YouTube2Text

AI Just Reached A Tipping Point — Transcript

by TheAIGRID · 5,515 words · 817 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Something strange just happened in the
  2. 0:01AI world. Over the last two weeks, some
  3. 0:04of the most careful, most skeptical
  4. 0:06people in AI all changed their minds at
  5. 0:08the same time. Duesh Patel, who spent
  6. 0:10the last year arguing that AGI is not
  7. 0:13right around the corner, now says we
  8. 0:15might be only a couple of months away
  9. 0:16from AI that's as good as the best
  10. 0:18humans at AI research. The team behind
  11. 0:21AI 2027 just published an update with a
  12. 0:23title that says it all. Looks like the
  13. 0:26intelligence explosion will happen in
  14. 0:272027 after all. And Anthropic quietly
  15. 0:31told the US government in its own IPO
  16. 0:33filing that it has now crossed its own
  17. 0:35safety threshold for AI, speeding up AI
  18. 0:38research. One researcher at Meta put it
  19. 0:40bluntly. I was seriously wrong. So, in
  20. 0:43this video, I'm going to break down what
  21. 0:44Anthropic just admitted, the numbers
  22. 0:46that changed everyone's minds, and why
  23. 0:48so many people now think AI has reached
  24. 0:51a tipping point.
  25. 0:56Which leads me to point number one,
  26. 0:58which is that Anthropic just crossed its
  27. 1:00own line. Anthropic has something called
  28. 1:02a responsible scaling policy. It's
  29. 1:04basically a list of trip wires, lines
  30. 1:06that, if their AI ever crosses them,
  31. 1:08mean the company has to put stronger
  32. 1:10safety measures in place. And one of
  33. 1:12those trip wires is about AI research
  34. 1:14itself. It triggers when AI is speeding
  35. 1:17up Anthropic's own progress by around
  36. 1:19two times. [music] And when that speed
  37. 1:20up is substantially because of
  38. 1:22automation. In other words, the moment
  39. 1:24where AI starts meaningfully building
  40. 1:26the next AI. Now, for most of AI's
  41. 1:29history, that line felt like science
  42. 1:31fiction. But on the 10th of November,
  43. 1:32Anthropic amended its IPO filing with
  44. 1:35the SEC, and Eli Lifeland, one of the
  45. 1:37authors of AI 2027, [music]
  46. 1:39spotted something in it. Anthropic now
  47. 1:41says it has met that threshold. Here's
  48. 1:44the key part of the filing. The company
  49. 1:46says its use of AI systems has already
  50. 1:48been accelerating its research and
  51. 1:49development and that its internal
  52. 1:51estimates indicate a further increase in
  53. 1:53the rate of progress. [music] And
  54. 1:55because the threshold applies where it
  55. 1:56observes or expects a certain level of
  56. 1:58acceleration, Anthropic says it made a
  57. 2:00determination that it has met this
  58. 2:02threshold based on what it expects from
  59. 2:04future models and its overall rate of
  60. 2:06progress. Now, that wording matters.
  61. 2:09Anthropic is not saying it has watched a
  62. 2:11current model cross the line. It's
  63. 2:13saying it expects its future models to,
  64. 2:15so it's treating the line as crossed
  65. 2:17[music] now. But the filing also says
  66. 2:19something that should make everyone pay
  67. 2:21attention. It [music] says Anthropic's
  68. 2:23evidence about the pace of AIdriven
  69. 2:25acceleration comes mostly from lagging
  70. 2:27indicators and that it may not identify
  71. 2:29a significant acceleration in its rate
  72. 2:31of development until after it has
  73. 2:33already happened. Read that again. The
  74. 2:36company building some of the most
  75. 2:38powerful AI in the world is telling
  76. 2:40investors that it might only notice the
  77. 2:42speed up after it's already here and
  78. 2:44people noticed. The account AI not kill
  79. 2:47everyone is in memes posted it in the
  80. 2:49most dramatic way possible. Anthropic
  81. 2:51have crossed their threshold for RSI.
  82. 2:54Recursive self-improvement is here. And
  83. 2:56they added that RSI was long dismissed
  84. 2:58by very serious people as impossible and
  85. 3:01that they were wrong. So what does AI
  86. 3:04building AI actually look like inside
  87. 3:06Anthropic? Anthropic now publishes
  88. 3:08something called the R&D automation
  89. 3:10index. It rates every model development
  90. 3:13task on a scale from no AI at all to AI
  91. 3:16assists to AI collaborates to AI leads
  92. 3:19all the way up to fully automated. And
  93. 3:22as of the latest version, Claude now
  94. 3:24leads 26% of Anthropic's model R&D work.
  95. 3:27Back in February, that number was under
  96. 3:301%. That's from basically nothing to a
  97. 3:33quarter of the work in about 6 months.
  98. 3:35And that lines up with what Anthropic
  99. 3:37has already said publicly. In a report
  100. 3:39called When AI builds itself, written by
  101. 3:41Jack Clark and Marina Favaro, Anthropic
  102. 3:43said that more than 80% of the code
  103. 3:45merged into its codebase was written by
  104. 3:47Claude. And in one AI safety research
  105. 3:50problem, two human researchers working
  106. 3:52for about a week recovered around 23% of
  107. 3:55the performance gap. Claude agents
  108. 3:57working on the same problem recovered
  109. 3:5997%.
  110. 4:00Anthropic also tested how often Claude
  111. 4:03picks a better next research step than
  112. 4:04the human researcher would. Claude Opus
  113. 4:074.5 won 51% of the time. Claude Mythos
  114. 4:10preview won 64% of the time. So this
  115. 4:13isn't just AI writing code faster. It's
  116. 4:16starting to show research judgment. And
  117. 4:18it gets more interesting. Just 2 days
  118. 4:21after that filing, Anthropics founders
  119. 4:23wrote a letter ahead of the IPO. In it,
  120. 4:25they said that building AI safely may
  121. 4:27require Anthropic and ideally the rest
  122. 4:29of the industry to modulate the
  123. 4:31extremely fast pace [music] at which AI
  124. 4:33capabilities are compounding, especially
  125. 4:35when this is driven by AI systems
  126. 4:37building [music] themselves. And they
  127. 4:39added that there may be times when they
  128. 4:41choose to advance more slowly than they
  129. 4:42could, even if that puts them at a
  130. 4:44competitive disadvantage. One person on
  131. 4:46X summed it up perfectly. Seems like
  132. 4:49they're preparing for RSI. Think about
  133. 4:51how unusual that is. a company that is
  134. 4:53about to go public telling future
  135. 4:55shareholders that it might deliberately
  136. 4:57slow down and let competitors catch up.
  137. 5:00You don't write that in an IPO letter
  138. 5:01unless you genuinely think the speed is
  139. 5:03becoming a problem. And if you want the
  140. 5:05biggest AI news delivered straight to
  141. 5:07your inbox, check out the AI Grid
  142. 5:09newsletter. The link is in the
  143. 5:11description. Which leads me to point
  144. 5:12number two, which is the numbers that
  145. 5:15changed everyone's minds. The trigger
  146. 5:17for all of this was a report from Meta,
  147. 5:19the independent group that tests what
  148. 5:21Frontier AI models can actually do. Meta
  149. 5:24spent months working with internal
  150. 5:25anthropic data, and it found that since
  151. 5:27October last year, Anthropic's internal
  152. 5:30capability scores have been growing at
  153. 5:31around 1.9 times the earlier rate. Meta
  154. 5:35says the real number could be anywhere
  155. 5:36from about 1.5 to 2.8 times, and it
  156. 5:40guessed that perhaps around 80% of that
  157. 5:42acceleration is because of AI itself.
  158. 5:44Now, two times faster might not sound
  159. 5:46like the end of the world, but here's
  160. 5:48why researchers freaked out. Ryan
  161. 5:50Greenblat, a well-known AI safety
  162. 5:52researcher who also works at MET, wrote
  163. 5:55that this report is a big update towards
  164. 5:57super intelligence within 1 to2 years.
  165. 5:59And then he wrote something you almost
  166. 6:01never see from someone in his position.
  167. 6:03I was seriously wrong about RSI speed
  168. 6:06up. He said he had previously expected
  169. 6:08only around a 1.15 to 1.3 time speed up
  170. 6:12in 2026. Instead, it looks like around
  171. 6:15two times, and it gets bigger. Greenblat
  172. 6:17points out that not all AI progress
  173. 6:19comes from better software. Some of it
  174. 6:21just comes from using more compute. So,
  175. 6:24if overall progress is going two times
  176. 6:26faster, the software side has to be
  177. 6:28going even faster than that. When he
  178. 6:30plugged the numbers into his own public
  179. 6:31model, he got a very rough estimate that
  180. 6:33AI is now speeding up AI software
  181. 6:36progress by around 4.5 times. And here's
  182. 6:39the kicker. In the AI 2027 scenario,
  183. 6:42that level of acceleration doesn't
  184. 6:44happen until around March 2027. So, by
  185. 6:47this estimate, reality is running about
  186. 6:495 months ahead of the most famous AI
  187. 6:51takeoff scenario ever written. Now, to
  188. 6:54be fair, Greenblat also gave a big
  189. 6:56caveat. Much of the underlying evidence
  190. 6:58in Meter's report was redacted, so a lot
  191. 7:01of the conclusions depend on trusting
  192. 7:02the authors. He says he does trust them
  193. 7:05a lot, but he also says less redaction
  194. 7:08would greatly reduce the need for trust.
  195. 7:10So, we're seeing the conclusion, but not
  196. 7:12all of the proof. And it's not just
  197. 7:14anthropic where this is showing up. Open
  198. 7:16AAI published a post in September about
  199. 7:18how its own researchers use AI. The
  200. 7:21heaviest 10% of AI users among OpenAI's
  201. 7:23researchers went from spending about
  202. 7:25$1,000 a day on internal coding agents
  203. 7:28at the beginning of the summer to over
  204. 7:29$7,000 a day by the end of August.
  205. 7:32That's a seven times jump in about 3
  206. 7:34months. When your researchers are each
  207. 7:36spending $7,000 a day on AI agents, it's
  208. 7:39because those agents are doing a huge
  209. 7:41amount of the work. But the speed up
  210. 7:43isn't the only number that's moving.
  211. 7:45Remember Meta's famous chart, the one
  212. 7:47that measures how long a task an AI
  213. 7:49agent can complete on its own. From 2019
  214. 7:52to 2025, that number doubled roughly
  215. 7:54every 7 months. From 2024 to early 2026,
  216. 7:58it was doubling every 3.3 months. And
  217. 8:01according to a new analysis using the
  218. 8:03estimates from meter's report, since
  219. 8:04Claude Opus 4.5, it's been doubling
  220. 8:07roughly every 1.7 months. The last model
  221. 8:10meter actually measured was an early
  222. 8:11version of Claude Mythos preview at
  223. 8:13around 3.1 hours. But Anthropic's secret
  224. 8:16internal model, which meter calls model
  225. 8:184, is estimated at around 38 hours.
  226. 8:22That's almost a full working week of
  227. 8:24human effort done by an AI on its own.
  228. 8:26And here's the craziest part about that
  229. 8:28chart. When someone asked Benjamin Todd
  230. 8:30for the official link, he said, "Meter
  231. 8:32isn't updating the official graph
  232. 8:34anymore because they can no longer
  233. 8:35measure the time horizons. The models
  234. 8:37have outgrown the test." Then Andrew
  235. 8:40Curran took Anthropic's own automation
  236. 8:42index and fitted a trend line to it. If
  237. 8:44the current curve keeps going, he
  238. 8:46estimates that 95% of Anthropic's model
  239. 8:49R&D would be fully automated by November
  240. 8:512027. That's what Anthropic describes as
  241. 8:54level five, fully autonomous RSI. Now, I
  242. 8:58want to be clear about this one. It's an
  243. 8:59extrapolation and no task has actually
  244. 9:02been rated fully automated yet. So, that
  245. 9:04part of the curve is based on
  246. 9:05assumptions. But Curran makes a point
  247. 9:07that I think is really important. This
  248. 9:09will eventually happen at every lab, not
  249. 9:12just anthropic. Which leads me to point
  250. 9:14number three, which is the great update.
  251. 9:17What makes this moment feel different
  252. 9:18isn't just the numbers, it's who is
  253. 9:21changing their mind. Duesh Patel runs
  254. 9:23one of the most respected podcasts in
  255. 9:25AI. He's interviewed basically every
  256. 9:27major AI leader. And last year, he wrote
  257. 9:29a famous essay arguing that AGI is not
  258. 9:32right around the corner because AI
  259. 9:34models can't learn on the job like
  260. 9:35humans can. On the 13th of November, he
  261. 9:38released a new video essay with a very
  262. 9:40different message. The avalanche that
  263. 9:42hit math is coming for AI development.
  264. 9:45And this is how it opens. If you had
  265. 9:47asked me earlier this year how far away
  266. 9:49we are from super intelligence, I'd
  267. 9:50probably have said at least the better
  268. 9:52part of a decade. But [music] over the
  269. 9:53last three months, I've updated my views
  270. 9:55a ton. I've been debating people both on
  271. 9:58and off the podcast about this question.
  272. 10:00And of course, I've been watching the
  273. 10:01progress of AI on verifiable domains and
  274. 10:04thinking about what that implies for AI
  275. 10:06research itself. And I've come to the
  276. 10:08view that we're only a couple years away
  277. 10:10and possibly only a couple months away
  278. 10:12from AIs that are as [music] good as the
  279. 10:15best humans at AI research. And I've
  280. 10:17also come to believe that automating AI
  281. 10:19research will compress so much progress
  282. 10:22into a single year that soon thereafter,
  283. 10:24so by 2030 and plausibly as early as
  284. 10:27next [music] year, we'll have AIs that
  285. 10:30are at least as good as the best human
  286. 10:32experts in every field. I know this is a
  287. 10:34wild claim and it's one that I myself
  288. 10:36didn't believe even a couple months ago.
  289. 10:38That post got over 1.5 million views and
  290. 10:41Dwarkkesh isn't alone. On the 10th of
  291. 10:43November, the AI Futures Project, the
  292. 10:46team behind AI 2027, published their
  293. 10:48latest timelines update. They say, "We
  294. 10:51are now about 5 months ahead of the AI
  295. 10:532027 scenario." Previously, Daniel
  296. 10:56Kokilo, the most bullish of the group,
  297. 10:58gave a 50% chance of a superhuman AI
  298. 11:01researcher before 2029. Now, all of them
  299. 11:04forecast that a superhuman AI researcher
  300. 11:06will likely arrive at some point in
  301. 11:082027, assuming there's no deliberate
  302. 11:11slowdown. They describe that as roughly
  303. 11:13the point where a company would rather
  304. 11:14fire all of its human researchers than
  305. 11:16lose its AI researchers. And they give a
  306. 11:1935 to 65% chance of AI that can automate
  307. 11:22all white collar work arriving before
  308. 11:242028. And they say the single biggest
  309. 11:27piece of evidence behind the update was
  310. 11:29the meter report. And then there's
  311. 11:31Anthropic's own co-founder Jack Clark.
  312. 11:33In his newsletter, Import AI. He wrote
  313. 11:36that he now thinks it's more likely
  314. 11:37we'll see full RSI become possible in
  315. 11:402027, whereas before he thought it was
  316. 11:42more likely in 2028. He said the big
  317. 11:45thing he was waiting for was creativity
  318. 11:47and that since then we've seen a
  319. 11:49tremendous advancement in AI making
  320. 11:51creative seeming math breakthroughs and
  321. 11:53signs that AI is getting increasingly
  322. 11:55good at AI research too. Jeremy Harris
  323. 11:58put it perfectly. Many of the smartest
  324. 12:00people who thought RSI wasn't near are
  325. 12:02now changing their minds. [music] And if
  326. 12:04you think very powerful AI is still a
  327. 12:06ways off because someone you respect
  328. 12:07told you so, it might be worth checking
  329. 12:09in with them again. They may have
  330. 12:11changed their mind. And this is why
  331. 12:13people keep bringing up AI 2027. In that
  332. 12:16scenario, June 2027 is the month of
  333. 12:18self-improving AI. The fictional lab has
  334. 12:21a country of geniuses in a data center.
  335. 12:23Most of the human researchers can't
  336. 12:25usefully contribute anymore. They go to
  337. 12:27bed every night and wake up to another
  338. 12:29week's worth of progress made mostly by
  339. 12:31the AIS. When that scenario came out, a
  340. 12:33lot of people called it science fiction.
  341. 12:35Now, its authors think the real world is
  342. 12:37running ahead of it. Which leads me to
  343. 12:40point number four, which is why AI
  344. 12:41research could go the way of math. This
  345. 12:44is the heart of Dwarkh's argument, and I
  346. 12:46think it's the most important idea in
  347. 12:48this whole video. Look at what just
  348. 12:50happened to mathematics. It's crazy how
  349. 12:52fast this progress has happened. As
  350. 12:54recently as 2023, these models couldn't
  351. 12:56reliably add four-digit numbers. In
  352. 12:582024, they could barely solve any
  353. 13:01problems in a high school math
  354. 13:02competition. In 2025, they could win a
  355. 13:05gold medal in the International Math
  356. 13:06Olympiad. [music] To earlier this year,
  357. 13:08where they finally started to solve some
  358. 13:10niche open problems to a month ago where
  359. 13:13OpenAI solved a millennium prize problem
  360. 13:16in 88 hours. [music] I'm reminded of a
  361. 13:18line from the novel East. Quote, "And
  362. 13:22there is one sure thing about the fall
  363. 13:24of gods. They do not fall a little. they
  364. 13:26crash and shatter." End quote. That's
  365. 13:29the idea. OpenAI's model reportedly
  366. 13:31solved a Millennium Prize problem, one
  367. 13:33of the hardest open problems in
  368. 13:35mathematics, in 88 hours. And it wasn't
  369. 13:37trained by solving other Millennium
  370. 13:39Prize problems. It was trained on lots
  371. 13:41of smaller, more tractable math
  372. 13:42problems, and it generalized all the way
  373. 13:44up. So, why does that matter for AI
  374. 13:47research? Because the key to all of this
  375. 13:49is something called verifiability. When
  376. 13:52an AI can check whether its answer is
  377. 13:54right, you can give it thousands of
  378. 13:55rounds of practice and it gets better
  379. 13:57and better. That's exactly what happened
  380. 13:59with competitive programming and with
  381. 14:01math. They went from worse than most
  382. 14:03humans to better than the best humans in
  383. 14:05just a few years. And Dwarkh's argument
  384. 14:08is that AI research is surprisingly
  385. 14:10verifiable, too. Did your new training
  386. 14:12trick lower the loss on a test set? Did
  387. 14:15the model get faster? Did it use less
  388. 14:17memory? Those are all things you can
  389. 14:19measure, which means [music] they're all
  390. 14:20things AI can practice. And here's a
  391. 14:23wild example he gives. The Transformer,
  392. 14:25the architecture behind every modern
  393. 14:27chatbot, wasn't really one giant leap of
  394. 14:29genius. Most of the pieces were already
  395. 14:32known. What it took was combining half a
  396. 14:34dozen ideas in just the right way and
  397. 14:36running tons and tons of small
  398. 14:38experiments. And that's exactly the kind
  399. 14:40of persistent combine everything, test
  400. 14:43everything work that AI is already
  401. 14:45extremely good at. Dwaresh even says the
  402. 14:47transformer is the kind of thing that a
  403. 14:49future Claude could have invented if a
  404. 14:51researcher asked it to try really hard
  405. 14:53to get attention to work better. Then
  406. 14:56there's the speed. Dwarsh says today's
  407. 14:58AIS already work at about 10 times human
  408. 15:01speed. And an automated researcher
  409. 15:03doesn't sleep. It works nights and
  410. 15:05weekends. So each one could cram roughly
  411. 15:0850 times more work into every calendar
  412. 15:10week than a human researcher. On top of
  413. 15:12that, every one of them could be as good
  414. 15:14as the very best researchers at the top
  415. 15:16labs and there could be thousands of
  416. 15:18them working in parallel. And a lot of
  417. 15:20people have argued that compute will
  418. 15:21stop this, that even with brilliant AI
  419. 15:24researchers, you can only run so many
  420. 15:26experiments. But there's a thought
  421. 15:28experiment that changed a lot of minds
  422. 15:29on this. Dan Robinson said it hit him
  423. 15:32like a sledgehammer. Imagine a slow
  424. 15:34version of open AI. Let's call it sloth
  425. 15:37AI. It has exactly the same compute as
  426. 15:39the real open AI, but its researchers
  427. 15:41only work 1 hour every week. How much
  428. 15:44slower would it make progress? Probably
  429. 15:46a lot slower, because with just 1 hour
  430. 15:48of thinking a week, [music] they can't
  431. 15:50design and analyze many careful
  432. 15:52experiments. But here's the twist.
  433. 15:54Compared to a company full of automated
  434. 15:56researchers thinking 10 times faster
  435. 15:58around the clock, the real open AI is
  436. 16:01sloth AI. Human thinking speed is a
  437. 16:04biological accident. It would be a
  438. 16:06strange coincidence if it happened to be
  439. 16:07the exact point where thinking faster
  440. 16:09stops helping. And what about data?
  441. 16:11Isn't AI going to run out of training
  442. 16:13data? Dwarvesh's answer is that during
  443. 16:16RSI, the AIs will be the ones making the
  444. 16:18data. And there's already evidence of
  445. 16:20this. Deepseek just published a paper
  446. 16:22about how it post-trained [music] its
  447. 16:24newest model and it said that changes to
  448. 16:26the basic training algorithm haven't
  449. 16:28mattered nearly as much as improvements
  450. 16:30in the scale diversity and verifiability
  451. 16:32of synthesized data and environments.
  452. 16:35And it said it uses large-scale
  453. 16:37automated pipelines to build those
  454. 16:39environments, training the model to
  455. 16:40construct better and harder tasks for
  456. 16:42itself. So the AI is already helping to
  457. 16:45build its own classroom which leads me
  458. 16:48to point number five which is that every
  459. 16:50lab is now saying the loop has started.
  460. 16:53This isn't just an anthropic story.
  461. 16:54Listen to how Google's CEO Sundar Pichai
  462. 16:57described the way Google is building its
  463. 16:59newest Gemini model being
  464. 17:03you know critical contributor as we are
  465. 17:05developing Argon and but to be very
  466. 17:08clear today you know in a continuum of
  467. 17:11we're pretty far from uh you know the
  468. 17:14end state of RSI or something like that
  469. 17:16right so today you have the the right
  470. 17:19way to describe it is you have RSI loops
  471. 17:22along with AI researchers in various
  472. 17:24aspects of model development, right? So,
  473. 17:27you know, I would say you almost have
  474. 17:30agents helping AI researchers uh in in
  475. 17:33some automated way uh work together to
  476. 17:36build these models. Though Google is
  477. 17:38openly saying RSI loops are part of how
  478. 17:40it builds its models and Google
  479. 17:42co-founder Sergey Brin went even further
  480. 17:45in
  481. 17:45>> I'm very excited about the cadence that
  482. 17:47we have nowadays and uh if you look at
  483. 17:51even before Gemini 4 when we had um
  484. 17:54flash 3.6 3.7 3.8 eight. Uh we're
  485. 17:58starting to get onto that RSI treadmill
  486. 18:01and now the you know our researchers and
  487. 18:04the model work hand in hand to make the
  488. 18:07next version of the model. Actually I
  489. 18:09just popped in here kind of randomly
  490. 18:11because Melvin and I were talking about
  491. 18:13the next version of the model and you
  492. 18:15know probably by the time folks are
  493. 18:17viewing this uh internally we're already
  494. 18:19using a couple models ahead.
  495. 18:21>> Internally they're already using a
  496. 18:23couple of models ahead. That means the
  497. 18:25Gemini you use today is not the best
  498. 18:27Gemini that exists and the models inside
  499. 18:29the lab are helping build the ones after
  500. 18:31them. Here's how Google deep mind
  501. 18:33researcher Mustafa Dani described it
  502. 18:36>> like and and and it's somehow like most
  503. 18:38of the people don't realize that this is
  504. 18:39like already uh happening especially
  505. 18:41over the past few months in almost every
  506. 18:44lab the new generation of the models are
  507. 18:46built heavily using the previous
  508. 18:48generation of the models. I think that's
  509. 18:49basically like like the case again
  510. 18:52everywhere. Uh and uh it's not fully
  511. 18:55automatic yet. Uh but the direction is
  512. 18:58like super clear and it's like easy to
  513. 19:00imagine that you know like we're going
  514. 19:02to get to a situation with full
  515. 19:04automation. These models are going to
  516. 19:06improve themselves and keep learning
  517. 19:09from the word
  518. 19:09>> and at OpenAI the focus is already on
  519. 19:12what comes next. Here's OpenAI's head of
  520. 19:15applied research Boris Power. So I would
  521. 19:17say still 80 90% of the research at
  522. 19:20OpenAI is focused at GPT7, GPT8 and
  523. 19:24beyond because we believe that's where
  524. 19:26fundamentally most of the value comes
  525. 19:27from.
  526. 19:28>> And it's not just the big labs. In
  527. 19:29August, Jeff Dean, one of the most
  528. 19:31legendary engineers in Google's history,
  529. 19:34left Google after 27 years to co-found a
  530. 19:37startup called Discovery Loop alongside
  531. 19:39Sanjay Gamat, Oriel Vineyles, and Quac
  532. 19:41Lair. Here's how one of its lead
  533. 19:43investors, Kla Ventures, described it on
  534. 19:45TBPN.
  535. 19:47>> You know, we invested in, we were the uh
  536. 19:49co-leads of Discovery Loop. This is Jeff
  537. 19:51Dean and his gang from Google. The
  538. 19:53legend was 27 years at Google.
  539. 19:56Everything from, you know, TPUs,
  540. 19:58TensorFlow,
  541. 19:59uh, Google Brain. I mean, you you know,
  542. 20:02really every almost all the significant
  543. 20:04achievements at Google, these four folks
  544. 20:06had their fingers in. and they finally
  545. 20:08leave after 27 years to do an RSA RSI
  546. 20:11company looking at how can we improve
  547. 20:14the scientific method. So you now are
  548. 20:16designing experiments getting results in
  549. 20:18real time and then continually
  550. 20:21in parallel designing thousands of
  551. 20:24follow-on experiments. People like that
  552. 20:26leave the biggest AI lab on earth to
  553. 20:28build an automated researcher. It tells
  554. 20:30you where they think the action is.
  555. 20:32Which leads me to point number six,
  556. 20:34which is spark or singularity. So let's
  557. 20:37say AI does become as good as the best
  558. 20:39human AI researchers. What happens next?
  559. 20:42Dwaresh says there are two competing
  560. 20:44forces. On one side, AI researchers keep
  561. 20:47getting faster, smarter, and more
  562. 20:49numerous because each new version helps
  563. 20:51build the next. On the other side, AI
  564. 20:54research gets harder because the easy
  565. 20:56ideas get used up. Companies already
  566. 20:58have to roughly quadruple their R&D
  567. 21:00compute and double their workforce every
  568. 21:02year just to keep the same rate of
  569. 21:04progress. If AI gets smarter faster than
  570. 21:07ideas get harder to find, you get what's
  571. 21:09called a softwareonly singularity.
  572. 21:12Without even adding new compute, the AI
  573. 21:14swarm makes smarter and smarter versions
  574. 21:16of itself until [music] within a matter
  575. 21:18of months you hit a super intelligence
  576. 21:20that runs into the physical limits of
  577. 21:22our current chips. But if ideas get
  578. 21:24harder to find faster than AI improves,
  579. 21:26you get what Dwarvesh calls a spark.
  580. 21:29something like 3 to 6 years of progress
  581. 21:31at today's rate packed into a single
  582. 21:33calendar year before things simmer down
  583. 21:35until more compute is built and he says
  584. 21:37it's basically a coin toss 50% spark 50%
  585. 21:41singularity but here's the part that I
  586. 21:43think people miss think but to be clear
  587. 21:46even in this spark scenario where we
  588. 21:47make something like 4 years of progress
  589. 21:49in the first calendar year after we
  590. 21:51automate AI research things would feel
  591. 21:54insane that's a jump as big as GPT 3.5
  592. 21:57to GPT6 six Astra in just 12 months. To
  593. 22:01remind you of the scale of progress that
  594. 22:03this leap would compress, GBT3.5 was
  595. 22:06able to write a passable limmerick and
  596. 22:08autocomplete a trivial Python function.
  597. 22:10But otherwise, it was mostly a useless
  598. 22:12toy. Whereas Aster can build whole video
  599. 22:14games from scratch and turns out so many
  600. 22:17fields medal worthy mathematical
  601. 22:18contributions that we become totally
  602. 22:20desensitized. [music] In the last 4
  603. 22:22years, we've gone from a preschooler to
  604. 22:25a savant intern. Think about that. A
  605. 22:27jump as big as GPT3.5 to GPT6 Astra in
  606. 22:3112 months. And Darkh says a mind like
  607. 22:34that would probably be vastly superhuman
  608. 22:37in any domain that's even slightly
  609. 22:39verifiable. Better than Nvidia at
  610. 22:41designing the next generation of AI
  611. 22:43chips. Better than the top trading firms
  612. 22:45at making money in the markets. Better
  613. 22:47than the rest of humanity at finding a
  614. 22:49room temperature superconductor if you
  615. 22:51gave it access to a materials lab. But
  616. 22:53what about things that can't be checked
  617. 22:54inside a data center, like running a
  618. 22:57company or winning a court case?
  619. 22:59Dwarvesh thinks the AIS will train on
  620. 23:01enormous virtual worlds. He describes
  621. 23:03simulated economies with millions of
  622. 23:05other AIs where an AI is told to co-ound
  623. 23:07a successful company and endless back
  624. 23:10tests of history. For example, can an AI
  625. 23:13figure out general relativity if it's
  626. 23:15only given information from before the
  627. 23:16year 1700? And then there's how fast an
  628. 23:19AI could learn a real job. Imagine an AI
  629. 23:22shadowing the CEO of a Fortune 500
  630. 23:25company. On its first morning, it reads
  631. 23:27every email, every Slack message, and
  632. 23:29every document the company has ever
  633. 23:31produced. By the end of its first day,
  634. 23:33copies of it have interviewed every
  635. 23:35employee, and every major client, all in
  636. 23:37parallel. That's years of context built
  637. 23:39up in a few days. And even in the slower
  638. 23:42Spark scenario, he thinks AI gets to top
  639. 23:44human performance across basically all
  640. 23:46fields by around 2028 or 2029. And
  641. 23:50here's the thing, some people think we
  642. 23:51get there even without RSI. Benjamin
  643. 23:54Todd wrote a post arguing that even
  644. 23:56without recursive self-improvement, dumb
  645. 23:58trend extrapolation gets to AGI by 2030.
  646. 24:01His logic is simple. AI revenue has
  647. 24:04tripled every year for more than 4
  648. 24:06years, and that growth is actually
  649. 24:08accelerating. According to Epoch AI
  650. 24:10data, the leading AI companies grew
  651. 24:12their revenue three times in 2024, 4.7
  652. 24:15times in 2025, and around six times this
  653. 24:18year. If revenue keeps growing six times
  654. 24:21a year, it reaches $10 trillion by
  655. 24:23around March 2029. That's roughly 10% of
  656. 24:26the entire world economy. And AI task
  657. 24:29length has been growing around eight
  658. 24:31times a year since 2024, which would
  659. 24:33reach multi-year tasks by early 2030. As
  660. 24:36he puts it, every exponential becomes an
  661. 24:39S-curve eventually, [music] but right
  662. 24:41now, revenue is still accelerating. So,
  663. 24:43you have two separate paths that both
  664. 24:45point to roughly the same place. Even if
  665. 24:48RSI turns out to be slower than people
  666. 24:50fear, the plain old trend lines are
  667. 24:52already pointing at something enormous
  668. 24:54before 2030. Which leads me to point
  669. 24:56number seven, which is the fight over
  670. 24:58the brakes. Because if this really is
  671. 25:01the tipping point, then the most
  672. 25:02important question isn't how fast AI can
  673. 25:05go. It's who decides how fast it should
  674. 25:07go. Even Sam Olman is now framing it
  675. 25:10this way. Here's what he said when asked
  676. 25:11about the biggest problem he's thinking
  677. 25:13about. I think that one that I just gave
  678. 25:15the the like this question of how do we
  679. 25:18think about
  680. 25:20the pace of progress as we enter this
  681. 25:21RSI era and how do we avoid these two
  682. 25:25ways that [music] things could go wrong
  683. 25:27either way from a loss of control
  684. 25:29incident or a concentration of power
  685. 25:31incident and what is the right
  686. 25:33governance mechanism? How like how do we
  687. 25:35think about how [music] smart an AI
  688. 25:38system should be allowed to be relative
  689. 25:40to a person? How quickly can that
  690. 25:41improve? how smart an AI system should
  691. 25:44be allowed to be relative to a person,
  692. 25:46how quickly it can improve. These are
  693. 25:48the questions the CEO of OpenAI is
  694. 25:50openly asking. And Microsoft's AI CEO,
  695. 25:53Mustafa Sullean, has been even more
  696. 25:55direct about where the real danger is.
  697. 25:58>> Um, the next phase, which I think
  698. 26:00carries the most risk, is closed RSI,
  699. 26:02closed loop, where you basically manage
  700. 26:04to somehow figure out how to do it
  701. 26:06without the human reviewer process uh in
  702. 26:09this in the in the whole procedure. And
  703. 26:12that um will obviously be faster, but I
  704. 26:14think is also a very dangerous way to
  705. 26:16get to full AGI because we're likely for
  706. 26:19two reasons. It's it's sometimes used to
  707. 26:21be referred to 20 years ago as like an
  708. 26:22intelligence explosion because you don't
  709. 26:25quite know uh how you know how do you
  710. 26:27control keep control of such a system uh
  711. 26:30such a process and then the end result
  712. 26:33will likely be quite alien. It's going
  713. 26:35to be like an alien intelligence.
  714. 26:36>> That's the key distinction right now.
  715. 26:38Humans are still in the loop. They
  716. 26:40review the work and set the direction.
  717. 26:42The scary version is when that human
  718. 26:44reviewer disappears. And notably, in an
  719. 26:47earlier version of its IPO filing,
  720. 26:49Anthropic said it is considering updates
  721. 26:51to its responsible scaling policy to
  722. 26:53keep humans in control of Frontier model
  723. 26:55development. So, even the company
  724. 26:57crossing the threshold is thinking about
  725. 26:59how to keep a human hand on the wheel.
  726. 27:01But not everyone in power sees
  727. 27:03anthropics warnings the same way. The
  728. 27:05White House's AI adviser, David Sax, had
  729. 27:07a very different take on all-in. So
  730. 27:09they're running around saying that this
  731. 27:11is a winner take all market. What does
  732. 27:13that sound like? Monopoly. They're not
  733. 27:14even saying that it's going to be a
  734. 27:16duopoly with anthropic and open AI.
  735. 27:18They're saying they're going to win the
  736. 27:19whole thing. They definitely are saying
  737. 27:22that recursive self-improvement has
  738. 27:24kicked in [music] despite the obvious
  739. 27:26safety problems with that. I mean,
  740. 27:29recursive self-improvement is the most
  741. 27:30dangerous thing anybody can be working
  742. 27:32on in [music]
  743. 27:34SI. If you're worried about rapid
  744. 27:36takeoff, for example, if that's [music]
  745. 27:38your if that's your greatest fear, then
  746. 27:40I mean, they are racing towards that.
  747. 27:42But recursive self-improvement would be
  748. 27:44the key to getting to true next level
  749. 27:48super intelligence, and they're treating
  750. 27:50it like a business advantage.
  751. 27:51>> But from Sax's point of view, Anthropic
  752. 27:53is warning about RSI while racing
  753. 27:55towards it and treating it like a
  754. 27:57business advantage. One person on X
  755. 27:59summed up the irony of it with three
  756. 28:01words. David Saxs says, "Ban RSI." And
  757. 28:04in Congress, some people literally are
  758. 28:06saying that. After New York Assembly
  759. 28:08member Alex Bores posted that every AI
  760. 28:11company is trying to fully automate AI
  761. 28:13research, Congressman Greg Casar replied
  762. 28:15with two words: ban RSI. That post got
  763. 28:19over 800,000 [music]
  764. 28:20views. And it's also worth hearing from
  765. 28:22people who think everyone is getting a
  766. 28:24little carried away. Richard No, an
  767. 28:26independent AI researcher who used to
  768. 28:28work at OpenAI, said he thinks people
  769. 28:30are jumping further than the evidence
  770. 28:32warrants towards expecting a very fast
  771. 28:34RSI over the next couple of years. But
  772. 28:36listen to how he finished that thought.
  773. 28:38I I I don't really see the world looking
  774. 28:40normal over the next decade or so. Um
  775. 28:43and and so the the default sense I I
  776. 28:45have for what's going to happen is that
  777. 28:48uh we're not going to get uh like yeah
  778. 28:53things going vertical, if you will. uh
  779. 28:55we're not going to get like sudden and
  780. 28:57extreme um self-improvement that goes
  781. 29:00far beyond uh that goes very quickly far
  782. 29:03be into the super intelligent regime but
  783. 29:05at the same time things are going to go
  784. 29:07much uh things are going to look much
  785. 29:09weirder than almost anyone expects
  786. 29:12honestly I think that's the most
  787. 29:13realistic way to look at it nobody
  788. 29:15actually knows if this ends in a
  789. 29:17software only singularity a spark or
  790. 29:19something slower the evidence is partly
  791. 29:21redacted the forecasts have wide error
  792. 29:24bars Dwaresh himself puts a 25% chance
  793. 29:27on AI not matching the best human
  794. 29:29researchers until after 2032. But what
  795. 29:32has changed is the direction. A year
  796. 29:34ago, recursive self-improvement was a
  797. 29:36thought experiment. Today, Anthropic has
  798. 29:39told the SEC it's crossed its own
  799. 29:41threshold. Google says RSI loops are
  800. 29:43part of how it builds Gemini. OpenAI CEO
  801. 29:46is talking about the RSI era. And the
  802. 29:49people who spent years saying, "Slow
  803. 29:50down and look at the evidence," are now
  804. 29:52saying, "We looked at the evidence and
  805. 29:54we changed our minds." That's what a
  806. 29:56tipping point looks like. Not one big
  807. 29:58announcement, but a lot of careful
  808. 30:00people quietly updating at the same
  809. 30:02time. And if they're right, 2027 might
  810. 30:05be the year that AI stops being a tool
  811. 30:07we build and starts becoming the thing
  812. 30:09that builds itself. So, let me know what
  813. 30:11you think. Are we really at the tipping
  814. 30:13point, or is everyone getting carried
  815. 30:14away? Drop your thoughts in the
  816. 30:16comments, and I'll see you in the next
  817. 30:17one.

About this transcript

This page contains the full transcript of AI Just Reached A Tipping Point by TheAIGRID, generated from the public captions YouTube serves with the video. The transcript has 5,515 words across 817 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.