Stop Being Tricked — Transcript

Full transcript

1. 0:00Fellas, is it over? Or are we so back?
2. 0:03Now, for the last few years, we've kind
3. 0:05of been in this limbo of we are so over,
4. 0:07no, actually we're so back. And you
5. 0:10know, it's it's a little difficult. The
6. 0:12latest version of it's so over comes
7. 0:14courtesy of this Claude Opus 5 demo.
8. 0:17Look at this. This is a one-shotted
9. 0:19game. The entire thing. One shot. One
10. 0:22prompt. This isn't the only one. There's
11. 0:24plenty of these all over the the
12. 0:26internet right now. Here's another one
13. 0:27with snowboarding. Another one that
14. 0:29looks like some sort of Starfield
15. 0:31variant or clone. RTS's in space. Or
16. 0:35this one that looks just like
17. 0:37Battlefield. Oh, that that that one's
18. 0:39actually just Battlefield. There's just
19. 0:41hundreds more of these and we're seeing
20. 0:43them all the time and the kind of common
21. 0:45thread I'm seeing is that game dev
22. 0:48Guess what? All of you guys that went to
23. 0:50school, you're dumb because AI games
24. 0:52it's going to be amazing. You don't even
25. 0:54know. And I wanted to take a moment and
26. 0:56I wanted to talk about this, okay? The
27. 0:57what I what I'm now referring to as demo
28. 1:00porn. Because I think there's two camps
29. 1:02of people right now. The it's fake,
30. 1:04you're faking it, oh, but it's actually
31. 1:06just super sucky and everything that
32. 1:08you're showing is stupid. Or B, oh my
33. 1:10gosh, it's the greatest thing ever, all
34. 1:12game devs are out of a job and
35. 1:13everybody's going to be able to create
36. 1:14their own triple-A game. Naturally, I
37. 1:16disagree with everybody and I think
38. 1:18everybody's completely wrong. I have the
39. 1:20mandate from heaven to give an opinion
40. 1:22on this matter and I'm going to give a
41. 1:24very strong opinion on this matter. Now,
42. 1:27before we do that, of course, naturally,
43. 1:29I would like to say thank you to the
44. 1:30sponsor.
45. 1:31>> Look at all these engineers sitting at
46. 1:33their neat little desks.
47. 1:35It takes dirty work to keep a code base
48. 1:37clean.
49. 1:38>> [music]
50. 1:38>> Every day sickos are out there
51. 1:40committing unreviewed code. And when
52. 1:42that happens,
53. 1:43linters won't save you. You need someone
54. 1:46like me.
55. 1:47>> Let's go!
56. 1:49>> Feature freeze, scrum bag.
57. 1:51>> Who you calling scrum bag?
58. 1:52>> What's this slop you're trying to push?
59. 1:54Unnecessary comments? Global state?
60. 1:57Nested ternaries?
61. 1:58>> Uh, my bad. I didn't even read the code
62. 2:00yet.
63. 2:01>> You disgust me. Step away from the
64. 2:03keyboard.
65. 2:03>> Just let me explain.
66. 2:05>> Is that a mouse? HE'S MERGING A PROD.
67. 2:07>> YOU HAVE THE RIGHT TO REMAIN SILENT.
68. 2:08Anything you push to GitHub can and will
69. 2:10be used against you. You have the right
70. 2:12to a debugger, but if you cannot afford
71. 2:14one, a public stack trace will be made
72. 2:15available to you.
73. 2:18>> And one more code criminal off the
74. 2:20streets where they belong.
75. 2:22HR.
76. 2:23>> [snorts]
77. 2:25>> Look, I didn't
78. 2:28I know I didn't review any of the code,
79. 2:30but I was going to have Code Rabbit
80. 2:31review it from the start. With one-click
81. 2:33fixes and style enforcement, I don't
82. 2:35need Merge Cop.
83. 2:36I would never merge unreviewed code, but
84. 2:38a first pass with Code Rabbit always
85. 2:40makes things go faster. Actually, you
86. 2:42can try it, too, at coderabbit.ai.
87. 2:45>> Next week on Merge Cop.
88. 2:47>> The diff ler's out there, and I'm going
89. 2:49to be the one to deprecate him.
90. 2:55>> Now, I already know what everybody wants
91. 2:56me to do at this point, right? Which is
92. 2:58they want me to [ __ ] on all these games.
93. 3:00Be like, "Dude, these people are idiots.
94. 3:02What they're creating sucks.
95. 3:03Everything's awful." Uh, the reality is
96. 3:05I I just don't I don't I just don't
97. 3:07think that. See, I remember a time when
98. 3:09the old one-shot examples were this
99. 3:11bouncing ball inside of some sort of
100. 3:13hexagon, sex octagon octagon situation.
101. 3:16But now, this is the new one. It's some
102. 3:19sort of really bizarre clone of Call of
103. 3:22Duty. And the thing is is that I don't
104. 3:24want to actually [ __ ] on that itself cuz
105. 3:26to me, it's very impressive. That is how
106. 3:28far we've come. That is what the new
107. 3:30models can do. All right, hey, that's
108. 3:32super cool. The fact that I can say few
109. 3:34word in English and big result come out.
110. 3:37Much wow, very great. To make my case,
111. 3:39let me first
112. 3:41describe a problem that I've seen with a
113. 3:42lot of AI art, generally speaking. When
114. 3:45I go onto the internets and I go to one
115. 3:47of these many AI generated items and I
116. 3:49say, "Hey, yo, give me me picture of a
117. 3:50crab." It doesn't really matter what the
118. 3:53crab that comes back looks like. An
119. 3:55infinite possibility of crabs look like
120. 3:58a crab. And I go, "Holy cow, look at how
121. 4:01realistic that looks. Oh my gosh, look
122. 4:03at the shading. Look at all this stuff."
123. 4:05But the reality is is I asked for a crab
124. 4:08and statistically speaking it gave me
125. 4:09something out that I cannot really
126. 4:11distinguish whether it is or is not a
127. 4:12crab. It looks like a crab. I'm like,
128. 4:14"Oh my gosh, it's a crab." But there's a
129. 4:15fundamental problem. I didn't have any
130. 4:17opinions on what the crab should look
131. 4:20like. I didn't go, "Okay, I want a crab
132. 4:22with this exact lighting angle. I want
133. 4:24all these little details. I want it to
134. 4:26move in this very specific manner."
135. 4:28Because I want something that I want to
136. 4:31create. Not just what could be. Not a
137. 4:34single point in the set of all possible
138. 4:36crabs. I want a very specific crab. And
139. 4:40anybody who's gone through this exercise
140. 4:42knows that there is a world of
141. 4:44difference between the exact thing I
142. 4:47want and just a point in space. And Matt
143. 4:50Schumer, the person that's really kind
144. 4:52of
145. 4:53leading the way on these one-shot demos
146. 4:55saying how amazing AI gaming is going to
147. 4:56be, writes a blog post and goes through
148. 4:58exactly how he did his one prompt. And I
149. 5:00think that this is the most important
150. 5:02point that you need to take away.
151. 5:04Minimal is better here. We want the
152. 5:06agent to decide the specifics. Again,
153. 5:09you can get anything you dream up, just
154. 5:12nothing you want. And what I mean by
155. 5:14that is look at this little stick man
156. 5:16figure, okay? Okay, that swing was
157. 5:18pretty nice, whatever. I made this game
158. 5:20and to generate this game took almost no
159. 5:22time, nothing. A few minutes and boom,
160. 5:24this all worked out. But the stick guy
161. 5:26looked horrible. Uh his swing was all
162. 5:28weird. He was facing the ball. All sorts
163. 5:31of strange like everything was wrong
164. 5:32about it. And for the next 2 hours I
165. 5:35tried to get it to look better and yes,
166. 5:36it does look better based on what you're
167. 5:38seeing, but still it's nothing what I
168. 5:40can envision inside of my head. And
169. 5:42getting it to go from super janky and
170. 5:44weird-looking into something that just
171. 5:46kind of swings and mostly looks normal,
172. 5:48that cost me $117
173. 5:51of Claude Opus 5, totaling 72 million
174. 5:55tokens. And that was only to get a
175. 5:57result that's like 25% of what I want.
176. 6:00And this is the fundamental with demo
177. 6:01porn. Demo porn just simply says, "Hey,
178. 6:04go look at Starfield and make
179. 6:06something." And it does an incredible
180. 6:08job cloning some basic visual version of
181. 6:11it in which you let it infinitely
182. 6:13iterate for some, you know, hundreds, if
183. 6:15not thousands of dollars. That's
184. 6:17awesome. It's It's It's a cool feat.
185. 6:19It's a cool world we live in. And I bet
186. 6:21you there will be some cool ideas and
187. 6:23games that actually get executed via
188. 6:24this way. But the problem is this, what
189. 6:26makes a good multiplayer shooter? Well,
190. 6:29the levels are a very important part.
191. 6:32Are you going to be able to just a AI
192. 6:34generate a level that's actually fun? If
193. 6:37you've spent any time playing a
194. 6:39multiplayer level, you know it takes
195. 6:41hundreds, if not thousands of hours of
196. 6:44really precise building to create the
197. 6:47perfect multiplayer experience just for
198. 6:51one level. The work is still all there.
199. 6:53The demo is neat. It's cool. It's
200. 6:56something, but it's not actually a game.
201. 6:59And that's why I call it demo porn. It's
202. 7:02like you're getting something, but
203. 7:04you're not getting anything real at all
204. 7:06because all the work is in making the
205. 7:08game fun. And you can see that with all
206. 7:10the responses. All the responses were
207. 7:12like, "Dude, if you know anything about
208. 7:14Unity, you'd be able to do it in like 2
209. 7:15hours without spending a gazillion
210. 7:17dollars." It's like, that's not the
211. 7:18point. The point is it's super cool that
212. 7:19we can build this. But the real point is
213. 7:22it doesn't matter that you can generate
214. 7:23it. What matters is that you can
215. 7:24generate exactly what you want. What
216. 7:27does that look like? I want to see
217. 7:29somebody make Call of Duty for actually
218. 7:31real and generate every single stage of
219. 7:34the way cuz what we'll probably end up
220. 7:36finding is, "Oh, no, actually
221. 7:38it's just 10 times easier, 10 times
222. 7:40faster, and 10 times better hiring an
223. 7:43artist to do a lot of the modeling and
224. 7:45getting things better. Oh, yeah, it
225. 7:46actually turns out you run to like all
226. 7:48sorts of insane performance problems and
227. 7:50you actually just need an expert to step
228. 7:52in. Oh, it turns out actually making a
229. 7:54fun level is super hard and actually
230. 7:56having somebody that understands game
231. 7:58mechanics is really, really important.
232. 8:00Uh, I'm going to need them to step in.
233. 8:02And then next thing you know, you didn't
234. 8:04actually solve the problem, but you were
235. 8:06able to generate something that looked
236. 8:08compelling. I've no I already referenced
237. 8:10this video, but I want to just show it
238. 8:12again because I think he says it so
239. 8:14perfect.
240. 8:15>> Speed isn't the issue. If I told you
241. 8:18David, with this technology, you can
242. 8:20create something that looks exactly like
243. 8:23GTA.
244. 8:25And it's going to take 3 years, not 30
245. 8:27seconds. You'd be like, "I'll spend 3
246. 8:28years on it. It's worth it." But and
247. 8:30that exists. You can in 3 years,
248. 8:33technology exists prior to AI to clone
249. 8:36GTA. But it won't be GTA. It'll be a
250. 8:39clone of GTA. Clone Stone Cell. All hits
251. 8:43are by their very nature unexpected.
252. 8:45>> Absolutely beautifully said and of
253. 8:47course very based. Absolutely love that
254. 8:50video, love that interview. I just
255. 8:51wanted to talk about this because I
256. 8:52think people who don't build the thing,
257. 8:55they see this and go, "Wow, it's so
258. 8:58impossibly impressive. I could never do
259. 9:00that." Reality is yes, of course you
260. 9:02could do that. You just would have to
261. 9:03learn some things and then boom, you
262. 9:04would be able to do that. But that's not
263. 9:06the point. Like the point is really that
264. 9:08the distance between demo to a finished
265. 9:11product is a gulf that is gigantic. I
266. 9:14don't think AI has solved the abandoned
267. 9:17project problem. In fact, I think
268. 9:19there's at least a magnitude more of
269. 9:21abandoned projects because even when you
270. 9:23fully generate something, the tedium,
271. 9:26the the actual like constantly going
272. 9:28back and forth continues to be the same
273. 9:30thing. Yes, you're moving faster, but
274. 9:32you just have a whole new set of
275. 9:34challenges. And again, the difference
276. 9:35between demo and reality and something
277. 9:38that is good that people actually want
278. 9:40is a golf that is gigantic. And good,
279. 9:43amazing decisions don't come from, "Hey
280. 9:45model, you decide." No, it comes from
281. 9:47you sitting down and thinking for days
282. 9:50or weeks about one small little feature.
283. 9:53There's a tower defense game that I'm
284. 9:55building and one of these small features
285. 9:56is that you get these upgrades and when
286. 9:57an upgrade happens, bullet time starts.
287. 10:00And coming up with that idea of bullet
288. 10:02time and exactly how it should look and
289. 10:04feel, it was like four or five hours of
290. 10:06experimenting just for one tiny little
291. 10:09feature. And games are just an
292. 10:12amalgamation of a million tiny little
293. 10:15features. Anyways, I just wanted to say
294. 10:16all these things cuz I really want to
295. 10:18encourage you guys that A, it's not so
296. 10:20over and B, don't be tricked by demo
297. 10:23porn. Stop being tricked. It's not the
298. 10:25end of the world. This isn't the most
299. 10:27amazing thing. Go look at all the
300. 10:28videos. Immediately you'll start seeing
301. 10:30that there's like 1 billion problems
302. 10:32within all the demos, but that's like
303. 10:33again, besides the point. Just don't
304. 10:35forget, something that is amazing that
305. 10:38actually captures people's attention
306. 10:40takes time and dedication, no matter how
307. 10:43easy it is to generate exactly what you
308. 10:45want. Because determining what is
309. 10:47actually the best is still really hard
310. 10:50in a world where everything is free. The
311. 10:52name
312. 10:54is the Prime Agent.