• Home
  • AI

OpenAI Backtracks, Gunning for Superintelligence: Altman Brings His AGI Timeline Closer – ’25 to ’29

Sam Altman unexpectedly brings his timelines to AGI forward, while OpenAI backtrack on superintelligence. None of these changes were heralded, but they are significant. Plus the new year brings new assessments of the true capability of models to automate 'large swathes of the economy'. I'll give my prediction on that front for 2025, announcement a new Simple Bench competition, and showcase Kling 1.6 vs Veo 2 vs Sora, and much more.

wandb.me/simple-bench

(Colab):

AI Insiders:

TheAgentCompany Paper:
Sam Altman Major Interview:
OpenAI Agent Coming Jan 2025:

Altman Singularity:

Altman Original Timeline:

OpenAI Original Emails:

DeepMind Sky News 2014 Article:

Altman Blog Reflections:

OpenAI Changes Who Gets AGI:

OpenAI 5 Levels:

Altman 2015:

OpenAI React to Anthropic:

Microsoft $100B Definition:
Epoch Scramble for Task Benchmark:

GPQA Progress:

Task Length Crucial for ARC-AGI:

RL Environment Tweet:

Jason Wei Talk:

Miles Brundage Tweet:

Jan Leike Tweet:

O1 Pro Losing Money:

Kling 1.6:

Chapters:
00:00 – Introduction
01:03 – Altman Timeline Moves Forward
04:33 – Superintelligence?
06:55 – AGI was not the only pitch
09:26 – AgentCompany and OpenAI New Agent
17:24 – SimpleBench Competition
23:03 – Kling 1.6 vs Veo 2 vs Sora

Non-hype Newsletter:

Podcast:

Joe Lilli
 

  • @JeffMcJunkin says:

    I can’t track Sam’s definitions on AGI. Is it matching Microsoft where it’s effectively impossible? Sigh. These goalposts are constantly in motion.

  • @fletchermclaughlin8971 says:

    This is the best channel for AI news, by far.

    Deep and comprehensive research on the latest developments with the technology itself, not just shallow coverage of yesterday’s drama, as so many others channels focus on.

    Keep up the fantastic work!

    • @aiexplained-official says:

      Wow thanks fletcher

    • @Juttutin says:

      Yeah, I second this. And so far, as far as I can tell, the only one really honing in on which significant aspects of what “general intelligence” might be still absent.

      (I wrote a couple of comments here, and many elsewhere on YouTube, on the current AIs seeming inability to doubt or really question themselves mid reasoning. This may tie very neatly into the lack of common-sense you discuss. )

    • @WillyJunior says:

      Exactly. Nobody SHOCKED or STUNNED, just hype-free, balanced reporting.

    • @Elongatedrider says:

      ​@@WillyJuniorikr it’s so refreshing 😂

    • @AfifFarhati says:

      This channel and byCloud are the best AI channels out there.

  • @jamaalstv says:

    I appreciate you catching out Sam on his contradictions😂

  • @JasoTheRed48F2 says:

    Sam Altman Not Being Super Shady Challenge: IMPOSIBLE

  • @ArianeQube says:

    “very skilled humans” = specialized AI that is better than humans, which in certain domains we already have. He realized they can’t get generalized intelligence and changed the definition to something a lot more realistic.

  • @BunnyOfThunder says:

    To really be AGI, a system should be able to establish a Type III Kardashev scale civilization.

    • @pandosann says:

      What?
      “A Type III civilization is able to capture all the energy emitted by its galaxy, and every object within it, such as every star, black hole, etc.”

      I think you mean type I – II and still you need ASI to achieve this level

    • @andybaldman says:

      No, he means Type III

    • @maciejbala477 says:

      @@pandosann it’s exaggerated on purpose to refer to the moving goalposts, and the 100 billion thing from Microsoft. So, no, he does mean III, because that’s clearly way above what an AGI should be able to achieve

    • @anonymes2884 says:

      Pfft, if you can’t capture all the energy from an ensemble of parallel universes then GTFO.

    • @ticketforlife2103 says:

      AT LEAST

  • @MikelBober says:

    15:05 Thank you for the shoutout! Glad you liked the article

  • @vectoralphaSec says:

    If AGI truly is achieved this year 2025, then that would change everything drastically and would make this year the most important in human history. Well see.

    • @maciejbala477 says:

      eh, sure, but also, not really, because it’s not a hard limit, it’s a spectrum. “AGI” is just an arbitrary designation which lots of people understand differently, as demonstrated by the people claiming we already have AGI now. Unless some kind of new model drops which overshadows everything else (not impossible), then it’s rather going to be a more gradual shift over several years

  • @khonsu0273 says:

    It’s telling that that Sam suddenly changed his tune about objectives and timelines just after the US election. Suddenly, timelines cane right in again. And the definition of AGI keeps shifting to a Microsoft tune.

  • @vethum says:

    Can’t wait to see how o3 scores on SimpleBench

    • @theWACKIIRAQI says:

      Isn’t the question of “how” they reach their score on these benches more important than the score itself?
      Is it actually reasoning? Or brute-forcing its way through without any actual understanding?

    • @vethum says:

      @ Chollet says o3 is a breakthrough in reasoning so idk. We don’t really know how humans reason either.

    • @41-Haiku says:

      @@theWACKIIRAQI The answer to your first question is a flat “No.” Is the important thing about the tiger that it has DNA, or that it has teeth and claws?

      With rare exception, we can’t look inside these systems and identify how they’re making decisions at all. We don’t know how they are doing what they are doing, but we can tell that they are extremely capable. If it is possible to solve novel PHD-level math problems, explain novel jokes, and beat most humans in the world at coding challenges, _all without reasoning,_ then we should be in awe of that ability, and treat it with the respect and caution that it deserves.

      (Reasoning isn’t just one thing, of course, either in humans or in AI. It’s a bundle of useful algorithms. But that’s a story for another time.)

    • @peter-rhodes says:

      o3 won’t be able to be tested “fully”. Because the compute costs are way too high.

    • @tornmap4385 says:

      @@peter-rhodes 03 mini is as cheap as o1 mini

  • @anonymes2884 says:

    Must admit, I don’t expect 2025 to be quiet, the opposite if anything – I suspect this will be a pivotal year because LLMs will either look a lot more hypey, _maybe_ even to the extent of a partial bursting of the bubble, OR they’ll be doing something actually useful by the end of this year (possibly due to _some_ sort of handle on hallucinations, even if only “known percentage reliability” or similar).

    • @netscrooge says:

      It’s already been “actually useful” and accurate for me. Not sure if that’s due to our having different tasks or different expectations.

    • @Novascrub says:

      It’ll be both. We are entering the trough of disillusionment.

    • @zoeherriot says:

      @@netscroogeit just indicates whatever you are doing is commonly found in the training data. I have very mixed results myself, because a lot of my work is creating novel code.

    • @skierpage says:

      AI is doing something actually useful for the people who will replace doubters in their jobs.

    • @Novascrub says:

      @@skierpage tell me you are really into crypto without telling me

  • @thanX says:

    “Cookie banners are the major obstacles between us and AGI” I find that quite illuminating somehow! 😁

    • @maciejbala477 says:

      just an example of silly everyday “problems” that are trivial to a human, but which AI, for all its good sides, struggles with somehow. It happens a whole lot

  • @raideveloper says:

    crazy to think that in 2000 Ray Kurzweil predicted 2029 for AGI, if he really nail it will be astonishing, and will make everyone wait for singularity on his terms in 2045

  • @MarcoRivis says:

    Gemini 2.0 Flash Thinking Exp 17:42

    “The key information here is that the letter transforms randomly into another letter of the alphabet each second. This means that at any given moment, there are 26 possible letters. Since the transformation is random, and Husain is far away and can’t see clearly enough to discern any patterns or slight visual differences, his best strategy is essentially a random guess.
    Therefore, the probability of him correctly guessing the current letter is simply 1 out of the 26 possible letters.
    The correct answer is D) 1/26.
    The information about the font size and distance is a distraction, setting up a scenario where Husain has to guess randomly. His past trouble with ‘w’ and ‘m’ is also a slight red herring, as he’s so far away he can’t even see the letter clearly enough to apply that previous difficulty.”

    • @CodyRay295 says:

      it says it will transform into *another* letter, implying it will be a different letter. Think the answer should technically be 1/25.

    • @authenticallysuperficial9874 says:

      ​@CodyRay295 So in your mind a man “slowly inches away” several football field lengths in less than a second?

    • @cosmiclounge says:

      This tracks with my experience of the Flash 2.0 “Thinking” variant.

    • @41-Haiku says:

      @@CodyRay295 Different from what? If you don’t know what the letter was before, the new letter could be any of them.

    • @coreywoodward4389 says:

      I’ve been using Gemini recently at work, specifically Experimental 1206 as I prefer larger models, and it seems to really understand nuance and novel tasks! Generally I’ll attach a bunch of documents for context (usually exported Notion pages) and a large multi step prompt like Task 1, task 2, task 3, etc. instructing to stop after each task for approval on the output. I’ve had this fail completely with gpt-4o, and Claude is usually at capacity when I try to use it but I’ve had good experience with it in the past. Fyi my boss fully supports these activities.

  • @AngeloWakstein-b7e says:

    Brilliantly done Phillip! I am always looking SO forward to any and all of your videos, Post more often pls.

  • @MiminNB says:

    Oh, there you are!! I have been waiting for this!!

  • @apgd81 says:

    This is superb video with tons of extremely good reference. What a beautiful work you are doing curating all this to us. Good luck with SimpleBench!

  • @stephenrodwell says:

    Thanks! Excellent content, as always. 🙏🏼

  • @exhibitD79 says:

    *A CEO saying anything should NOT be the key reference* …..His job is to present their progress and potential in the most attractive and postive way possible to keep investors interested and the media hype primed.

  • @danverzhao9912 says:

    13:48 This is the reason why I watch the best AI YouTube channel – AI Explained

  • >