Claude Opus 5.5 Rebuilt Asteroids in Godot — Transcript

Full transcript

1. 0:02For one day, my laptop belonged to AI agents.
2. 0:06I gave them one job: rebuild a classic.
3. 0:09Plus my whole workflow.
4. 0:10They asked me a few questions…
5. 0:12and then I got out of their way.
6. 0:15Twenty-three hours later:
7. 0:17this.
8. 0:18The 1979 arcade classic, rebuilt in 3D.
9. 0:23Here's what happened on my laptop in those
10. 0:25twenty-three hours.
11. 0:27Three agents built it, stage by stage.
12. 0:30One builds. One tests.
13. 0:32One judges. They never talk to each other.
14. 0:35They built a ship that stayed dead for six
15. 0:38minutes. And forty-one flying saucers that
16. 0:40never made it across.
17. 0:43Once, they figured out what laptop they were
18. 0:45actually on. The answer broke their speed
19. 0:47test.
20. 0:49Two of five stages never passed.
21. 0:51And the bill?
22. 0:52We'll get there.
23. 0:54On this channel, we build games with AI.
24. 0:57We push it to its limits.
25. 0:58And we learn by playing.
26. 1:01If you like seeing the real process,
27. 1:03subscribe.
28. 1:05This was a test of a pipeline I've been
29. 1:07building. An idea goes in.
30. 1:09A playable game comes out.
31. 1:12My idea: a little game in space.
32. 1:14You fly a ship, smash asteroids and enemy
33. 1:17ships, and try to survive.
34. 1:19A bit like a game I remember from childhood.
35. 1:22The orchestrator asked me eight questions.
36. 1:26One answer went beyond the options.
37. 1:28In 1979, rocks split into two equal halves.
38. 1:31I wanted something more spectacular:
39. 1:35rocks breaking into irregular pieces,
40. 1:37and dust.
41. 1:39The orchestrator had GPT-6 Astra draw the
42. 1:41first concepts.
43. 1:43Not bad. But I kept iterating on my own,
44. 1:46also with GPT, and sent a more polished
45. 1:49picture instead.
46. 1:50It became the target.
47. 1:52Then the spec was frozen.
48. 1:54Five stages, measurable rules for each.
49. 1:57From then on, the laptop was theirs.
50. 1:59Fully autonomous.
51. 2:01Inside, the orchestrator runs the show.
52. 2:04It plans, hands out the work,
53. 2:05and checks it.
54. 2:07It never writes code.
55. 2:09The builder writes the game.
56. 2:10When it says done, the code is frozen.
57. 2:14QA plays the frozen build with its own
58. 2:16scripts, and measures everything.
59. 2:18The critic reads only the evidence,
60. 2:20and scores it from zero to a hundred.
61. 2:23Fail, and the builder gets another try.
62. 2:26Three tries per stage, max.
63. 2:27And nobody talks to anybody,
64. 2:30except through the orchestrator.
65. 2:32And here's the trick.
66. 2:34They all keep a diary.
67. 2:35Every decision with a timestamp.
68. 2:37Every review.
69. 2:38A screenshot of every attempt.
70. 2:41That's how I know what happened while I was
71. 2:43away.
72. 2:44Stage one: flight.
73. 2:46Turning, thrust, drifting,
74. 2:49wrapping around the edges of the screen.
75. 2:52The builder shipped it with thirty-five
76. 2:54passing tests.
77. 2:55QA found what none of them checked.
78. 2:58Every time the ship wraps around the edge,
79. 3:01it freezes for two frames.
80. 3:0325 milliseconds.
81. 3:04You'd barely see it.
82. 3:06But you'd feel it.
83. 3:08Critic: pass.
84. 3:0982.
85. 3:11The hitch goes on the list for stage two.
86. 3:14Stage two: the rocks.
87. 3:16The builder generates every one in Blender,
88. 3:18from code. The first ones looked like
89. 3:21potatoes.
90. 3:23Now the break.
91. 3:24Every rock has a mass.
92. 3:25Hit it, and it splits into two to five pieces
93. 3:28of different sizes.
94. 3:30Together they're always less than the whole.
95. 3:33The rest turns to dust.
96. 3:35But no piece can be too small.
97. 3:37Otherwise one shot could turn a big rock
98. 3:40straight into dust.
99. 3:42Let me show you how to prove a test really
100. 3:44works.
101. 3:45The builder switched the 1979 rule back on.
102. 3:49Equal pieces.
103. 3:51203 failures.
104. 3:54Switch it back: zero.
105. 3:55If a test can't fail, it proves nothing.
106. 3:59QA did something the builder didn't.
107. 4:01It left the game running.
108. 4:03After a few waves, new rocks appeared right on
109. 4:06top of the ship.
110. 4:08Critic: fail.
111. 4:0963.
112. 4:11New rocks only where there's room.
113. 4:13But to stop broken pieces overlapping,
114. 4:16the builder gave them a little push.
115. 4:18And about one in ten flew back at the shooter.
116. 4:21Fail again. 79.
117. 4:24Last attempt: push removed.
118. 4:26Zero pieces fly back.
119. 4:28Then QA found the real problem.
120. 4:30The field now stayed full of big rocks,
121. 4:32and the ship needed a clear spot to respawn.
122. 4:35In one test game, it stayed dead for six
123. 4:38minutes. The game couldn't even end.
124. 4:41Three attempts used.
125. 4:43Stage two gets the stamp:
126. 4:45needs manual review.
127. 4:47The bug had to wait for stage three.
128. 4:50New rule: wait for a clear spot,
129. 4:53but not forever.
130. 4:54After a few seconds, the ship comes back
131. 4:56anyway, with a pulse that pushes the rocks
132. 4:58away. Longest time dead before:
133. 5:02361 seconds. Now: 4.
134. 5:05Stage three also brought the saucers,
135. 5:08scoring, and power-ups.
136. 5:11And the agents keep testing their own work.
137. 5:14After a change, they launch the game
138. 5:16themselves.
139. 5:17Fly it on autopilot or with scripted keys,
140. 5:20and grab screenshots and recordings.
141. 5:23Meanwhile, the critic watched forty-one
142. 5:25saucers fly in.
143. 5:27Zero made it across.
144. 5:28Eight died within a twentieth of a second of
145. 5:30appearing.
146. 5:32Attempt two: saucers only enter on a clear
147. 5:35path, and plan their route.
148. 5:37Eleven of thirteen made it across.
149. 5:40Later the builder tried making them faster.
150. 5:43More died. What worked was caution.
151. 5:46Critic: pass.
152. 5:47On to stage 4.
153. 5:49Stage four: effects and sound.
154. 5:51Fifteen sounds, all generated in Python.
155. 5:54No sample library.
156. 5:56At least for now.
157. 5:57We'll get back to them after a human review.
158. 5:59I think you can already guess why.
159. 6:02Right after a hit, the pieces first jumped
160. 6:05almost 4 metres in one frame.
161. 6:07Then the glow.
162. 6:08The first looked like light bulbs.
163. 6:11The next, like paint.
164. 6:13And finally, hot stone with glowing cracks,
165. 6:16cooling from yellow to dark red in under a
166. 6:19second.
167. 6:27Every stage also had a speed test:
168. 6:2960 frames a second, no stutters.
169. 6:32The agents kept waiting for the screen to
170. 6:34switch to its high refresh rate,
171. 6:37120 Hz, to measure it properly.
172. 6:40Then the critic checked what laptop this
173. 6:42actually was.
174. 6:43A MacBook Air.
175. 6:45Its screen has no high refresh rate.
176. 6:4760 Hz is the limit, so no build could pass.
177. 6:52Stage four was still a bit slower,
178. 6:54and everyone blamed the new effects.
179. 6:56The builder switched them all off.
180. 6:58Nothing changed.
181. 6:59The real culprit: the saucers,
182. 7:02planning their routes.
183. 7:04In the morning, the verdict.
184. 7:06Effects and sound: pass.
185. 7:08Speed: unproven.
186. 7:10Needs manual review.
187. 7:12Stage five: screens and scores.
188. 7:15Title, pause, game over, name entry,
189. 7:19a top ten.
190. 7:21The critic put it next to the concept.
191. 7:23The screens match closely.
192. 7:25The builder even proved that pause really
193. 7:27stops everything.
194. 7:291.67 seconds of pause.
195. 7:310 pixels change.
196. 7:32And the audio: pure digital silence.
197. 7:37And the tip. If you hand your laptop to AI,
198. 7:40check your sound settings.
199. 7:41My volume was close to max.
200. 7:43And suddenly, without any warning,
201. 7:46QA launched the game.
202. 7:48For 5 minutes, my home was under alien attack.
203. 7:51Can you imagine?
204. 7:53And just before handing the game over to me,
205. 7:56QA found one last bug.
206. 7:58The title screen plays a demo game.
207. 8:00Its lasers kept flying into your new game,
208. 8:02and scoring. Twenty points,
209. 8:05no key pressed.
210. 8:07The restart cleared everything except the
211. 8:09ship's lasers.
212. 8:10Fixed. Critic: pass.
213. 8:12Ninety-two. The best score of the run.
214. 8:18So here it is.
215. 8:20Twenty-three hours on the clock.
216. 8:22Almost seventeen hours of actual work,
217. 8:25spread over a day and a night.
218. 8:28Five stages. Eleven attempts.
219. 8:30Three passed.
220. 8:31Two closed as needs manual review.
221. 8:33Every review kept.
222. 8:35In total, the agents burned through about
223. 8:38eight hundred and twenty million tokens.
224. 8:40At API list prices, that would be around two
225. 8:43hundred and ninety-eight dollars.
226. 8:45Then I got my laptop back,
227. 8:47and for the first time, I played it myself.
228. 8:50My reaction? Real excitement.
229. 8:53Honestly, I'm impressed and happy.
230. 8:55The game looks nice and clean,
231. 8:58and it's a working remaster of an old classic.
232. 9:02Then my own list of fixes.
233. 9:04Some of the sounds had to go.
234. 9:07And the final bill, with my fixes included?
235. 9:10What I would keep: an independent QA and
236. 9:13critic. They found the bugs that mattered.
237. 9:15And once, the builder caught them missing one.
238. 9:19And if you ever hand your laptop to AI agents:
239. 9:22tell them what laptop it is.
240. 9:24If you want more builds like this,
241. 9:26let me know in the comments,
242. 9:27and subscribe.
243. 9:30See you in the next one.