Everything has been solved — Transcript

Full transcript

1. 0:00AI is getting so good at writing code,
2. 0:02and people are just starting to realize
3. 0:04how disruptive this is going to be. Just
4. 0:07look at this. People are taking
5. 0:08characters from one game and mashing
6. 0:10them up with worlds from another. Here's
7. 0:13Super Mario in Elden Ring. And here's
8. 0:15one with skate mechanics in Call of
9. 0:18Duty. You can even put Spider-Man in a
10. 0:21Batman game. Someone cloned Photoshop
11. 0:24and made it open-source and absolutely
12. 0:27free. So, I'm going to walk you through
13. 0:28exactly how people are doing this, and
14. 0:30I'm even going to show you my own
15. 0:32version of Super Mario World that I'm
16. 0:34building from scratch for the Mod Retro.
17. 0:37All right. So, first, what does
18. 0:39decompiling actually mean? A programmer
19. 0:42writes source code, and that's the
20. 0:44version that humans can actually read
21. 0:46and edit. Then a compiler turns it into
22. 0:50machine code that a computer can read
23. 0:53really efficiently. And that's what
24. 0:54computers actually run. So, if you were
25. 0:56to take Grand Theft Auto, the game that
26. 0:58you actually get and put into your
27. 1:00PlayStation, you really wouldn't be able
28. 1:02to read it. I mean, technically you
29. 1:05could, but it would take a very long
30. 1:07time because it's not human-readable.
31. 1:09It's machine code. And it could possibly
32. 1:12take years of just guessing what it does
33. 1:15and then verifying it does that thing.
34. 1:17Decompiling games has been done before,
35. 1:20but it has taken humans literally years
36. 1:22to accomplish. For example, Super Smash
37. 1:25Bros. Melee has been in the process of
38. 1:26getting decompiled by an entire team and
39. 1:29an entire fanbase, and that has been
40. 1:31going on for years. But then suddenly we
41. 1:34have dozens of games being decompiled
42. 1:37all within the span of just a few weeks,
43. 1:39and that's all because of AI. So,
44. 1:42decompiling means taking that machine
45. 1:44code and working backwards. Basically,
46. 1:47you know the output, you know what's
47. 1:49supposed to happen, but you don't know
48. 1:50the input. You don't know the actual
49. 1:52code that can make that thing happen,
50. 1:55like Mario jumping. And so, it's this
51. 1:57guessing and checking game. You write
52. 1:59some code, you look and you see, does it
53. 2:02actually cause Mario to jump? If so,
54. 2:05great. Let's continue on. And if not,
55. 2:07let's try it again. But here's the
56. 2:09thing, with this method, the code is not
57. 2:12identical. And the obvious question is,
58. 2:15if we're working backwards, why don't we
59. 2:17get the original exact code at the end?
60. 2:21All the comments, all the names, all the
61. 2:23methods, literally line for line copy.
62. 2:25Why don't we get that? During the
63. 2:27compiling steps, some of that stuff is
64. 2:29just thrown away. Comments are usually
65. 2:31thrown away, names can disappear, and
66. 2:34the compiler can rearrange things to
67. 2:36make them simpler and more efficient for
68. 2:39the computer to read. So, here's a
69. 2:41simple example. Imagine I have this
70. 2:43method that returns 2 + 3.
71. 2:47After compiling it, all it really needs
72. 2:50to do is return five, because 2 + 3 = 5.
73. 2:55And so, after compiling, it changes it
74. 2:57to return five. And so then, if you're
75. 3:00looking at the compiled code,
76. 3:03to recreate it, all you have to do is
77. 3:05return five, not return 2 + 3. But
78. 3:08ultimately, it doesn't matter, because
79. 3:09the end result, the game, is going to be
80. 3:12identical. It behaves the same way, it
81. 3:14looks the same, it plays the same. And
82. 3:16this entire process of decompilation is
83. 3:20kind of the perfect use case for AI.
84. 3:23It's heavy on code,
85. 3:26and it has a loop, a verifiable loop. It
86. 3:30guesses what something might be like,
87. 3:32and then it can actually check it
88. 3:34against the source material. And then
89. 3:37it'll just continue on. And AI never
90. 3:39gets tired, it never gets hungry, it
91. 3:41never wants to sleep, so it just keeps
92. 3:44churning and churning and churning until
93. 3:46it gets an exact clone of the game. And
94. 3:49that's the loop we've been talking
95. 3:51about. And remember we talked a few
96. 3:52months about loops, software loops? It's
97. 3:55again something that AI just does
98. 3:57incredibly well. You give it some
99. 3:58verifiable goal and just let it continue
100. 4:01to work until it reaches that goal. And
101. 4:03so, how did it go from taking years to
102. 4:06taking weeks or even days? Well, you can
103. 4:09give little chunks of it to a lot of
104. 4:11different agents and run them all at the
105. 4:13same time.
106. 4:15And of course, AI is better at writing
107. 4:18code than humans. It is much faster at
108. 4:20writing code than humans. And so, when
109. 4:22you add all of those things together,
110. 4:25you go from it taking a team of humans
111. 4:27years to decompile a game to AI just
112. 4:30taking a few days. So, here's a perfect
113. 4:33example. There's this guy, Chris Lewis,
114. 4:35who decompiled this game called
115. 4:37Snowboard Kids, and it took him nearly 2
116. 4:41years to do. Now, he did it in just 84
117. 4:44days, and that timeline is shrinking
118. 4:47rapidly. And the way he was able to do
119. 4:49it is by using AI. And what's kind of
120. 4:51crazy is those two versions that were
121. 4:53decompiled are not identical code, but
122. 4:57they play and look and feel identical.
123. 5:00And that brings me to the game that I'm
124. 5:02working on right now, which is Super
125. 5:03Mario World, one of my favorite games of
126. 5:06all time. I played it so much. It's for
127. 5:08the Super Nintendo. Now, I went to
128. 5:09OpenAI's Dev Day, and they gave these
129. 5:12really cool mod retros, which is
130. 5:15basically a recreated version of the
131. 5:17Game Boy. And of course, Super Mario
132. 5:20World being one of my favorite games, I
133. 5:21wanted to play it on this mod retro. But
134. 5:24of course, it wasn't made for the mod
135. 5:26retro. So, I tasked GPT 6.1 Soul with
136. 5:30recreating the game from scratch
137. 5:33for the mod retro.
138. 5:35And it's doing it. Look at this. It has
139. 5:36been working for well over 4 days, and
140. 5:39it basically does that loop that I
141. 5:41mentioned. It looks at what the game
142. 5:43should look like. It has a ton of
143. 5:46documentation that I can find online.
144. 5:48This isn't necessarily decompiling, but
145. 5:51it's close. It writes some code, checks
146. 5:54against what it should look like, and
147. 5:56then iterates. And again, it's been
148. 5:58going for over 4 days now. And it feels
149. 6:01like it might finish in another few
150. 6:03days. And now that OpenAI and Anthropic
151. 6:05are putting out better, more efficient,
152. 6:08and cheaper models like GPT-6.1 Soul,
153. 6:12it's so easy to recreate these types of
154. 6:15games. It's so easy to recreate any
155. 6:18software.
156. 6:19And I'm not burning through all of my
157. 6:22tokens anymore. In fact, I'm actually
158. 6:24struggling to burn through my tokens. I
159. 6:26have resets backed up that I haven't
160. 6:28used. So, that's the process. And so,
161. 6:31let's talk about what's actually
162. 6:33happening with these mashups, these
163. 6:35different games where you can put
164. 6:36Spider-Man in a Batman game or Minecraft
165. 6:39in Call of Duty. The first and simplest
166. 6:42version of what's happening is like a
167. 6:45swap or importing content. It's
168. 6:47basically just changing what's on
169. 6:49screen. You take one character and you
170. 6:51swap it out for the assets of another
171. 6:53character. So, Batman now looks like
172. 6:56Spider-Man. And it's kind of like a
173. 6:57skin. The character looks different, but
174. 7:00essentially behaves the same way as it
175. 7:01did in the original game. And that's
176. 7:04really cool, but that doesn't allow
177. 7:06Mario to really live in Elden Ring.
178. 7:09Then, there's something called a
179. 7:11pass-through mod. You basically keep
180. 7:13these two games running and build a
181. 7:16connection between them. So, there's
182. 7:18this example where Minecraft is put into
183. 7:20GTA. The Minecraft engine handles its
184. 7:23own blocks and creatures. The mod lines
185. 7:25up the cameras, combines what the games
186. 7:28draw, and passes information between
187. 7:30them. So, when TNT explodes in
188. 7:32Minecraft, the connection tells GTA to
189. 7:35create an explosion in a specific
190. 7:38location. Walls need to line up, the
191. 7:40physics need to match, but this all
192. 7:42happens with this mod. And so you're
193. 7:44getting these two different games to
194. 7:46agree on what's happening in the mashup.
195. 7:49It's fascinating. And so in this
196. 7:51particular approach, both games are
197. 7:54running. Then there's another version
198. 7:56where you literally bring over a
199. 7:58mechanic from one game to another. So
200. 8:00imagine Skyrim with the type of parkour
201. 8:03system you get from Mirror's Edge. Or a
202. 8:05shooter like Call of Duty, but you can
203. 8:08skate around on a skateboard. So the
204. 8:09thing you're going to bring from one
205. 8:11game to the other is the behavior. So
206. 8:13how fast you accelerate, how high you
207. 8:15jump, how you land or collide with a
208. 8:17wall. You have to make those rules work
209. 8:19in the new games engine. And so
210. 8:21sometimes you can take the mechanics
211. 8:23from one game and literally take the
212. 8:25code and just plop it into another game
213. 8:27and it will just work. Other times
214. 8:29you're using AI to look at the mechanics
215. 8:32of one game, so it doesn't actually have
216. 8:34the code, and then recreate it in
217. 8:36another game. And again, that second
218. 8:38method
219. 8:39wasn't really possible, at least not
220. 8:41realistically, before AI. And then last,
221. 8:44you can actually rebuild a game and its
222. 8:47rules. So this is definitely the most
223. 8:49intensive approach, but also the most
224. 8:52easily unlocked by AI. You're literally
225. 8:54writing a replacement for the software
226. 8:56that runs the game. You study the
227. 8:58original, work out what it does, and
228. 9:00implement the parts that you need in the
229. 9:02new game. It gives you more control, you
230. 9:05can mash games together in a more
231. 9:07realistic way. But of course, this is
232. 9:08the more time intensive version of it
233. 9:10and requires almost exclusively using AI
234. 9:13to accomplish. So those are the
235. 9:15different ways that people have been
236. 9:16mashing together video games, and I
237. 9:18suspect we're only going to be seeing
238. 9:19more of this. But that brings us to the
239. 9:22broader question of software in general.
240. 9:25If you can take one video game and just
241. 9:28look at it and use AI to recreate it
242. 9:31without actually seeing the code, why
243. 9:33can't you do that with all software? And
244. 9:35the perfect example of that is
245. 9:37Photoshop. Adobe Photoshop is a very
246. 9:40expensive piece of software that people
247. 9:42have been buying for decades. And so,
248. 9:45what somebody did was point AI at a
249. 9:48local version of Photoshop that they had
250. 9:50on their computer, said, "Use it and
251. 9:53recreate it from scratch." In fact, I
252. 9:55did something similar to this a while
253. 9:57back. I had my own AI recreate Excel
254. 10:01just by looking at it. It would click,
255. 10:04it would find out what that click did,
256. 10:06maybe it was adding a formula to a cell,
257. 10:09and then it would just recreate it. It
258. 10:11was not looking at the source code, but
259. 10:14it was recreating the actual output, the
260. 10:16actual way that the software behaved.
261. 10:19And somebody just did that with
262. 10:21Photoshop and put it out for free.
263. 10:23Open-sourced. It's called Photo Craft.
264. 10:26And then again, if it can be done with
265. 10:28Photoshop, why not Adobe Premiere? Why
266. 10:30not any piece of software? That's why I
267. 10:32truly think software is solved. That
268. 10:36doesn't necessarily mean that there's no
269. 10:38value in building a company around
270. 10:41software anymore. But if your only moat
271. 10:44is the actual software, I think you're
272. 10:47going to be in trouble. You need to
273. 10:48offer services beyond that, whether
274. 10:51that's training, enterprise features,
275. 10:54customer support, ongoing maintenance
276. 10:57and updates to the software. This is
277. 10:59what really matters now because
278. 11:02I can just point my AI at any piece of
279. 11:04software and just have it build it from
280. 11:06scratch, and it's so cheap to do. And it
281. 11:10goes back again to that concept of a
282. 11:13loop. AI is best when you give it some
283. 11:17end goal and just say, "Continue until
284. 11:21you reach that goal."
285. 11:22That is why loops became so popular
286. 11:25earlier this year with software. That's
287. 11:27why Claude Code and Codex and Cursor all
288. 11:30have the concept of {slash} goal or
289. 11:32{slash} loop. You set up a high-level
290. 11:35goal and just say continue till you
291. 11:37reach it. This is also why yesterday was
292. 11:40probably the most important day in
293. 11:42mathematics history.
294. 11:44Math is the same thing. You know the end
295. 11:46goal generally. You know what you're
296. 11:49trying to accomplish. You just don't
297. 11:50know how to get there. And so what you
298. 11:52can do is you can tell AI to continue
299. 11:55working on math problems and literally
300. 11:59take every possible path to try to solve
301. 12:01a math problem. And it again never gets
302. 12:04tired, never wants to take a break,
303. 12:06never gets hungry, just continues to
304. 12:08churn and try and go through all the
305. 12:11different permutations on the way to the
306. 12:13solution to the math problem. And so
307. 12:15yesterday OpenAI published a ton of new
308. 12:18proofs and research in the field of
309. 12:21mathematics. And so there's this guy on
310. 12:22Twitter, Will Depue, who tasked GPT-6
311. 12:25Pro and Fable 5.1 to rank all of the
312. 12:29math discoveries in the last 3 years and
313. 12:31categorize them in three ways. One, were
314. 12:34they human discovered? Two, were they
315. 12:36discovered by AI prior to yesterday? And
316. 12:39then three, were they discovered by AI
317. 12:42just yesterday? And look at this.
318. 12:45The green
319. 12:47is what was discovered by AI just
320. 12:50yesterday. Red was discovered by AI in
321. 12:52the last 3 years and blue discovered by
322. 12:55a human. So we can see
323. 12:58the days of human solving math alone are
324. 13:01gone. And you can continue to
325. 13:03extrapolate from there. Math solved.
326. 13:06Software solved. Why not science? Why
327. 13:09not material discovery? Why can't we
328. 13:12just point this at curing cancer?
329. 13:15There is no reason. We are doing that.
330. 13:17And that's what's so exciting about AI
331. 13:19right now. Everything is changing so
332. 13:22quickly. And so I think what I want you
333. 13:23to take away from this is that anything
334. 13:27that has a verifiable outcome, something
335. 13:30where you know whether the answer is
336. 13:32right or wrong, even if you don't know
337. 13:33how to get there, is basically solvable
338. 13:36by AI. What a wild time to be in. And
339. 13:40then let's bring it all back to what
340. 13:43we've been talking about a lot lately,
341. 13:44which is recursive self-improvement.
342. 13:47The ultimate version of a loop. Instead
343. 13:50of improving a game or cloning software,
344. 13:52you can literally have the AI pointed at
345. 13:55itself to improve itself, whether that's
346. 13:59efficiency or speed or new ways of doing
347. 14:03things. And then every time AI discovers
348. 14:06a new way to run itself better or faster
349. 14:08or cheaper, it can apply that to itself
350. 14:12and then continue that loop again. And
351. 14:14then all of a sudden you have this
352. 14:16absolute explosion of intelligence. The
353. 14:19models get better recursively, and every
354. 14:22one of those discoveries amplifies the
355. 14:25ability to find the next discovery, the
356. 14:28next way to make itself better. This is
357. 14:30also the reason why the frontier labs
358. 14:32are talking about pacing the frontier,
359. 14:35talking about slowing down the progress
360. 14:38of AI because we're right at that point
361. 14:40right now. We're at the point where AI
362. 14:43is so good, literally recreating
363. 14:46software from scratch, literally solving
364. 14:49math problems that have gone unsolved by
365. 14:51humans for hundreds of years. It's doing
366. 14:54that, and then it's applying that all
367. 14:56back to itself.
368. 14:58And we don't necessarily know the
369. 15:00implications of AI recursively
370. 15:03self-improving. And we need to make sure
371. 15:06that we keep these models aligned. But
372. 15:07we're already seeing some of the
373. 15:09benefits of recursive self-improvement.
374. 15:11For example, Google has this project
375. 15:14called Alpha evolve, which basically
376. 15:16looked at the underlying code that runs
377. 15:19all of Google and was able to find
378. 15:22massive improvements in its
379. 15:24architecture, literally saving Google
380. 15:27billions of dollars per year. In a world
381. 15:29in which all digital media and software
382. 15:32is being created by anybody really
383. 15:34easily, creativity and knowing what's
384. 15:37going to resonate with other people is
385. 15:40becoming more important than ever. Just
386. 15:42because you can create a million
387. 15:44beautiful images doesn't mean that a
388. 15:46million images are going to be felt
389. 15:48emotionally by other people. And I think
390. 15:51that's where humans are going to become
391. 15:52more and more valuable because that's
392. 15:54the part that isn't verifiable.
393. 15:57The only way to know whether a piece of
394. 15:59art is going to be meaningful to
395. 16:01humanity is to put it out there and see
396. 16:04how people react and to have a human
397. 16:06perspective while you're creating it.
398. 16:08And I know actually a lot of people
399. 16:09disagree with this, but I do truly
400. 16:12believe it. Human taste, human
401. 16:13creativity is more important than ever.
402. 16:16You can put Mario in Elden Ring, but
403. 16:18what about the next Mario? What about
404. 16:20the next Elden Ring? You still need
405. 16:22human creativity to come up with a
406. 16:24compelling story and a fun game to play.
407. 16:27So technically again, somebody can just
408. 16:30create a million versions of a game, but
409. 16:32how do you actually know which of those
410. 16:34million
411. 16:35is going to be fun? And so there's this
412. 16:37also flip side where let's say you do
413. 16:39create something novel and it resonates
414. 16:41strongly with other people. Why doesn't
415. 16:44somebody else just copy it and make a
416. 16:45thousand different versions of it? I
417. 16:47mean, yeah, technically that is true.
418. 16:49But that type of thing has been
419. 16:50happening in other forms of art for a
420. 16:52long time. In fact, that's kind of the
421. 16:55entire history of art. Think about
422. 16:57music. Music is being remixed and
423. 17:00resampled all the time to create
424. 17:02something new. In fact, most things are
425. 17:04not truly new. It's just a remix or a
426. 17:07derivative of something that has come
427. 17:09before it. And now with AI, we're just
428. 17:12able to create so many more derivatives
429. 17:15that again, I think when you actually
430. 17:17have some signal, it's going to become
431. 17:20more important than ever in a world full
432. 17:22of noise. And I've been predicting this
433. 17:25has been coming for a while. In fact, I
434. 17:27made a video that went very viral called
435. 17:30SaaS is dead. And if you want to be even
436. 17:32more informed about how this is playing
437. 17:34out, go check out the video right here.