GPT-4.5 shocks the world with its lack of intelligence...

00:04:16
https://www.youtube.com/watch?v=FW2XOIxaNqg

Resumo

TLDRThe video critiques the release of OpenAI's GPT 4.5, labeling it as expensive and underwhelming compared to previous models. Despite its claims for natural conversation and a new Vibes Benchmark for creative thinking, it lacks significant improvements, performing poorly in tasks like programming. The host expresses disappointment in the current trajectory of AI development and questions the potential for future advancements, while promoting educational tools for those interested in programming and computer science.

Conclusões

  • 💰 GPT 4.5 is the most expensive model yet at $75 per million input tokens.
  • 🛑 It fails to outperform existing benchmarks or introduce novel features.
  • 🌊 The model aims to create 'Chill Vibes' in conversations but lacks substantial improvements.
  • 👩‍💻 Despite lower hallucination rates, it still makes many silly mistakes.
  • 🤖 Comparison reveals it performs worse in programming tasks than previous models.
  • 🚨 Concerns are raised about the future capabilities of GPT 5.
  • 📉 OpenAI's valuation may decline without significant advancements.
  • 📚 Viewers are encouraged to enhance their programming skills with resources like Brilliant.

Linha do tempo

  • 00:00:00 - 00:04:16

    The excitement around AI has recently diminished following the release of GPT-4.5, which, despite being the most expensive AI model, lacks substantial improvements or novel features. OpenAI's launch was criticized for its lack of engagement, indicated by the absence of CEO Sam Altman at the event. GPT-4.5 is notably more expensive than its predecessors, costing $75 per million input tokens and $150 for output tokens, and is only accessible to Pro users. While it was claimed to have a 'Vibes Benchmark' for measuring creative thinking, user experiences showed it still made basic errors and has a lower performance in coding benchmarks compared to more established models. In light of these issues, there is speculation about OpenAI's ability to maintain its industry lead as they transition to a for-profit model amidst rising competition. The discussion also touched upon the educational opportunities for programmers as AI tools evolve, paired with the promotion of online learning platforms like Brilliant, aimed at empowering users to understand deep learning and AI fundamentals.

Mapa mental

Vídeo de perguntas e respostas

  • Why is GPT 4.5 considered underwhelming?

    It fails to surpass benchmarks, lacks novel capabilities, and is expensive without significant improvements.

  • How much does GPT 4.5 cost?

    It costs $75 per million input tokens and $150 per million output tokens, available to Pro users for $200/month.

  • What are the new features of GPT 4.5?

    It has a 'Vibes Benchmark' aimed at measuring creative thinking and claims to emit 'Chill Vibes' during interactions.

  • How does GPT 4.5 compare to other models in programming?

    It is worse at programming than previous deep learning models and is significantly more expensive.

  • What are the expectations for GPT 5?

    There are concerns it may not show significant improvements and might operate more like a router selecting models based on prompts.

  • What recommendations are given for viewers?

    Viewers are encouraged to enhance their programming skills using educational platforms like Brilliant.

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
    it's official the AI hype train just
  • 00:00:01
    went on life support with the
  • 00:00:03
    underwhelming release of GPT 4.5
  • 00:00:06
    yesterday open AI unveiled the most
  • 00:00:08
    expensive AI model ever produced yet it
  • 00:00:10
    fails to crush any benchmarks win any
  • 00:00:12
    awards or offer any novel capabilities
  • 00:00:14
    whatsoever its only real selling point
  • 00:00:16
    is Vibes and is supposed to chat in a
  • 00:00:18
    more natural human-like way don't get me
  • 00:00:20
    wrong it's a good model but not good
  • 00:00:22
    enough to feed the AI hype monster and
  • 00:00:24
    it looks increasingly likely that we're
  • 00:00:26
    not headed into a technological
  • 00:00:28
    singularity but rather a sigmoid of
  • 00:00:30
    sorrow the Sam ultman couldn't even be
  • 00:00:31
    bothered to leave his newborn kid in the
  • 00:00:33
    hospital to show up to the product
  • 00:00:34
    launch and instead send in a bunch of
  • 00:00:36
    interns to demo it and that's crazy
  • 00:00:38
    because we're talking about Orion here
  • 00:00:40
    in 2023 Tech leaders signed a petition
  • 00:00:42
    to stop training big models like this
  • 00:00:44
    mman himself begged the government to
  • 00:00:45
    regulate it and the only thing more
  • 00:00:47
    disappointing than GPT 4.5 is the
  • 00:00:49
    release of the Epstein files in today's
  • 00:00:51
    video we'll find out if we just reach
  • 00:00:52
    the limits of pre-training in generative
  • 00:00:54
    pre-trained Transformers it is February
  • 00:00:56
    28th 2025 and you're watching the code
  • 00:00:59
    report I didn't want to make another
  • 00:01:00
    crappy AI video today but the bat signal
  • 00:01:03
    was triggered anytime an official video
  • 00:01:04
    gets ratioed like this I have no choice
  • 00:01:07
    but to make a video before you
  • 00:01:08
    unsubscribe though I've got an
  • 00:01:09
    interesting postgress video on the way
  • 00:01:11
    the first thing to know about GPT 4.5 is
  • 00:01:13
    that it's extremely expensive if you
  • 00:01:15
    thought Claude was expensive at $15 per
  • 00:01:17
    million tokens GPT 4.5 is five times
  • 00:01:20
    more expensive at $75 per million output
  • 00:01:23
    tokens actually no correction that's
  • 00:01:25
    input tokens it's $150 per million
  • 00:01:28
    output tokens and to chat with it it's
  • 00:01:29
    it's currently only available to the
  • 00:01:31
    $200 per month Pro users I tried it out
  • 00:01:33
    myself and it does seem to emit Chill
  • 00:01:35
    Vibes but the problem is that's highly
  • 00:01:37
    subjective however in the launch open
  • 00:01:39
    aai talked about a new Vibes Benchmark
  • 00:01:42
    that's supposed to measure creative
  • 00:01:43
    thinking the best way to get a f for the
  • 00:01:45
    model is to talk to it so let's jump
  • 00:01:47
    into a demo a lot of people on the
  • 00:01:48
    internet criticize this presentation but
  • 00:01:50
    as an introvert myself I think they did
  • 00:01:52
    a great job in addition it apparently
  • 00:01:53
    has a far lower hallucination rate but
  • 00:01:55
    what I found is that it still makes a
  • 00:01:57
    lot of silly mistakes it's not
  • 00:01:58
    self-aware and has has no idea what gbt
  • 00:02:00
    4.5 even is and says its training cut
  • 00:02:03
    off is October 2023 it was however able
  • 00:02:06
    to tell me how many RS are in Strawberry
  • 00:02:08
    that felt like a huge leap forward but I
  • 00:02:10
    quickly became disappointed when it gave
  • 00:02:11
    me the wrong number of L's in laap paloa
  • 00:02:14
    now when it comes to programming and
  • 00:02:15
    science I didn't even try because we
  • 00:02:17
    already know it's not going to perform
  • 00:02:18
    as well as the deep thinking models like
  • 00:02:20
    03 then to make matters worse on the AER
  • 00:02:22
    polyglot coding Benchmark it's not only
  • 00:02:24
    worse at programming than deep seek but
  • 00:02:26
    also hundreds of times more expensive
  • 00:02:28
    now if you're an Elon Musk hater you'll
  • 00:02:29
    want take a bong rip of copium right now
  • 00:02:31
    because currently xai's Gro is the best
  • 00:02:33
    model in the world that's not my opinion
  • 00:02:35
    it's the opinion of the betting Market
  • 00:02:37
    although by the end of 2025 open AI is
  • 00:02:39
    still the favorite to have the best
  • 00:02:41
    model but its odds are on the decline
  • 00:02:43
    that's problematic for open AI though
  • 00:02:44
    because they're raising billions and
  • 00:02:46
    billions of dollars as they transition
  • 00:02:47
    to for-profit and will need to maintain
  • 00:02:49
    a massive valuation Alman says there is
  • 00:02:51
    no wall and believes they can scale
  • 00:02:53
    these models almost infinitely that's
  • 00:02:55
    assuming he gets trillions of dollars
  • 00:02:56
    from soft Bank in the Saudis to build
  • 00:02:58
    these data centers my theory as an
  • 00:03:00
    unqualified ship poster is that they
  • 00:03:01
    failed to train GPT 5 with any
  • 00:03:03
    significant Improvement despite scaling
  • 00:03:05
    up the number of parameters in compute
  • 00:03:07
    GPT 4.5 is the biggest model they've
  • 00:03:09
    ever created and now they're lowering
  • 00:03:10
    the bar for gbt 5 which ultman described
  • 00:03:13
    a few weeks ago being more like a router
  • 00:03:15
    that automatically chooses the best
  • 00:03:16
    model based on your prompt and that's
  • 00:03:17
    highly disappointing because I was
  • 00:03:19
    expecting to be a post-apocalyptic
  • 00:03:21
    warlord by now the battling robots and
  • 00:03:23
    barbecuing rats over burning garbage
  • 00:03:24
    cans for dinner but instead I live in
  • 00:03:26
    this dystopia where artificial super
  • 00:03:28
    intelligence never comes and nothing
  • 00:03:30
    ever happens but if you're a computer
  • 00:03:31
    science student the plateau is great
  • 00:03:33
    news AI coding tools are incredible but
  • 00:03:35
    they're most useful to real human
  • 00:03:37
    programmers who know what they're doing
  • 00:03:38
    and I don't see that changing anytime
  • 00:03:40
    soon and you can start getting really
  • 00:03:41
    good at programming for free thanks to
  • 00:03:43
    this video sponsor brilliant their
  • 00:03:45
    platform provides interactive Hands-On
  • 00:03:47
    lessons that demystify the complexity of
  • 00:03:49
    deep learning with just a few minutes of
  • 00:03:51
    effort each day you can understand the
  • 00:03:53
    math and computer science behind this
  • 00:03:55
    seemingly magic technology I'd recommend
  • 00:03:57
    starting with python then check out
  • 00:03:59
    their full how large language models
  • 00:04:01
    work course if you really want to look
  • 00:04:02
    under the hood of chat gbt try
  • 00:04:04
    everything brilliant has to offer for
  • 00:04:06
    free for 30 days by going to
  • 00:04:08
    brilliant.org fireship or use the QR
  • 00:04:11
    code on screen this has been the code
  • 00:04:12
    report thanks for watching and I will
  • 00:04:14
    see you in the next one
Etiquetas
  • GPT 4.5
  • OpenAI
  • AI development
  • programming
  • vibes benchmark
  • technology critique
  • AI performance
  • deep learning
  • cost
  • future of AI