This GPT-5 NEWS Could Change EVERYTHING...

00:26:31
https://www.youtube.com/watch?v=I3Wo-23WlQw

概要

TLDRIn the video, the speaker elaborates on the speculation surrounding OpenAI's GPT-5, suggesting that the model has been developed but is withheld from public release to maximize internal benefits. They discuss the recent evolution of AI technologies and model distillation processes, where powerful models improve the performance of smaller, less costly models. The impact of AI industry competition, advancements in smaller models, and the strategic decisions behind delaying releases are examined, ultimately raising questions about the future accessibility of cutting-edge AI systems and the trajectory that OpenAI and other companies might take.

収穫

  • 🔍 Rumors suggest GPT-5 is ready but unreleased.
  • 🧪 Model distillation enhances performance efficiently.
  • 🤔 OpenAI may prioritize internal AI capabilities over public releases.
  • 💡 Smaller models outperform larger predecessors through distillation.
  • 📉 High operational costs impact AI model availability.
  • 🔮 Future AI may increasingly rely on internal improvements.
  • 🛡️ Public access to advanced AI may diminish over time.

タイムライン

  • 00:00:00 - 00:05:00

    The video discusses rumors surrounding GPT-5 and its implications for the AI industry. The central claim is that OpenAI may have developed GPT-5 but is keeping it internal due to better return on investment, hinting at undisclosed advancements that could reshape AI applications.

  • 00:05:00 - 00:10:00

    The article mentions the mysterious absence of Anthropic's Claude Opus 3.5, which was anticipated as a competitor to GPT-4, suggesting that AI model development may involve internal use rather than public release to optimize performance and reduce costs while maintaining competitive edge.

  • 00:10:00 - 00:15:00

    Details on the performance of Claude 3.5 indicate that its results weren't satisfactory enough for public release. Instead, Anthropic might be using the training data generated from this model to enhance its existing models, such as Claude Sonnet 3.6, implying a cycle where models improve incrementally through distillation.

  • 00:15:00 - 00:20:00

    The concept of model distillation emerges as a crucial strategy, wherein powerful models are utilized to enhance smaller models' capabilities. This allows companies to reduce operational costs while simultaneously improving performance, which might be a common practice among leading AI labs, including OpenAI.

  • 00:20:00 - 00:26:31

    The discussion concludes with speculations that OpenAI may be internally refining GPT-5 without releasing it, potentially due to the high costs of inference and improving performance on smaller models. There is also a suggestion that the trajectory of AI model development may shift away from larger models to more efficient and smaller ones, emphasizing the business strategy behind AI releases.

もっと見る

マインドマップ

ビデオQ&A

  • What is GPT-5 rumor about?

    The rumor suggests that OpenAI has developed GPT-5 but is keeping it internal for strategic reasons.

  • How does model distillation work?

    Model distillation is a process where a powerful model generates data to enhance the performance of a smaller, cheaper model.

  • Why might OpenAI choose not to release GPT-5?

    OpenAI may keep GPT-5 internal to leverage its capabilities for distilling knowledge into more user-friendly models and for cost control.

  • What did Anthropic do with Claude Opus 3.5?

    Anthropic allegedly utilized Claude Opus 3.5 internally to improve the performance of Claude Sonnet 3.6.

  • How does AI industry competition affect model releases?

    The lack of necessity to publicize cutting-edge models can lead to a strategic focus on improving internal models rather than external offerings.

  • What is the significance of reduced model sizes?

    The trend shows that newer models can outperform larger predecessors while being smaller and cheaper, thanks to distillation.

  • What are the implications of AI companies prioritizing internal models?

    Companies may focus on using advanced models internally to drive innovation without releasing them publicly.

  • How does distillation help in AI development?

    Distillation allows companies to improve performance without incurring high costs from running larger models.

  • What future advancements are anticipated in AI technology?

    Further enhancements in AI may stem from internal models training smaller public-facing models, continuing a cycle of improvement.

  • How does this trend of AI model development impact user access?

    As companies become more self-sufficient, public access to the latest advancements could diminish.

ビデオをもっと見る

AIを活用したYouTubeの無料動画要約に即アクセス！
字幕
en
オートスクロール:
  • 00:00:00
    so there has been a rumor floating
  • 00:00:02
    around about gbt 5 and in this video
  • 00:00:04
    I'll explain the rumor articulated by
  • 00:00:06
    Alberto Romero in a long winded article
  • 00:00:09
    that dives into the details of where
  • 00:00:11
    exactly GPT 5 is and this isn't just a
  • 00:00:14
    video on GPT 5 it's more about where
  • 00:00:16
    exactly the AI industry is headed as a
  • 00:00:18
    whole because there could be something
  • 00:00:20
    secret that these AI labs are doing that
  • 00:00:22
    you ought to know so you can see right
  • 00:00:24
    here the article starts that this rumor
  • 00:00:26
    is about you know gb5 and it does change
  • 00:00:28
    everything that it isn't an OV
  • 00:00:30
    exaggeration it actually does change
  • 00:00:32
    everything when I get into the details
  • 00:00:34
    because it means that the way that these
  • 00:00:36
    models going to be released is going to
  • 00:00:37
    be completely different so let's take a
  • 00:00:40
    look at what this article says so it
  • 00:00:42
    talks about GPT 5 internally and it says
  • 00:00:44
    that what if I told you that gbt 5 is
  • 00:00:47
    real not just real but already shaping
  • 00:00:49
    the world from where you can't see it so
  • 00:00:51
    this is the hypothesis okay this is the
  • 00:00:53
    entire hypothesis for this entire video
  • 00:00:56
    and credit goes to this user but it says
  • 00:00:58
    the open AI has already built GPT 5 but
  • 00:01:01
    it's kept it internally because the
  • 00:01:03
    return on investment is far greater than
  • 00:01:05
    if they released it to millions of chat
  • 00:01:07
    GPT users and also the ROI they're
  • 00:01:10
    getting is not money but something else
  • 00:01:13
    as you see the idea is simple enough but
  • 00:01:16
    the challenge is connecting the
  • 00:01:17
    breadcrumbs that lead up to it and this
  • 00:01:19
    article is a deep dive into why I
  • 00:01:20
    believe it all adds up so overall this
  • 00:01:23
    article is basically say that look
  • 00:01:25
    opening I have created GPT 5 and it is a
  • 00:01:28
    stunning model like the model is truly
  • 00:01:30
    stunning but they've decided not to
  • 00:01:32
    release this model because of a reason
  • 00:01:34
    that we're going to get into into this
  • 00:01:36
    article and that reason is particularly
  • 00:01:39
    interesting so let's go ahead and take a
  • 00:01:40
    look at this so one of the things they
  • 00:01:42
    actually talk about here is they talk
  • 00:01:44
    about the mysterious disappearance of
  • 00:01:45
    Opus 3.5 says before going into gbt 5 we
  • 00:01:48
    have to pay a visit to its distant
  • 00:01:49
    cousin also missing an action anthropics
  • 00:01:52
    clae Opus 3.5 and as you know the top
  • 00:01:55
    three AI labs open eye Google and deep
  • 00:01:57
    mind and anthropic offer a range of
  • 00:01:59
    models designed to span the price
  • 00:02:01
    latency versus performance Spectrum now
  • 00:02:03
    open AI provides options like GPT 40 gp2
  • 00:02:06
    40 mini as well as 01 and 01 mini and
  • 00:02:09
    Google Deep Mind offers Gemini Ultra Pro
  • 00:02:12
    and Flash while in thoric has Claude
  • 00:02:14
    Opus son and Hau and basically talking
  • 00:02:16
    about how with all of these different
  • 00:02:17
    models you're trying to cater to as many
  • 00:02:19
    customers as possible some prioritize
  • 00:02:21
    top T performance no matter the cost
  • 00:02:23
    While others seek affordable Solutions
  • 00:02:25
    which are you know good enough now
  • 00:02:26
    they're talking about the mysterious
  • 00:02:27
    disappearance of this model the Opus 3.5
  • 00:02:30
    model that was meant to be coming in the
  • 00:02:32
    series you see right here it says
  • 00:02:33
    something strange happened in October
  • 00:02:35
    2024 everyone was expecting anthropic to
  • 00:02:37
    announced claw 3.5 Opus as a response to
  • 00:02:40
    GPT 40 instead they released an updated
  • 00:02:43
    version of claw 3.5 Sonic that people
  • 00:02:45
    started a course on in 3.6 now Opus 3.5
  • 00:02:49
    was nowhere to be found remember this
  • 00:02:51
    was supposed to be like the GPT 5 type
  • 00:02:53
    model and it says this model was nowh to
  • 00:02:55
    be found seemingly leaving anthropic
  • 00:02:56
    without a direct competitor to GPT 40
  • 00:02:59
    but when we take a look at actually what
  • 00:03:00
    happened with Opa 3.5 it's rather
  • 00:03:02
    fascinating so they talk about how you
  • 00:03:03
    know um previously in October he wrote
  • 00:03:06
    in his nuclear weat post that there are
  • 00:03:08
    rumors that Sonic 3.6 which is you know
  • 00:03:10
    a really good model is just you know a
  • 00:03:12
    checkpoint of a failed training run on
  • 00:03:14
    the much anticipated Opus 3.5 and that
  • 00:03:16
    was because at the time Claude Opus 3.5
  • 00:03:19
    according to the web pages a lot of
  • 00:03:21
    people did think at the time that it was
  • 00:03:23
    scrapped because previously they were
  • 00:03:24
    stating that look claw 3.5 Opus is
  • 00:03:26
    coming it's coming and then they kind of
  • 00:03:28
    just removed the fact that the model was
  • 00:03:30
    going to be coming from a lot of
  • 00:03:31
    different web pages and recently there
  • 00:03:33
    was actually an interview with the CEO
  • 00:03:35
    of anthropic and this stuff about Opus
  • 00:03:37
    3.5 is actually all going to relate to
  • 00:03:38
    GPT 5 in a moment but the reason we're
  • 00:03:41
    talking about Opus 3.5 is because that
  • 00:03:43
    is the model that is supposed to be a
  • 00:03:45
    gp5 type model so just prefacing what
  • 00:03:48
    we're about to go into here is that
  • 00:03:49
    we're looking at how Opa 3.5 which is
  • 00:03:52
    supposed to be the GPT 5 type competitor
  • 00:03:54
    how anthropic have done that and how
  • 00:03:57
    revealing what anthropic have done
  • 00:03:58
    reveals kind of what done so they talk
  • 00:04:01
    about how you know Dario amade basically
  • 00:04:02
    says that you know on Opus 3.5 not
  • 00:04:05
    giving you an exact date but as far we
  • 00:04:06
    as we know the plan is still to have a
  • 00:04:08
    claw 3.5 Opus cautious yet ambiguous yet
  • 00:04:12
    valid well ridiculous timeline question
  • 00:04:14
    uh when is cloud Opus uh 3.5 coming up
  • 00:04:18
    uh not giving you an exact date uh but
  • 00:04:20
    you know they're they uh you know as far
  • 00:04:23
    as we know the plan is still to have a
  • 00:04:24
    Claude 3.5 Opus so right here you can
  • 00:04:27
    see that this is where we talk about the
  • 00:04:29
    fact that you know in Bloomberg there
  • 00:04:30
    were many articles at the time that were
  • 00:04:32
    interestingly enough talking about you
  • 00:04:34
    know open AI Google and anthropic are
  • 00:04:36
    struggling to build more advanced Ai and
  • 00:04:38
    this was going from the jump from GPT 4
  • 00:04:39
    to a GPT 5 type model and this article
  • 00:04:41
    is really fascinating cuz I remember at
  • 00:04:43
    the time I was reading it and it was
  • 00:04:44
    like Gemini 2 was you know disappointing
  • 00:04:46
    they were saying the GPT 5 was
  • 00:04:47
    disappointing and the OPA 3.5 they they
  • 00:04:50
    all performed essentially better but not
  • 00:04:52
    better enough so he talks about here you
  • 00:04:54
    know Bloomberg wrote an article that you
  • 00:04:56
    know after training it at anthropic
  • 00:04:57
    found that 3.5 Opus which is upt 5
  • 00:05:00
    competitor performs better on
  • 00:05:02
    evaluations than the older version but
  • 00:05:03
    not buy as much as it should given the
  • 00:05:05
    size of the model and how costly it was
  • 00:05:07
    to build and run and it seems that Dario
  • 00:05:08
    refrained from giving a date because
  • 00:05:10
    although the Opus 3.5 training run
  • 00:05:11
    hadn't failed its results were
  • 00:05:13
    underwhelming and the basically what
  • 00:05:15
    they actually stating about this was
  • 00:05:16
    that a lot of people missed was the fact
  • 00:05:18
    that these next iterational models gbt 5
  • 00:05:20
    Gemini 2 those models were you know good
  • 00:05:23
    but the only bad thing about them was
  • 00:05:25
    that they were just quite expensive for
  • 00:05:27
    what they were and TR trust me you want
  • 00:05:28
    to remember this fact because it you
  • 00:05:30
    know it ties into you know a part of why
  • 00:05:32
    they didn't release these models now
  • 00:05:33
    here's where we get some Insider
  • 00:05:35
    information SL some leaks and it says on
  • 00:05:37
    December 11th semiconductor expert Dylan
  • 00:05:39
    Patel and his semi analysis team
  • 00:05:41
    delivered the final plot twist
  • 00:05:43
    presenting an explanation that weaves
  • 00:05:45
    all the data points into a coherent
  • 00:05:46
    story says anthropic finished training
  • 00:05:49
    claw 3.5 Opus and it did perform well
  • 00:05:52
    with it scaling appropriately yet for
  • 00:05:53
    some reason anthropic didn't release it
  • 00:05:55
    so anthropic managed to get their GPT 5
  • 00:05:58
    model performing well doing very well
  • 00:06:00
    but apparently they didn't release it
  • 00:06:01
    and he says this is because instead of
  • 00:06:03
    releasing it publicly and this is where
  • 00:06:04
    we get to the crazy theories that
  • 00:06:06
    anthropic used claw 3.5 Opus to generate
  • 00:06:09
    synthetic data and for reward modeling
  • 00:06:12
    to improve Claude 3.5 Sonic
  • 00:06:15
    significantly alongside user data now
  • 00:06:18
    that's a crazy crazy statement but this
  • 00:06:20
    is from a reputable Source within the
  • 00:06:22
    industry there's previously spoken about
  • 00:06:24
    things before they've been released and
  • 00:06:25
    later it's come out that that
  • 00:06:27
    information has been true and this is
  • 00:06:31
    basically one of the key theories
  • 00:06:33
    driving behind why many people think
  • 00:06:35
    that GPT 5 is Alive and Well internally
  • 00:06:38
    but isn't currently yet being released
  • 00:06:40
    to the public and I mean it kind of
  • 00:06:42
    makes sense when we actually take a look
  • 00:06:43
    at more things because we all know just
  • 00:06:45
    how good claw 3.6 Sonet is or claw 3.5
  • 00:06:48
    Sonet is whatever you want to call it
  • 00:06:49
    but many people can't figure out why is
  • 00:06:51
    claw 3.5 Sonet so good or claw 3.6 on it
  • 00:06:54
    so good but they haven't released Claude
  • 00:06:56
    Opus just yet and so you see here it
  • 00:06:57
    says in short anthropic did TR CLA 3.5
  • 00:07:00
    Opus they dropped the name because it
  • 00:07:02
    wasn't good enough and Dario confident a
  • 00:07:04
    different training run could improve the
  • 00:07:05
    results avoided you know giving a date
  • 00:07:08
    and blueberg confirmed that the results
  • 00:07:10
    were better than existing models but not
  • 00:07:12
    enough to justify the inference cost
  • 00:07:13
    which is of course how much these models
  • 00:07:15
    cost to run when people are using the
  • 00:07:17
    models and it says here that Dylan and
  • 00:07:19
    his team uncovered the link between the
  • 00:07:21
    mysterious Sonet 3.6 and the missing
  • 00:07:23
    Opus 3.5 the latter was being used
  • 00:07:25
    internally to generate synthetic data to
  • 00:07:27
    boost the former's performance and so
  • 00:07:29
    this is where we have this kind of thing
  • 00:07:31
    right here and trust me it's going to
  • 00:07:32
    link to GPT 5 in a moment but we can see
  • 00:07:34
    here that this GPT file star model from
  • 00:07:36
    anthropic the one of the biggest Rivals
  • 00:07:37
    to you know one of the leading Labs we
  • 00:07:39
    can see that that model right there is
  • 00:07:41
    you know going down into Sonic 3.6 and
  • 00:07:44
    this is basically where um you know
  • 00:07:45
    they're distilling the model so
  • 00:07:47
    essentially you could say that they
  • 00:07:48
    internally built a really really smart
  • 00:07:50
    AI system but they probably haven't
  • 00:07:52
    released it to the public just yet and
  • 00:07:53
    we can see right here that you know
  • 00:07:54
    publicly Sonic 3.6 that's what's going
  • 00:07:56
    on but um yeah it's pretty crazy now it
  • 00:07:59
    gets crazier here so this is where they
  • 00:08:00
    start to explain model distillation this
  • 00:08:02
    is what they say better but also smaller
  • 00:08:03
    and cheaper the process of using a
  • 00:08:05
    powerful expensive model to generate
  • 00:08:07
    data that enhances the performance of a
  • 00:08:09
    slightly less capable model is known as
  • 00:08:11
    distillation and it's a common practice
  • 00:08:13
    where this technique allows a labs to
  • 00:08:15
    use their smaller models beyond what
  • 00:08:17
    could be achieved through additional
  • 00:08:18
    pre-training alone so one way that a lot
  • 00:08:20
    of these companies are managing to get
  • 00:08:22
    these smaller models to be better and
  • 00:08:23
    better is through model distillation so
  • 00:08:25
    you have a teacher model then you
  • 00:08:27
    distill that knowledge and you transfer
  • 00:08:29
    to a student model and this is something
  • 00:08:31
    that just performs a lot better than the
  • 00:08:33
    standard you know pre-training and that
  • 00:08:35
    entire Paradigm that we used to do so
  • 00:08:37
    this is something that has been you know
  • 00:08:38
    going pretty well for AI companies and
  • 00:08:40
    it's going to be something that we
  • 00:08:41
    continue to do now you can see right
  • 00:08:43
    here that it says that there are various
  • 00:08:44
    approaches to distillation but we're not
  • 00:08:46
    getting into them what you need to
  • 00:08:47
    remember is that a strong model acting
  • 00:08:49
    as a teacher turns student models from
  • 00:08:51
    small cheap and fast and weak into small
  • 00:08:54
    cheap fast and Powerful so distillation
  • 00:08:56
    turns a strong model into a gold mine
  • 00:08:58
    and Dylan explains why it made sense for
  • 00:09:00
    anthropic to do this with Opus 3.5 and
  • 00:09:02
    Sonet 3.6 so it talks about how the
  • 00:09:04
    inference costs of the new Sonet versus
  • 00:09:06
    the old Sonic they didn't change
  • 00:09:08
    drastically but the model's performance
  • 00:09:10
    did why release Opus 3.5 on a cost basis
  • 00:09:13
    when it does not make economic sense to
  • 00:09:14
    do so relative to releasing a 3.5 Sonic
  • 00:09:17
    with further Pro training from said 3.5
  • 00:09:20
    Opus basically saying that look there's
  • 00:09:22
    no point us going ahead and releasing
  • 00:09:23
    Opa 3.5 when it's so expensive and
  • 00:09:25
    costly to run why don't we just you know
  • 00:09:27
    train SL distill a lot of the
  • 00:09:29
    capabilities from this Opus model which
  • 00:09:31
    is so amazing into a 3.6 Sonic model
  • 00:09:34
    that they would love anyways so this is
  • 00:09:36
    super super fascinating here and of
  • 00:09:38
    course this is like what I said already
  • 00:09:40
    anthropic chose not to release this you
  • 00:09:41
    know not because of the poor results but
  • 00:09:43
    because it's just more valuable
  • 00:09:44
    internally for them to use so they
  • 00:09:46
    basically use this AI to train the next
  • 00:09:48
    set of AIS and this is why this is kind
  • 00:09:50
    of changing everything because a lot of
  • 00:09:51
    people have speculated for quite some
  • 00:09:53
    time that this is going to be the moment
  • 00:09:55
    where these AI systems are going to just
  • 00:09:58
    keep getting better and better better
  • 00:09:59
    because you use one to train the next
  • 00:10:01
    version and then use that version to
  • 00:10:02
    train next version and then you know so
  • 00:10:04
    far you're just going to you know
  • 00:10:06
    continue to expand in terms of the
  • 00:10:07
    intelligence so what's crazy about this
  • 00:10:09
    is that you know he also says that you
  • 00:10:11
    know Dylan Guy the one that we're
  • 00:10:12
    referencing here he says that that is
  • 00:10:14
    why the open source Community caught up
  • 00:10:16
    to GPT 4 so quickly they were basically
  • 00:10:18
    taking the gold straight from opening
  • 00:10:19
    eyes mind so people were essentially
  • 00:10:21
    distilling GPT 4's capabilities into
  • 00:10:24
    smaller models and that's you know
  • 00:10:25
    pretty interesting because GPT 40 is a
  • 00:10:28
    smaller model it's much faster and it's
  • 00:10:29
    much better but it's also really really
  • 00:10:31
    useful and that's essentially what these
  • 00:10:33
    other companies are doing as well
  • 00:10:34
    they're using those capabilities and
  • 00:10:36
    that's how they're getting these open
  • 00:10:37
    source projects up to GPT 4's level so
  • 00:10:40
    you can see here that you know one of
  • 00:10:41
    the things that we all know is that
  • 00:10:42
    Sonic 3.6 wasn't just good it was stated
  • 00:10:45
    the art good and better than GPT 40 and
  • 00:10:47
    of course anthropics mid-tier model
  • 00:10:49
    outperformed opening eyes flank ship
  • 00:10:50
    models thanks to distillation from Opus
  • 00:10:53
    3.5 and of course other reasons as well
  • 00:10:55
    so that was something that was like
  • 00:10:56
    pretty crazy and then this is where we
  • 00:10:58
    talk about you know bigger and better is
  • 00:11:00
    no longer the Paradigm so once these top
  • 00:11:02
    labs are you know no longer talking
  • 00:11:03
    about higher parameter counts being you
  • 00:11:05
    know better and the last time we got
  • 00:11:07
    knowledge about the parameter counts we
  • 00:11:08
    actually knew that GPT 3.5 was 175
  • 00:11:11
    billion parameters and GPT 4 there were
  • 00:11:13
    rumors saying that GPT 4 was 1.8
  • 00:11:15
    trillion parameters in a mixture of
  • 00:11:17
    experts but the craziest thing about
  • 00:11:18
    this is that they actually now speculate
  • 00:11:21
    that the future models like gpc5 and
  • 00:11:22
    Sonet 3.6 or like whatever distilled
  • 00:11:24
    models that we're getting like GPT 4 and
  • 00:11:26
    Sonic 3.6 are significantly smaller than
  • 00:11:28
    GPT 4 despite them both being better
  • 00:11:30
    than gbt 4 across both benchmarks so gbt
  • 00:11:33
    4 is 1.8 trillion parameters that's what
  • 00:11:35
    I'm trying to explain to you guys gbt 4
  • 00:11:36
    1.8 trillion parameters great model at
  • 00:11:38
    the time of release I think around two
  • 00:11:40
    years from now or one year and a half
  • 00:11:41
    ago and at the time that model was
  • 00:11:43
    considered crazy steady up but now we
  • 00:11:45
    got smaller models like gbt 40 and Son
  • 00:11:48
    3.6 that are significantly smaller than
  • 00:11:50
    this large model but they are better
  • 00:11:52
    because of distillation and that just
  • 00:11:54
    means that the knowledge in them is a
  • 00:11:55
    lot more efficient and the models are a
  • 00:11:58
    lot more smarter so you see right here
  • 00:11:59
    that it says you know current models
  • 00:12:00
    such as the original gbt 40 are probably
  • 00:12:03
    an order of magnitude smaller than gbt 4
  • 00:12:06
    with 40 having around 200 billion and
  • 00:12:08
    3.5 Sonet having around 400 billion
  • 00:12:09
    parameters though this estimate could be
  • 00:12:11
    off by a factor of two given the rough
  • 00:12:13
    we I've arrived at it and the point here
  • 00:12:14
    is that like the thing that you want to
  • 00:12:16
    pay attention to is that both these
  • 00:12:18
    companies are following a similar
  • 00:12:19
    trajectory their latest models are not
  • 00:12:21
    only better but also smaller and cheaper
  • 00:12:23
    than the previous generation we know how
  • 00:12:25
    anthropic pulled it off by distilling
  • 00:12:27
    Opus 3.5 into Sonet 3.6 but this is
  • 00:12:30
    where they get into something
  • 00:12:31
    interesting so what did open AI do
  • 00:12:33
    because of anthropic we know that they
  • 00:12:35
    trained op 3.5 and put the knowledge of
  • 00:12:37
    that model distilled it down into Sonic
  • 00:12:39
    3.6 what on Earth did opening I do and
  • 00:12:42
    this is where the GPT 5 Rumor comes into
  • 00:12:44
    account so we can take a look at this
  • 00:12:45
    diagram and basically shows that you
  • 00:12:46
    know we have the teacher models and we
  • 00:12:49
    have the distillation models and it's
  • 00:12:50
    basically saying that maybe opening eye
  • 00:12:52
    has a very secret internal model it
  • 00:12:54
    might be gbt 5 it might be a much
  • 00:12:56
    smarter model considering you know all
  • 00:12:58
    the things we've heard recently but we
  • 00:12:59
    can see here that these models are going
  • 00:13:01
    to be models that are very smart but
  • 00:13:02
    they're only used for distillation to
  • 00:13:05
    public models that are better smaller
  • 00:13:06
    and cheaper to actually run and this is
  • 00:13:09
    pretty crazy when you think about it so
  • 00:13:11
    internally at anthropic it's quite
  • 00:13:12
    likely they have Opus 3.5 which would be
  • 00:13:14
    incredible to use but they managed to
  • 00:13:16
    distill it down into Sonic 3.6 and Sonic
  • 00:13:18
    3.6 as you already know is pretty
  • 00:13:20
    credible and this is where they're
  • 00:13:21
    saying look with open eye it's
  • 00:13:23
    potentially true that we have a GPD 5
  • 00:13:26
    type model that's distilling all of this
  • 00:13:28
    information that the these you know
  • 00:13:29
    models that we're getting now these much
  • 00:13:31
    smaller ones that we're currently using
  • 00:13:32
    and I do know that there was some kind
  • 00:13:33
    of distillation going on with the
  • 00:13:35
    strawberry model so opening ey are
  • 00:13:36
    definitely familiar with that
  • 00:13:37
    distillation process it's quite like how
  • 00:13:39
    we had the 01 model and we had the 01
  • 00:13:41
    previews models and how effective those
  • 00:13:43
    ones were now when we get onto chapter 3
  • 00:13:45
    it says that one might assume the
  • 00:13:46
    anthropic distillation approach was
  • 00:13:47
    driven by unique circumstances an
  • 00:13:49
    underwhelming training run for Opus 3.5
  • 00:13:52
    but that is something that all of these
  • 00:13:54
    companies have experienced and the key
  • 00:13:55
    thing about this is that like the news
  • 00:13:57
    kind of flipped on everyone's head
  • 00:13:58
    because everyone took the wrong thing
  • 00:14:00
    away from the news they said they had
  • 00:14:02
    subpar training ones but subpar it
  • 00:14:04
    doesn't mean that it was worse because
  • 00:14:06
    think about like this okay like you can
  • 00:14:08
    have something that subpar but it
  • 00:14:09
    doesn't mean that it was worse than you
  • 00:14:10
    did before it just means that it didn't
  • 00:14:11
    live up to your expectations like
  • 00:14:13
    imagine you know you're expecting to win
  • 00:14:14
    $10,000 in a competition and you only
  • 00:14:17
    won $11,000 more than you usually win
  • 00:14:19
    like that's probably subpar but it
  • 00:14:21
    doesn't mean that it wasn't an
  • 00:14:22
    improvement and you have to remember
  • 00:14:23
    that like with intelligence every inch
  • 00:14:25
    that you gain unlocks a whole new host
  • 00:14:27
    of capability so they say say that the
  • 00:14:29
    causes for this don't matter to us
  • 00:14:31
    diminishing returns for lack of data
  • 00:14:32
    whatever the case is you know it doesn't
  • 00:14:34
    really matter because all of these
  • 00:14:35
    companies are going through at the same
  • 00:14:36
    time and basically this is why they talk
  • 00:14:38
    about you know one of the key things for
  • 00:14:39
    these companies that most people don't
  • 00:14:40
    understand is that you know if 300
  • 00:14:42
    million people are using your product
  • 00:14:43
    weekly the operational expenditures can
  • 00:14:45
    you know suddenly kill your company
  • 00:14:47
    which is really important like costs
  • 00:14:49
    matter costs really do matter for these
  • 00:14:51
    companies because there are so many
  • 00:14:52
    people using them there's so many
  • 00:14:53
    servers and you know this is something
  • 00:14:55
    that continually is expanding week on
  • 00:14:56
    week so it's pretty hard to keep up with
  • 00:14:59
    said demand and so this is why they talk
  • 00:15:00
    about how distillation was really good
  • 00:15:02
    because whatever drove anthropic to
  • 00:15:03
    distill Sonet 3.6 to Opus 3.5 is
  • 00:15:06
    affecting openingi several times over
  • 00:15:09
    and talks about you know it basically
  • 00:15:10
    just talks about here how distillation
  • 00:15:11
    works because it Bridges these two
  • 00:15:13
    Universal challenges into an advantage
  • 00:15:15
    you solve the inference cost Problem by
  • 00:15:16
    serving people a smaller model and avoid
  • 00:15:18
    the public backlash for underwelming
  • 00:15:19
    performance by not releasing a larger
  • 00:15:21
    one now of course as most of you guys
  • 00:15:23
    might know one of the things that you
  • 00:15:25
    can't do anymore is of course
  • 00:15:26
    overtraining so these AI Labs have
  • 00:15:28
    actually exhausted all the high quality
  • 00:15:29
    data sources for pre-training and this
  • 00:15:31
    is actually something that Elon Musk and
  • 00:15:32
    Elia satk have admitted in recent weeks
  • 00:15:35
    he says we're back at distillation I
  • 00:15:37
    think that both GPT 40 and Claw 3.5
  • 00:15:39
    Sonic have been distilled down from
  • 00:15:41
    larger models so this is something that
  • 00:15:42
    you know they both actually talk about
  • 00:15:44
    as a realistic things to the point where
  • 00:15:46
    like they're saying that the way how we
  • 00:15:48
    now get to a next level of model is that
  • 00:15:50
    we can't just put more and more data in
  • 00:15:51
    new Innovations are needed and the basic
  • 00:15:53
    thing that look distillation is probably
  • 00:15:55
    the only way that these models are going
  • 00:15:57
    to be getting better so you can see here
  • 00:15:58
    it says every piece of the puzzle so far
  • 00:16:00
    suggests that open ey is doing what
  • 00:16:01
    anthropic did with Opus 3.5 train the
  • 00:16:04
    model and hide the model in the same way
  • 00:16:06
    through distillation and for the same
  • 00:16:08
    reasons because of course there are poor
  • 00:16:09
    results in terms of the cost and that's
  • 00:16:12
    quite the discovery because Opus 3.5 is
  • 00:16:14
    still hidden but we have to think about
  • 00:16:16
    it where is opening eyes and nagalas
  • 00:16:18
    model is it hiding in the company's
  • 00:16:19
    basement care to venture a name and this
  • 00:16:22
    is why open ey are currently distilling
  • 00:16:23
    that model now what's really interesting
  • 00:16:25
    as well is that they talk about how he
  • 00:16:26
    who blazes the trail must clear the pass
  • 00:16:28
    so he he said I started this by
  • 00:16:30
    analyzing anthropics Opus 3.5 story
  • 00:16:32
    because it's the one where we have more
  • 00:16:33
    information then I traced a bridge to
  • 00:16:35
    open ey with the concept of distillation
  • 00:16:37
    and explained why the underlying forces
  • 00:16:39
    pushing anthropic are also pushing open
  • 00:16:41
    ey but there's a new obstacle in theory
  • 00:16:44
    because openi is the Pioneer they might
  • 00:16:46
    be facing obstacles that like anthropic
  • 00:16:48
    simply haven't found yet because if
  • 00:16:49
    you're innovating you're going to you
  • 00:16:51
    know face problems that nobody else has
  • 00:16:52
    seen yet so you need to clear that path
  • 00:16:54
    and it says here that one obstacle is
  • 00:16:55
    the hardware requirements to train GPT 5
  • 00:16:58
    so at 3 2.6 is comparable to GPT 40 but
  • 00:17:01
    it was released with a 5mon lag we
  • 00:17:03
    should assume GPT 5 is on another level
  • 00:17:05
    more powerful and bigger also more
  • 00:17:07
    expensive not only to inference but also
  • 00:17:09
    to train we could be talking about a
  • 00:17:11
    half billion dollar training run would
  • 00:17:13
    it even be possible to do such a thing
  • 00:17:15
    with current hardware and yes that would
  • 00:17:17
    be possible but the crazy thing about
  • 00:17:19
    this is that you know you wouldn't be
  • 00:17:20
    able to have inflence over that model so
  • 00:17:22
    basically what they stating here is that
  • 00:17:23
    these companies are probably still
  • 00:17:25
    scaling the models internally and doing
  • 00:17:27
    insane training runs but the only way
  • 00:17:29
    that they can actually you know provide
  • 00:17:31
    the inference you know to get these
  • 00:17:32
    models out into the public is to distill
  • 00:17:34
    those capabilities down into a smaller
  • 00:17:36
    model and it says you know in principle
  • 00:17:38
    our Cent Hardware is good enough to
  • 00:17:39
    serve models much bigger than GPT 4 for
  • 00:17:42
    example a 50 times scaled up version of
  • 00:17:44
    GPT 4 having around 100 trillion
  • 00:17:46
    parameters could probably be served at
  • 00:17:49
    $3,000 per million token and 10 to 20
  • 00:17:51
    tokens per second of output speed
  • 00:17:53
    however for this to be viable those big
  • 00:17:55
    models would have to unlock a lot of
  • 00:17:56
    economic value for the customers using
  • 00:17:58
    them so so of course this is the kind of
  • 00:18:00
    reason why they don't release it and
  • 00:18:01
    this is super interesting because we
  • 00:18:02
    know that these companies are always
  • 00:18:04
    currently struggling for inference CU
  • 00:18:05
    they're trying to do a lot of research
  • 00:18:06
    and all these kind of things but it will
  • 00:18:08
    be interesting to see what is actually
  • 00:18:09
    going on behind the scenes here because
  • 00:18:11
    if they are training up a model that is
  • 00:18:12
    that big then they could do a lot of
  • 00:18:15
    things with that but the only thing I
  • 00:18:16
    would say that this article doesn't
  • 00:18:17
    cover at the moment is the fact that gbt
  • 00:18:19
    40 I remember reading a few posts about
  • 00:18:21
    this and I do remember reading that gbt
  • 00:18:23
    4 is a model that is an omn model that
  • 00:18:26
    built from the ground up to be
  • 00:18:27
    multimodel in and multimodo out so I'm
  • 00:18:30
    not sure like it was just an llm but
  • 00:18:32
    maybe there are more details on that but
  • 00:18:34
    I do remember reading about GPT 40 being
  • 00:18:36
    like this Omni model so it was trained
  • 00:18:38
    on audio in audio out and I do remember
  • 00:18:39
    like reading the you know research paper
  • 00:18:41
    the entire full thing where I did you
  • 00:18:43
    know 30 40 minute you know details going
  • 00:18:45
    into it where you can actually see that
  • 00:18:46
    the model is capable of a lot of
  • 00:18:48
    different things so I do think that if
  • 00:18:50
    they do have a gbt 5 type model I don't
  • 00:18:52
    think they've just D it down into gbt 40
  • 00:18:54
    yet I think that model is definitely
  • 00:18:55
    going to be coming in the future and
  • 00:18:56
    they talk about you know spending that
  • 00:18:58
    kind of inference money is not even just
  • 00:18:59
    aable for Microsoft Google or Amazon
  • 00:19:01
    because of course they need to unlock a
  • 00:19:03
    lot of economic value if they plan to
  • 00:19:04
    serve this several trillion parameter
  • 00:19:06
    model to the public so they don't so
  • 00:19:08
    they train it they realize it performs
  • 00:19:09
    better than their current offerings but
  • 00:19:11
    they have to accept it as it hasn't
  • 00:19:12
    Advanced enough to justify the enormous
  • 00:19:15
    cost of keeping it running and that's
  • 00:19:16
    essentially what the Wall Street you
  • 00:19:17
    know Journal said on gbt 5 a month ago
  • 00:19:20
    and what Bloomberg said about Opus 3.5
  • 00:19:22
    now other the thing that they said if
  • 00:19:23
    opening ey were hypothetically
  • 00:19:24
    withholding GPT 5 under the preex status
  • 00:19:26
    not already they would achieve one more
  • 00:19:28
    thing besides cost control and
  • 00:19:29
    preventing public backlash so if openi
  • 00:19:31
    were hypothetically withholding GPT 5
  • 00:19:33
    under the pretext that it's not ready
  • 00:19:35
    they would achieve one more thing
  • 00:19:36
    besides the cost control and preventing
  • 00:19:38
    the public backlash that actually
  • 00:19:40
    sidestep the need to declare whether or
  • 00:19:42
    not it meets the threshold for being
  • 00:19:43
    categorized as AI as you know they have
  • 00:19:45
    a contract with Microsoft that says you
  • 00:19:47
    know AGI is a system that can generate
  • 00:19:49
    at least $100 in profits maybe Microsoft
  • 00:19:52
    would say that if people are able to
  • 00:19:53
    build rappers out of that and that's
  • 00:19:55
    able to get them to 100 billion maybe
  • 00:19:57
    they wouldn't mind triggering the a
  • 00:19:58
    clause and parting ways with Microsoft
  • 00:20:01
    so you know potentially if they were
  • 00:20:02
    looking at 100 billion in annual
  • 00:20:04
    recurring revenue from gbt 5 maybe they
  • 00:20:07
    wouldn't care now this is where the
  • 00:20:09
    theory starts to evolve into something
  • 00:20:10
    really crazy and it starts to finalize
  • 00:20:12
    so this is where the theory starts to
  • 00:20:14
    basically say that look they might not
  • 00:20:15
    need us so even if that were true no
  • 00:20:17
    skeptic has stopped to think that openi
  • 00:20:19
    may have a better internal use case than
  • 00:20:21
    whatever they'd get from it externally
  • 00:20:23
    there's vast differences between
  • 00:20:24
    creating an excellent model and creating
  • 00:20:26
    an excellent model that can be served
  • 00:20:27
    cheaply to 300 million people if you
  • 00:20:29
    can't you don't but also if you don't
  • 00:20:32
    need to you don't now it says here and
  • 00:20:33
    this is crazy okay that they were giving
  • 00:20:35
    us access to their best model because
  • 00:20:37
    they needed our data but not so much
  • 00:20:39
    anymore they're not chasing our money
  • 00:20:40
    either that's Microsoft but not them
  • 00:20:43
    they want a AGI and then they want ASI
  • 00:20:45
    and they want a legacy they're basically
  • 00:20:47
    stating that look before they needed to
  • 00:20:48
    train on user data and they needed to
  • 00:20:50
    figure out where to scale the model but
  • 00:20:52
    now that they have gbt 5 and whatever
  • 00:20:53
    internal models they're using to distill
  • 00:20:55
    down the knowledge into simple products
  • 00:20:57
    that we can use and get a decent amount
  • 00:20:59
    of value from on a day-to-day basis they
  • 00:21:01
    don't really need to provide us with the
  • 00:21:02
    Cutting Edge pieces of Technology
  • 00:21:04
    anymore simply because they don't need
  • 00:21:06
    to serve them anymore they just need to
  • 00:21:08
    use them themselves in order to develop
  • 00:21:09
    better products and better technology
  • 00:21:11
    and that's kind of like saying that you
  • 00:21:13
    know when open AI develops their own
  • 00:21:15
    internal AGI systems then they're just
  • 00:21:16
    going to use that themselves in order to
  • 00:21:18
    actually make money so this is where you
  • 00:21:20
    can see here that this is what they're
  • 00:21:21
    talking about internally their private
  • 00:21:23
    models are Opus 3.5 and privately they
  • 00:21:25
    have gbt 5 and they distill those
  • 00:21:27
    capabilities into these models I'm not
  • 00:21:29
    sure about the distilling into gbt 40
  • 00:21:32
    but maybe it's definitely being used to
  • 00:21:34
    help with some of the reinforcement
  • 00:21:35
    learning for future models and some of
  • 00:21:37
    the synthetic data generation and I have
  • 00:21:39
    heard that that is the case with 01/03
  • 00:21:42
    and you can see right here it says we're
  • 00:21:43
    nearing the end I believe i' laid out
  • 00:21:45
    enough arguments to make a solid case
  • 00:21:47
    open ey likely has gbt 5 working
  • 00:21:49
    internally just as anthropic does with
  • 00:21:51
    Opus 3.5 but it's quite plausible that
  • 00:21:53
    openi never releases GPT 5 at all the
  • 00:21:56
    public now measures performance again 01
  • 00:21:58
    sl3 Not Just gbt 4 or Claude sonit 3.5
  • 00:22:02
    now with the test Time new scaling laws
  • 00:22:04
    the bar for gbt 5 to clear keeps Rising
  • 00:22:07
    how could they ever release a GPT 5 that
  • 00:22:09
    truly outshines 01 and 03 and the coming
  • 00:22:12
    O Series models at the pace they're
  • 00:22:14
    producing them besides that they no
  • 00:22:16
    longer need our money or our data
  • 00:22:18
    anymore basically saying that look when
  • 00:22:19
    we take a look at how the fact that
  • 00:22:21
    these 01 series models are just so
  • 00:22:22
    incredible when it comes to the raw
  • 00:22:24
    capabilities of reasoning why on Earth
  • 00:22:26
    would they release a GPT 5 level model
  • 00:22:27
    at all when those smaller models that
  • 00:22:29
    we're currently getting with gbt 4 type
  • 00:22:31
    systems are just a lot more expensive
  • 00:22:33
    for incremental gains so you can see
  • 00:22:34
    right here it says training new base
  • 00:22:36
    models like gbt 5 gbt 6 and Beyond will
  • 00:22:38
    always make sense for op ey internally
  • 00:22:40
    but not necessarily as products that
  • 00:22:42
    part might be over the only goal that
  • 00:22:45
    matters to them from now on is to keep
  • 00:22:47
    generating better data for the next
  • 00:22:49
    generation of models from here on the
  • 00:22:52
    base models May operate in the
  • 00:22:53
    background empowering other models to
  • 00:22:55
    achieve Feats that they couldn't on
  • 00:22:56
    their own like an old hermit pass down
  • 00:22:58
    wisdom from a secret Mountain cave
  • 00:23:00
    except the that cave is a massive Data
  • 00:23:02
    Center and whether or not we meet him or
  • 00:23:04
    not is you know something we're going to
  • 00:23:05
    have to see so it's quite likely that
  • 00:23:07
    maybe we're going to have these internal
  • 00:23:09
    models gbt 6 gb7 gb5 producing the
  • 00:23:12
    synthetic data for these future models
  • 00:23:14
    to be trained on it could definitely be
  • 00:23:15
    the case considering the fact that it's
  • 00:23:17
    quite likely that we might not get these
  • 00:23:19
    models and so what about if you know gbt
  • 00:23:21
    5 suddenly gets released they basically
  • 00:23:22
    say that even if gbt 5 is eventually
  • 00:23:24
    released opening eye and the Tropic have
  • 00:23:26
    already initiated the operation of curse
  • 00:23:28
    of self-improvement with humans in the
  • 00:23:30
    loop and it doesn't really matter what
  • 00:23:32
    they give us publicly they're going to
  • 00:23:33
    be pulling further and further ahead
  • 00:23:35
    like the universe expanding so fast that
  • 00:23:37
    distant galaxies can no longer reach us
  • 00:23:39
    and that's probably how they jumped from
  • 00:23:41
    01 to 03 in barely 3 months and that's
  • 00:23:44
    how they're going to jump to 04 to 05 so
  • 00:23:47
    it's probably why they've been so
  • 00:23:48
    excited on social media because they've
  • 00:23:50
    implemented a new way to scale
  • 00:23:53
    incredibly so you can see right here
  • 00:23:54
    they actually talk about something that
  • 00:23:55
    I actually spoke about quite a long time
  • 00:23:57
    before like ages ago in I first launched
  • 00:23:58
    my ever AI Community I spoke up about
  • 00:24:00
    the fact that you know even if AGI does
  • 00:24:02
    arrive we probably won't get access to
  • 00:24:04
    it because the economic value for the
  • 00:24:05
    average person just simply doesn't make
  • 00:24:07
    sense and you can see right here that
  • 00:24:08
    they state did you really think
  • 00:24:09
    approaching AGI would mean gaining
  • 00:24:12
    access to increasingly powerful AI at
  • 00:24:13
    your fingertips that they'd release
  • 00:24:15
    every advancement for us to use surely
  • 00:24:17
    you don't believe that they meant it
  • 00:24:18
    when they said their models would push
  • 00:24:20
    them too far ahead for anyone else to
  • 00:24:22
    catch up and each new generation model
  • 00:24:24
    is an engine of escape velocity from the
  • 00:24:26
    stratosphere they're already w goodbye
  • 00:24:28
    and they're basically saying that look
  • 00:24:30
    with every time that they make a new
  • 00:24:31
    model it's going to become harder and
  • 00:24:33
    harder to catch up to open a ey because
  • 00:24:34
    they have something else that can
  • 00:24:36
    generate synthetic data and that can
  • 00:24:38
    also help them further the entire cycle
  • 00:24:40
    of increasing intelligence within those
  • 00:24:42
    models now recently there was also this
  • 00:24:43
    very cryptic tweet that has been going
  • 00:24:45
    viral on Twitter and it says just got
  • 00:24:47
    reading to some info what's happening
  • 00:24:48
    globally happening internally at openi
  • 00:24:51
    and holy mother of God I don't even know
  • 00:24:53
    how to express my feelings without
  • 00:24:54
    sounding like hype but I don't know what
  • 00:24:57
    to say but I will share this the
  • 00:24:59
    innovators are coming the problem is we
  • 00:25:01
    don't know how they got there now I will
  • 00:25:02
    say this isn't like a Jimmy apples kind
  • 00:25:04
    of post I haven't really seen any people
  • 00:25:05
    coating this statement but not just this
  • 00:25:08
    person has been stating this thing I've
  • 00:25:09
    been seeing time and time again from
  • 00:25:10
    people at open AI stating crazy things
  • 00:25:12
    about super intelligence open AI on
  • 00:25:14
    their blog have said that you know
  • 00:25:15
    they're now chasing artificial super
  • 00:25:17
    intelligence instead of AGI and they
  • 00:25:18
    know exactly how to get to AGI so all of
  • 00:25:20
    these statements coming around the same
  • 00:25:22
    time at this new paradigm isn't honestly
  • 00:25:24
    surprising and as I was making this
  • 00:25:25
    video there was actually a tweet about
  • 00:25:27
    GPT 5 you can see someone said can you
  • 00:25:29
    comment something about GPT 5 we know
  • 00:25:31
    you won't be able to say anything just
  • 00:25:33
    anything at all and he respond saying
  • 00:25:35
    what would you like to know and chubby
  • 00:25:37
    comments back saying when any time in
  • 00:25:39
    terms of the estimate of time arrival
  • 00:25:41
    and of course the performance how much
  • 00:25:42
    better is this going to be than gbt 40
  • 00:25:44
    and will this GPT series merge with the
  • 00:25:46
    O Series and he says he's still figuring
  • 00:25:48
    out when the estimated type of arrival
  • 00:25:49
    is and of course the performance and in
  • 00:25:51
    2025 they actually talk about merging
  • 00:25:53
    the 01 series and the GPT Series so
  • 00:25:55
    overall still a very vague response not
  • 00:25:57
    much to for there but it definitely does
  • 00:26:00
    make sense considering the fact that
  • 00:26:02
    they did create these models they were
  • 00:26:03
    going to make them anyways and of course
  • 00:26:05
    we know that they haven't released them
  • 00:26:07
    so it's quite likely that they are using
  • 00:26:08
    them internally to generate data and do
  • 00:26:10
    many other things just imagine a version
  • 00:26:12
    of Claude 3.6 that's even better than it
  • 00:26:14
    is now imagine what they could be using
  • 00:26:16
    that for and we've seen literally how
  • 00:26:18
    much Claude 3.6 Sonic has changed entire
  • 00:26:21
    Industries in terms of like cursor and
  • 00:26:23
    coding and what people are able to do
  • 00:26:24
    with that so it's super super
  • 00:26:25
    interesting to see where things go from
  • 00:26:27
    here with that being said hopefully you
  • 00:26:29
    guys enjoyed this video and I'll see you
  • 00:26:30
    in the next one
タグ
  • GPT-5
  • OpenAI
  • AI development
  • model distillation
  • Anthropic
  • Claude Opus 3.5
  • AI industry
  • internal models
  • technology trends
  • cutting-edge AI