.
706. 33:21The last thing that worked with full confidence, where it made no sense.
707. 33:24Planting on dirt, punching air, a second table, an empty chest, twice.
708. 33:29And the fix was never making it smarter.
709. 33:31It was showing the situation that it had never seen.
710. 33:33The bald patch, the second table, the field with nothing to do.
711. 33:36It never needed to be cleverer.
712. 33:38It needed to have been there before.
713. 33:40I think that's true of more than fly's.
714. 33:42Anyway, it's a fly, it has a farm, and I'm going to bed.
715. 33:45If you thought this video was good at all, liking or subscribing would
716. 33:49genuinely help me a lot, and I'm praying that this video doesn't flop.
717. 33:53Bye, guys.