How China's DeepSeek upends the AI status quo

00:11:08
https://www.youtube.com/watch?v=3T0lg8cWeEc

Resumo

TLDR视频探讨了DeepSeek——一种中国开发的低成本AI模型——在当前AI市场中的重要性。专家们分析了该模型的成功如何影响美国AI公司的战略和市场反应，同时讨论了美国的芯片出口禁令和对其是否有效的辩论。随着AI行业的演变，许多参与者对未来的创新机会表示关注，并认为开放源代码模型可能是未来的关键。此外，视频显示了对AI行业持续投资和技术进步的迫切需要，分析了市场波动的原因与风险。

Conclusões

  • 🧠 DeepSeek是一个中国的低成本AI模型。
  • 💸 许多投资者考虑使用DeepSeek的技术，因为其成本低。
  • ⚖️ 美国市场面对AI技术的重大变化与挑战。
  • 🌍开放源代码模型可能是AI发展的未来方向。
  • 📉 AI行业的市场波动性显著，需谨慎对待投资。
  • 📊 专家们认为技术创新是保持竞争力的关键。
  • 📉DeepSeek使用Nvidia的低端芯片，打破了预期。
  • 🚀 投资者需适应快速变化的AI市场与新的商业模式。
  • 💡 创新需要大量资金投入，以维持技术领先地位。
  • 🛡️ 政策对芯片出口的调整可能影响美国AI公司的竞争。

Linha do tempo

  • 00:00:00 - 00:05:00

    此次讨论聚焦于中国低成本的AI模型，即Deep Seek的崛起和其对美国AI市场的冲击。与会者强调，美国在AI创新方面的滞后，市场对Deep Seek的反应和投资者的信心不足，表明人们对当前市场状况的担忧。同时，Deep Seek的成功催生了关于AI模型及其应用的广泛讨论，尤其是开放源代码与封闭源代码的竞争。

  • 00:05:00 - 00:11:08

    随着Deep Seek的成功，业界开始质疑美国公司在AI领域的高投入是否真正有效。讨论中提到，虽然美国巨头仍会在模型开发上持续投入，但Deep Seek所展示的低成本、高效率的能力，使得竞争格局发生变化。各方专家一致认为，未来AI技术的发展将更加迅速和多样化，对现有市场产生重大影响。

Mapa mental

Vídeo de perguntas e respostas

  • DeepSeek是什么？

    DeepSeek是一种低成本的AI模型，由中国公司研发，旨在提供高效的AI解决方案。

  • 美国对AI领域的最新反应是什么？

    美国市场对于DeepSeek的影响有不安，许多投资者考虑使用其低成本的AI技术。

  • AI技术未来的发展方向是什么？

    未来的发展可能会偏向于开放源代码的AI模型，同时还需要对当前的训练和推理技术进行创新。

  • DeepSeek如何与Nvidia芯片相关？

    DeepSeek使用了Nvidia的低端芯片h800，其性能表现引发了关于美国尖端芯片是否应该继续出口的讨论。

  • 为何AI市场的波动性如此大？

    AI市场在快速变化的创新环境中，投资者对公司股票的价值判断存在不确定性和风险。

Ver mais resumos de vídeos

Obtenha acesso instantâneo a resumos gratuitos de vídeos do YouTube com tecnologia de IA!
Legendas
en
Rolagem automática:
  • 00:00:00
    let's dive in here to the real
  • 00:00:01
    significance of China's quote lowcost AI
  • 00:00:03
    model this deep seek R1 is the selloff
  • 00:00:06
    across those names he mentioned deserved
  • 00:00:08
    Raymond James Tavis mcord is here to
  • 00:00:10
    look at that for us today VC Bradley
  • 00:00:12
    Tusk on why it's going to be hard not to
  • 00:00:14
    use deep seek cyber security CEO Ivan
  • 00:00:17
    serini is back to talk about the hack
  • 00:00:19
    and how deep seat got so good so quickly
  • 00:00:21
    and dear jaosa who was all over the
  • 00:00:23
    story before everyone else welcome to
  • 00:00:25
    all of you Bradley I I just thought it
  • 00:00:27
    was interesting what you said about
  • 00:00:29
    people using this model I by the way I
  • 00:00:31
    tried I tried finally to download it
  • 00:00:33
    this morning and couldn't maybe that's
  • 00:00:35
    because of the hack but what's your
  • 00:00:36
    experience Ben and and what do you think
  • 00:00:39
    U you know the takeup ultimately will be
  • 00:00:41
    here in the US of this Chinese offering
  • 00:00:43
    yeah I mean look as an early stage
  • 00:00:45
    Venture investor there's no such thing
  • 00:00:47
    as investing in a company that isn't
  • 00:00:49
    either an AI company or using AI in some
  • 00:00:51
    way or at least has an extremely good
  • 00:00:53
    reason if they're not but if they're
  • 00:00:55
    using AI in some way then the cost
  • 00:00:58
    obviously becomes very material and
  • 00:01:00
    really important and using open AI or
  • 00:01:02
    any of these other companies is really
  • 00:01:04
    really expensive and yes I understand
  • 00:01:07
    the Deep SE because a Chinese company
  • 00:01:08
    but if any of my portfolio companies
  • 00:01:10
    said listen um we can build what we need
  • 00:01:12
    at a fraction of the cost on an open
  • 00:01:14
    source system it's really hard as
  • 00:01:16
    fiduciary to tell them no you shouldn't
  • 00:01:18
    do that and so I think the problem isn't
  • 00:01:20
    just that we have uh essentially
  • 00:01:22
    regulatory to trade problem and the
  • 00:01:24
    problem is we got out innovated um every
  • 00:01:27
    those sort of this group think in the US
  • 00:01:28
    where all of the companies approach this
  • 00:01:30
    thing in exact same way and it really
  • 00:01:32
    was all about sort of scale and cost and
  • 00:01:36
    you know like Alex mik scale AI has made
  • 00:01:38
    some really good points about this which
  • 00:01:39
    is we should be a lot more aggressive
  • 00:01:41
    Innovation we've kind of lost the
  • 00:01:42
    Innovation war on this and we should ask
  • 00:01:45
    you know why is it that no one in
  • 00:01:47
    America thought to do this well I think
  • 00:01:50
    people would probably say hey we're
  • 00:01:51
    working on it like that's the next phase
  • 00:01:53
    of all this and there's a whole shift in
  • 00:01:54
    here from kind of training to
  • 00:01:56
    inferencing which is an important part
  • 00:01:57
    of the story and makes it interesting
  • 00:01:58
    that broad comes down but before we get
  • 00:02:00
    or maybe I should turn to Tavis uh to
  • 00:02:02
    talk more about that we saw 15 20%
  • 00:02:05
    declines in some of these exposed names
  • 00:02:07
    do you think that's warranted given what
  • 00:02:09
    we've learned here about what deep seeks
  • 00:02:10
    new model can do or do you think this is
  • 00:02:13
    overdone well if it's just deep seek
  • 00:02:16
    probably not but I think the reality is
  • 00:02:17
    what we're seeing is capitalism at work
  • 00:02:20
    and and nobody knows exactly how AI is
  • 00:02:21
    going to play out but there's a lot of
  • 00:02:23
    money that's gone into the space and
  • 00:02:25
    what that means there going to be a lot
  • 00:02:26
    of innovation and some of that
  • 00:02:27
    Innovation isn't necessarily going to be
  • 00:02:29
    great for profits of certain companies
  • 00:02:31
    and so um you know when I when I see
  • 00:02:34
    some of these names down 15 20% it just
  • 00:02:36
    screams to me that there's very little
  • 00:02:38
    conviction out there so even though
  • 00:02:40
    these names have been straight up for
  • 00:02:42
    the last two years people are buying
  • 00:02:44
    charts not businesses and um that's
  • 00:02:47
    always a dangerous situation that said
  • 00:02:49
    it's not like they're selling
  • 00:02:50
    indiscriminately Meadow's higher today
  • 00:02:51
    Apple's higher both of them are
  • 00:02:53
    beneficiaries of lower cost inferencing
  • 00:02:55
    models right they are the ones who are
  • 00:02:57
    kind of deploying this technology if
  • 00:02:59
    it's less expensive to make it even if
  • 00:03:00
    China got there first we're going to get
  • 00:03:02
    there eventually as well so there is a
  • 00:03:04
    logic tab as to what the market is doing
  • 00:03:06
    here yeah I would agree with that I
  • 00:03:08
    think in in general if you had to place
  • 00:03:10
    a bet is the uh is the real profit pool
  • 00:03:14
    going to be on chips or models or
  • 00:03:17
    applications that use those models like
  • 00:03:19
    today we're shifting over towards the
  • 00:03:21
    application front um if if AI is can be
  • 00:03:24
    a lot cheaper to deploy than than uh
  • 00:03:26
    than than most would have thought a
  • 00:03:27
    couple months ago that makes all the
  • 00:03:28
    sense in the world yeah dear J let's
  • 00:03:30
    talk for a second about kind of how they
  • 00:03:32
    got here and one of the big
  • 00:03:33
    controversies has been did they do it
  • 00:03:35
    with these inferior Nvidia chips these
  • 00:03:38
    h800 which one's a better one h100 so
  • 00:03:40
    they did it with h800 let's just take it
  • 00:03:42
    at face value because analyst say look
  • 00:03:44
    it's possible that they did this and
  • 00:03:46
    that the export controls Force this
  • 00:03:47
    level of innovation so thank you deep
  • 00:03:49
    seek for showing the rest of the world
  • 00:03:50
    it's possible the question now is do we
  • 00:03:53
    keep with the export ban on these
  • 00:03:55
    Leading Edge chips or not and in fact
  • 00:03:57
    David saxs that what they call the
  • 00:03:58
    crypto a the AI just weighed in on this
  • 00:04:01
    on Twitter he says this shows that the
  • 00:04:03
    AI race will be very competitive and
  • 00:04:05
    that President Trump was right to resend
  • 00:04:07
    Biden's EO which hamrun American AI
  • 00:04:09
    companies without asking whether China
  • 00:04:11
    would do the same it didn't he he says
  • 00:04:13
    I'm confident in the US we can't be
  • 00:04:14
    complacent there's a whole fight in here
  • 00:04:16
    about open source AI models and whether
  • 00:04:19
    we should be kind of going in the
  • 00:04:20
    direction that open AI would prefer
  • 00:04:21
    versus what deep seek and others are
  • 00:04:23
    doing kind of leaving this technology
  • 00:04:24
    more open to the public so um how do you
  • 00:04:26
    think policy-wise this is all going to
  • 00:04:28
    play out well let's be clear the export
  • 00:04:31
    band did not work in fact it backfired
  • 00:04:33
    it forced the Chinese to innovate in new
  • 00:04:35
    ways that companies and Giants here
  • 00:04:38
    didn't have to I call them the good
  • 00:04:40
    chips and the dumb down chips the good
  • 00:04:41
    ones are the h100s that's the Nvidia
  • 00:04:44
    gpus that everyone wants NV 800s are the
  • 00:04:47
    ones that the Chinese get because of the
  • 00:04:49
    export ban and what deep seek says it
  • 00:04:51
    was able to do is create a model that
  • 00:04:53
    outperforms some of our best models
  • 00:04:56
    built here in America with those dumb
  • 00:04:58
    down chips I mean that is a point of
  • 00:05:00
    debate some people say that the hedge
  • 00:05:02
    fund of which deep seek was born out of
  • 00:05:05
    had been sort of hoarding the better
  • 00:05:07
    chips um but at the end of the day Kelly
  • 00:05:09
    it doesn't really matter we've moved on
  • 00:05:11
    past that we know now what deep seek
  • 00:05:13
    proved is that these models and open
  • 00:05:16
    source nonetheless I want to stress how
  • 00:05:18
    important it is that deep See's model is
  • 00:05:20
    open source can be built for a fraction
  • 00:05:23
    of the price whether that's 6 million or
  • 00:05:25
    60 million that is far less than the
  • 00:05:27
    hundreds of millions or the billions of
  • 00:05:29
    dollars that open AI Gemini um Google
  • 00:05:33
    and others have poured into this so I
  • 00:05:34
    think that's also maybe why you're
  • 00:05:36
    seeing meta um in the positive today
  • 00:05:38
    because it's also working on open source
  • 00:05:40
    metas in fact um I can't remember who
  • 00:05:42
    was one of the people we spoke to in the
  • 00:05:44
    valley on this said that meta is
  • 00:05:46
    probably racing right now to figure out
  • 00:05:49
    what deep SE did and build on top of it
  • 00:05:51
    and so there's a new race that started
  • 00:05:53
    in open- Source development which is
  • 00:05:56
    really key there and that's why the
  • 00:05:58
    companies that were working on Clos
  • 00:05:59
    Source models like an open AI real
  • 00:06:01
    questions about their MO and Tavis this
  • 00:06:03
    all gets into this big pivot that's
  • 00:06:05
    happening in the AI space from the
  • 00:06:07
    massive clusters that were required to
  • 00:06:09
    train these models to the ones that can
  • 00:06:11
    do more inferencing maybe have a little
  • 00:06:13
    bit lighter that's why you're seeing the
  • 00:06:14
    power stocks down you know yes early
  • 00:06:16
    last week it was all about Stargate and
  • 00:06:17
    it's going to have a million gpus and
  • 00:06:19
    it's going to require you know nuclear
  • 00:06:20
    power and now it's about okay and even
  • 00:06:23
    if it were 5 to six million to kind of
  • 00:06:25
    make this shift to inferencing do you
  • 00:06:26
    think that's enough for the likes of
  • 00:06:28
    broadcom which should in a way benefit
  • 00:06:30
    from that transition to be down 18%
  • 00:06:33
    today uh I don't it seems extreme but
  • 00:06:36
    like look 3 weeks from now we could be
  • 00:06:37
    talking about something completely
  • 00:06:38
    different innovation's happening really
  • 00:06:40
    fast it's really hard to predict in what
  • 00:06:42
    ways it's going to happen and what
  • 00:06:43
    company's going to benefit on a specific
  • 00:06:45
    form of innovation I think all we know
  • 00:06:48
    is that you know we've this Market has
  • 00:06:50
    has thrown about 10 trillion dollars uh
  • 00:06:53
    of market cap into the AI theme and um
  • 00:06:58
    it's going to be really volatile for
  • 00:06:59
    some time that there's there's no set in
  • 00:07:01
    stone way AI is going to get to Market
  • 00:07:04
    and to know to to think we know who the
  • 00:07:05
    winners and losers are going to be at
  • 00:07:07
    this point uh or even to the degree
  • 00:07:09
    they're public is is is somewhat um
  • 00:07:12
    unknowable yeah there's a look at some
  • 00:07:14
    of the biggest decliners all in the
  • 00:07:15
    energy space again the the people who we
  • 00:07:17
    thought would need kind of the near-term
  • 00:07:20
    um benef amount of spend to support all
  • 00:07:22
    of this maybe don't Sam leson we talked
  • 00:07:24
    to a few different people about this
  • 00:07:25
    this morning he said $6 million in
  • 00:07:26
    training has now destroyed half a
  • 00:07:28
    trillion in mem driven market cap I
  • 00:07:30
    don't know if it's fair to call it meme
  • 00:07:31
    driven but in two business days Roger
  • 00:07:33
    momy said look it's obvious the Chinese
  • 00:07:35
    approach is going to produce more value
  • 00:07:37
    um it's not as wasteful he says it
  • 00:07:39
    exposes the ludicrous approach that
  • 00:07:41
    American companies have taken to Ai and
  • 00:07:43
    calls the path unsustainable Bradley do
  • 00:07:45
    you
  • 00:07:46
    agree yeah I mean ultimately I think D
  • 00:07:49
    got this right if open source is here
  • 00:07:50
    and it's what the market wants and it's
  • 00:07:52
    hard to see how that wouldn't be you
  • 00:07:54
    know we have to able to adapt to that so
  • 00:07:56
    in part that's what I think David sax
  • 00:07:57
    was getting at in his in his tweet which
  • 00:07:59
    was look we're going to have to both
  • 00:08:01
    move past the Biden executive order and
  • 00:08:03
    possibly allow open source and the
  • 00:08:06
    American companies are going to have to
  • 00:08:07
    adapt really quickly um but the other
  • 00:08:09
    thing is you know it always feels like
  • 00:08:11
    human beings have this tendency to think
  • 00:08:13
    whatever is right now is what will
  • 00:08:15
    always be and that's never the case
  • 00:08:17
    right it's really not the case with AI
  • 00:08:19
    but I remember you know coming on CNBC
  • 00:08:20
    multiple times in the past year where
  • 00:08:23
    everyone was super excited about Nvidia
  • 00:08:25
    and nothing can ever stop them and all
  • 00:08:26
    of a sudden $6 million just did so uh I
  • 00:08:29
    just really think it's important for
  • 00:08:30
    people to be mindful here that we are in
  • 00:08:32
    the early stages of something totally
  • 00:08:34
    disruptive and transformative and
  • 00:08:36
    whatever you're doing you know be
  • 00:08:38
    prepared to PFF at it in I was listening
  • 00:08:40
    to one of the emergency podcasts from
  • 00:08:41
    the weekend it was one with Miles
  • 00:08:42
    Brundage Bradley and you know he talked
  • 00:08:45
    about how we we shouldn't assume that
  • 00:08:47
    just because these kind of inferencing
  • 00:08:49
    models can be done with lighter compute
  • 00:08:50
    that that people don't still have a need
  • 00:08:52
    for the chips he likened it to creating
  • 00:08:54
    Einstein and said don't wouldn't you
  • 00:08:56
    rather have a thousand Einstein than
  • 00:08:58
    just one I mean wouldn't you rather even
  • 00:08:59
    deploy multiple you know types of this
  • 00:09:02
    technology at the same time in order to
  • 00:09:04
    get it whether it's a homework problem
  • 00:09:06
    or you know a cyber security problem or
  • 00:09:07
    whatever it is so that's why I find this
  • 00:09:10
    this issue around Nvidia okay maybe in
  • 00:09:12
    the very very near term you could call
  • 00:09:14
    some of the financials into question but
  • 00:09:16
    I'm sure deep seek I'm sure across China
  • 00:09:18
    I'm sure everybody would rather have
  • 00:09:19
    access to more Nvidia gpus at this point
  • 00:09:21
    than fewer yeah that that's absolutely
  • 00:09:24
    true but I think the problem is the
  • 00:09:27
    disparity we're talking about is not6
  • 00:09:29
    million versus $60 million it's $6
  • 00:09:31
    million versus you know $500 billion
  • 00:09:34
    that Stargate was talking about you know
  • 00:09:36
    Trump the other day the disparity is so
  • 00:09:38
    massive it's not the whole number though
  • 00:09:40
    6 million was only the amount that was
  • 00:09:42
    just the last the last that's an
  • 00:09:44
    important caveat you're totally right
  • 00:09:45
    you're to but but overall I just think
  • 00:09:47
    that you know yes but because this world
  • 00:09:51
    because the cost of compute is has been
  • 00:09:53
    so expensive and the cost of energy for
  • 00:09:55
    this has been so expensive that I think
  • 00:09:57
    that just creates massive demand for
  • 00:09:59
    some cheaper but yeah your point right
  • 00:10:02
    we don't all we know is where we are
  • 00:10:03
    today and it's going to evolve in 20
  • 00:10:05
    different ways we can't even think of
  • 00:10:06
    right now dear final word yeah on this
  • 00:10:09
    point what I hear from people in Tech is
  • 00:10:11
    that you're not going to see any kind of
  • 00:10:13
    slowdown from open AI or Google or
  • 00:10:16
    Microsoft the companies that want to
  • 00:10:18
    build the best models in America because
  • 00:10:20
    now it's just all the more existential
  • 00:10:22
    for them I mean what deep seek showed is
  • 00:10:24
    you can get right to the frontier and
  • 00:10:26
    build something on that and innovate but
  • 00:10:28
    if you actually make more advances in
  • 00:10:30
    the pre-training model which has become
  • 00:10:32
    a lot harder to do then the US is back
  • 00:10:35
    on top so maybe that drive um is just is
  • 00:10:39
    even more urgent let's say and you do
  • 00:10:40
    need to spend all that money half a
  • 00:10:43
    trillion dollars hundreds of millions in
  • 00:10:44
    capex maybe you do need to do that just
  • 00:10:46
    to keep the US ahead I mean I think now
  • 00:10:49
    it's more uncertain but the stakes have
  • 00:10:51
    risen I think that's exactly right I
  • 00:10:52
    appreciate you all joining us today
  • 00:10:54
    Bradley Tusk Tavis morda Raymond James
  • 00:10:56
    our dear drosa of course and remember
  • 00:10:58
    you can watch dearra full deep dive into
  • 00:11:00
    deep seek and China's AI Breakthrough by
  • 00:11:02
    scanning that QR code on your screen or
  • 00:11:04
    go to cnbc.com TCT takes we said she was
  • 00:11:07
    all over it
Etiquetas
  • DeepSeek
  • AI技术
  • 低成本模型
  • 投资
  • 开放源代码
  • 中国
  • Nvidia芯片
  • 市场波动
  • 创新
  • 推理技术