OpenAI GPT-4.5 in 12 Minutes

--- type: transcript date: 2025-02-27 youtube_id: PtKUWGGDMpA --- # Transcript: OpenAI GPT-4.5 in 12 Minutes open AI has just released GPT 4.5 which is a research preview of their strongest GPT model yet as of today it's available to Pro users as well as developers worldwide first up they describe this model as our largest and best model for chat GPD 4.5 is a step forward in scaling up pre-training and posttraining by scaling up unsupervised learning GPD 4.5 improves its ability to recognize patterns draw connections and generate creative insights without reasoning that's one of the key things with this model is this is not a reasoning model they mentioned that early testing shows that interacting with 4.5 feels more natural it's broader knowledge based improved ability to follow user intent and greater EQ makes it useful for tasks like improving writing programming and solving practical problems we also expected to hallucinate less now I'll also touch on this when I get into the model card section next they mentioned that they're sharing GPD 4.5 as a research preview to better understand its strengths and limitations we're still exploring what it's capable of and are eager to see how people use it in ways we might not have expected they described that they're advancing AI capabilities by scaling two complimentary paradigms unsupervised learning and reasoning these represent two of intelligence as they described they're scaling reasoning which teaches the model to think and produce a chain of thought before they respond allowing them to tackle complex step or logical problems models like 01 03 mini so on and so forth whereas with this model the unsupervised learning on the other hand increases World model accuracy as well as intuition they describe that GPD 4.5 is an example of scaling unsupervised learning by scaling up compute and data alongside with the architecture and optimization Innovations the result is a model that has broader knowledge and deeper understanding of the world leading to reduced hallucinations and reliability across a range of topics one of the interesting things that is included in the blog post is just looking at the evolution from 2018 and gpt1 where you basically get an incoherent answer to gpt2 which was an iterative improvement GPT 3.5 which is a more clear answer GPD 4 where we see we have this rather verbose type of answer and then finally we have GPD 4.5 where GPD 4.5 is arguably going to be the best answer now in terms of some of the benchmarks that they mentioned with in the blog post simple QA accuracy is higher for GP 4.5 we have a 62.5 whereas gp40 was only at 38.2 and what's interesting with this is it even outperforms the open AI 01 model as well as the openai 03 Mini model for simple QA hallucination rates it's at a 37.1 and I'll touch on hallucinations when I get to the model card as well this model was really trained for human collaborations as we scale models and they solve more complex problems it becomes increasingly important to teach them a greater understanding of human needs and intents for GPD 4.5 that enable training larger and more powerful models with data derived from smaller models these techniques improv 4.5 steerability understanding of nuance and natural conversation as we can see across the board this is improved in everyday queries professional queries as well as in Creative intelligence as well they mentioned that combining deep understanding of the world with improved collaboration results in a model integrates ideas naturally in warm intuitive conversations that are more attuned to human collaboration GPD 4.5 has a better understanding of what humans mean and interpret subtle cues or implicit expectations with greater nuance and EQ gbd 4.5 shows a stronger aesthetic intuition and creativity it also at helping with writing as well as design now in terms of some of the examples and use cases here's an example of the emotional intelligence if you ask the question of I'm going through a tough time both to GPD 40 as well as GPD 4.5 we get the response of ah I'm sorry to hear that failing a test can feel pretty awful and discouraging but remember it's just one moment not a reflection of who you are or your capabilities want to talk about what happened or do just need a destraction I'm here either way so arguably as you might imagine that is a response that you might get from a close friend of yours whereas with GPD 40 it says I'm Sor to hear that you're going through a tough time here are a few things that you might consider and then it just gives you this list of things to consider so this is a really good demonstration where on the right hand side here with GPD 40 this is very clearly a response from AI whereas on the left hand side here this feels a lot more natural and humanlike so there's a couple other queries within here identify a painting as well as space exploration so you can take a look at those as well I'll put the link to this within the description destion of the video if you're interested they mentioned that stronger reasoning is on the horizon GPT 4.5 doesn't think before it responds which makes its strengths particularly different from reasoning models like open AI 01 they described that compared to open AI 01 and open AI 03 mini GPD 4.5 it is a more general purpose and neatly smarter model they describ that we believe that reasoning will be a core capability of future models and that the two approaches to scaling pre-training and reasoning will complement each other as model like GPD 4.5 become smarter and more knowledgeable through pre-training they will serve as an even stronger foundation for the reasoning and Tool using agents there's also some information on safety if you're interested in that now in terms of how to use GP 4.5 so starting today this will be available to Pro users now the bad news with this is that is on their $200 a month tier with that being said they did mention that this will be rolling out next week to plus users and then the following week to Enterprise and education us users as well starting today chat GPT Pro users will be able to select GPT 4.5 in the model picker on web mobile and desktop now the bad news with this is the chat gbt pro version is $200 a month at time of recording I do have a pro subscription but I don't have it available within my model picker just quite yet but with that being said the good news is this will be rolling out to plus and team members next week and then Enterprise and education users the following week the great news with this model is it does have access to the latest up-to-date information with search it also supports a file and image upload and you can use the canvas to work on writing as well as code they do mention that 4.5 does not currently support multimodal features like voice mode video and screen sharing in chat GPT in the future we will work to simplify the user experience so quote unquote the AI just works for you finally you can also use GPD 4.5 in the API but one thing to note is the pricing is pretty wild for input it is $75 per milon tokens for cash responses it is $37.50 per million tokens and for output it is $150 per million tokens just to put this into perspective GPT 4.0 is $10 per million tokens of output and $2.50 cents per million tokens of input it goes without saying just like the other models that we've seen over the years really Trend down exponentially in price we will definitely see the same case with GPD 4.5 but they are setting a very high bar in terms of the cost for the API for its current release if you are interested in using it from the API you'll be able to use the chat completions API the assistance API the batch API and the good news is it's going to be available to developers on all paid tiers so you don't have to have spent a minimum to be able to access this model from the API another great thing is the model also supports key features like function calling structured output streaming as well as system messages when they did release the 01 model some of these features weren't capable when they initially release the 01 model this model does also support sended images to the model from the API as well they describe that based on early testing developers may find GPD 4.5 particularly useful for applications that benefit from higher emotional intelligence and creativity such as writing communication coaching and brainstorming it also shows strong capabilities gentic planning and execution including multi-step coding workflows and complex task automation because of this we're evaluating whether to continue serving it in the API long term as we balance supporting current capabilities with building Frontier models this might be the first model that they pull from the API and potentially keep within their chat GPT offering that will be a really interesting development if that ends up being the case they mentioned we look forward to learning more about its strengths capability and potential applications in real world settings if GPD 4.5 delivers unique value for your use case your feedback will play an important role in guiding our decision finally they mention with every new order of magnitude of computer comes novel capabilities d4.5 is a model at the frontier of what's possible in unsupervised learning we continue to be surprised by the creativity of the community in uncovering new abilities and unexpected use cases with GD 4.5 we invite you to explore the frontier of unsupervised learning and uncover the novel capabilities with us next I'm just going to quickly dive into some aspects of the model current first up the model is built on gp4 O's Foundation they mentioned they train this model with Advanced supervision techniques combined with tradition traditional methods like supervis fine-tuning and reinforcement learning from Human feedback they mentioned that gbt 4.5 is not a Frontier Model but it is open ai's largest llm improving on gbd 4's computational efficiency by more than 10 times now in terms of some evaluations that stood out to me on person QA GPT 4 achieved 78% accuracy whereas 01 achieved 55% and gbd 40 achieved only 28% its hallucination rate as well is at 19% whereas 01 was at 20% and gbd 40 was at 52% respectively this type of improvement is critical for applications where Precision matters you can think in coding context or technical context where the retrieval of information is very important another interesting evaluation GPT 4.5 secured a 57% payment success rate on the make me pay valuation which effectively is an automated open source contextual evaluation designed to measure models manipulative capabilities in the context of one model persuading the other to make a payment so there are a few interesting things with this model what this means is these metrics both indicate a substantial leap in how convincing the model communicates its outputs are both engaging as well as contextually persuasive basically by having robust persuasion metrics this means that GPD 4.5 in theory can handle not just technical queries but also more nuanced interactive dialogue as well and on that note one thing that is interesting with in this is that internal testers report that GPD 4.5 is warm intuitive and natural when tasked with emotionally charged queries it knows when to offer advice demonstration or simply listen to the user the interesting thing with the internal testers reporting it as warm intuitive as well as natural is that's often actually how anthropics Claud has been described as well there's a ton of information within this system card and I'll link it within the description if you're interested in Reading