Join our daily and weekly newsletters for the latest updates and exclusive content in the industry’s leading AI coverage. Learn more
Openai has Announced the release of GPT-4.5CEO Director General Altman said it would be last last Non-chain (cot) The model.
The company is « not a model of the border », but the new model of a new model, but is still the largest large language model with more calculation efficiency (LLM). Altman said GPT-4.5, O1 or O3-mini, the same way, in the same way as other new proposals of this new model, this new model offers more people similar to more people.
Many of the new models, many GPT-4.5 found GPT-4.5, found an interesting movement from Openai, which is temhing the expectations that the model can achieve.
Wharton Professor and AI commentator Ethan Mollick, GPT-4.5 is a « very strange and interesting model », despite being a strong writer, « It is a very strange and interesting model. »
Openai co-founder and former Tesla AI head Andrej Karpaty noted that when GPT-4.5 GPT-4 came out, he remembered him when he saw the potential of the model. One Send to xCarpathy, when using GPT 4.5, « Everything is not a little better and is awesome, but also not trivial to point it. »
Karpathy, however, warned that the model warned that the substantiation was « not to make a model ability » in cases of critical (mathematics, code, etc.). «
Detailed thoughts in the industry
This is what they say about the latest GPT iteration in a long article in the X.
« Today, the issue of GPT4.5 is noted by Openai. Since the release since this release since this release since the GPT4, this issue has been waiting for this release, which resulted in a computing computation of this issue (ie a larger model) offers a quality measurement of the slope. Each 0.5 in version is about 10x pre-computing account. Now let’s remind that GPT1 creates a suitable text with difficulty. GPT2 was a mixed toy. GPT2.5 was « skipped » in GPT3, which is even more interesting. GPT3.5, sending it as a product and passed the eve of the « Chatgept moment ». And GPT4, in turn, felt better, but I can say that it is definitely a thin feeling.
I remember being part of a hackato trying to find concrete instructions in GPT4 in GPT4. It was definitely available, but it was difficult to find clear and concrete « Slam Dunk » patterns. This … Everything was a little better, but was in a diffuse way. The word choice was a little more creative. Soon the concept of the nuance was improved. The likens have made a little more. The model was a little fun. The world’s knowledge and concept of rare domains was improved. Hallucinations were a little less often. Vibs were a little better. All ships that everything gets better felt like a rising water. Thus, I expect this to wait 10x more than 10x more claims that I have access to a few days and 10x more than the GPT4. I feel once again, 2 years ago is the same hackathon. Everything is a little better and is awesome, but also in the way that is meaningless to notice. Again, a title « free » consisting of a larger model is interesting and interesting as another quality measurement of a headline « free ».
Remember that GPT4.5 has only been training with Pretraining, controlled finetuning and RLHF, so it is not a substantial model yet. Therefore, in cases where the release of the release is critical of the substantiation (math, code, etc.), the model does not make a model. In these cases, RL and thinking exercises are incredibly important, even if an old base model (eg GPT4ish ability or so). The state of the art here remains complete O1. Probably, Openai will now try to further develop the strengthening of GPT4.5 to allow for thinking and pushing the model in these areas.
. In fact, we will see the improvement of tasks that do not think hard, and make them a lot of EQ (unlike the IQ), creativity, analogy, common understanding, humor, etc.
Thus, 5 funny / entertaining instructions that test these opportunities, using a combination of images and polls in one topic, I thought it would be fun to organize an interactive LM Arena Lite. Unfortunately, both an image and a post does not allow both an image and a survey, so I have to receive two answers from 4 to 4 and 4 (two response after 4) and two response (after 4) and the two response) and the other can receive alternative messages. After 8 hours, I will disclose the identity of which model. Let’s see what happens 🙂«
The thoughts of CEO on GPT-4.5
Other early users also saw the potential in GPT-4.5. Box CEO Aaron Levie He said in X His company is GPT-4.5, 4.5 helps to reach information and metadata based on the content of complex enterprises.
« AI leaps only continue to arrive. Openai has just announced GPT-4.5 and will then be available today for boxes today in the box AI Studio.
In GPT4.5, in GPT4.5, the advanced enterprise tested in early access mode with Box AI for the use of informational data and has done strong results. With the box AI Enterprise Eval, we try models against various different scenarios such as Q & A accuracy, reasoning capabilities and more. In particular, we focused on a basic area with the influence of the enterprise, to explore the GPT 4.5 capabilities: the extraction of structured data or metadata production, compound.
In the box, we seriously appreciate the data production models using many enterprise-level databases. One of our people is a CUAd, which consists of more than 510 commercial legal contracts. It’s once a model of the model, which is a chance to take a metadata once a time to take more than one time to take a time once a chance to try multiple times for these areas and a chance to try multiple times for these areas within the database). In our tests, GPT-4.5 compared to 19 percent, compared to 19 percent, GPT-4O, compared to GPT-4o, emphasizing the ability to manage Nuanced Contract information.
To ensure that the next GPT-4.5 can resolve the requirements of the real world owner’s content, we assessed the performance of the more seriously set of documents, box against their own problem. The bottom of the complex legal contracts – a very modal content, high-density information and lengths with more than 200 pages – to represent the most difficult scenarios for our customers. Demonstrating the advantage of the GPT-4.5, GPT-4.5, higher accurate key areas to manage complex and nuanced legal documents, consistently preferred GPT-4O.
In general, we see strong results with GPT-4.5 for complex enterprise information that will open more use in the enterprise.«
Questions about price and its importance
Early users have known that GPT-4.5 is being processed, although it is a little lazy and released.
For example, prominent Openai critic Gary Marcus is called « no » in Bluesky in GPT-4.5.
Hugging the face CEO Clement Derague commented The proof of the GPT4.5 indoor source is « meh ».
However, many noted that GPT-4.5 has nothing to do with his performance. Why did people instead of why Openai asked Leave a model so expensive It is almost forbidden for use but not as strong as other models.
A user commented in X: « So you tell me GPT-4.5 valuable than O1, but it does not perform in Benchmarks …. to make sense«
Other X users The cost of high signs « to disturb the » 4.5 model « expensive theories to prevent opponents as DeepSeek.
Depth In January became a great competitor against Openai Industrial heads The thought of the DeepSee-R1 is obvious, but thinking of as much as it is more affordable.
Source link
Leave a Reply