Since the release of the large language model GPT-4 by OpenAI, people have been trying to use this exciting application. GPT-4 can generate HTML code from hand-drawn website models. Many users have proven that it can find physical addresses from credit card transactions, generate lawsuit drafts, pass SAT math tests, aid in education and training, and even create first-person shooter games.
The power of GPT-4 is truly amazing, and as more and more users access its multi-modal version, one can expect more large-scale language models to be launched . However, while people celebrate the progress scientists have made in the field of large-scale language models, their limitations must also be noted.
Large language models like GPT-4 can perform many tasks, but they are not necessarily the best tools for them. If they complete a task successfully, it doesn't mean they are more reliable in that field.
After the release of GPT-4, many users criticized OpenAI, many of which were justified. With each release of GPT, their technical details become increasingly opaque. The technical report released by OpenAI when it released GPT-4 contained few details about the model’s architecture, training data, and other important aspects. There are various signs that OpenAI is gradually transforming from an artificial intelligence research laboratory into a company selling artificial intelligence products.
However, this does not diminish the fascinating technological breakthroughs enabled by large language models. The company OpenAI has played an important role in the development of these technologies. In just a few years, we have gone from the most mediocre deep learning models handling language tasks to large language models that can generate text that is very human-like, at least on the surface.
Furthermore, given enough parameters, computing power, and training data, Transformer (an architecture used in large language models) can learn to perform multiple tasks using a single model. This is important because until recently, deep learning models were thought to be suitable for performing only one task. Large language models can now perform several tasks by learning with zero-shot and few snapshots, and even show emergent capabilities when scaling.
ChatGPT fully demonstrates the latest capabilities of large language models. It can perform coding, Q&A, text generation and many other tasks in a single conversation. It does a better job of following instructions thanks to a training technique called Reinforcement Learning from Human Feedback (RLHF).
GPT-4 and other multimodal language models are showing a new wave of capabilities, such as including images and voice messages in conversations.
Once you look beyond the scientific achievements, you can start to think about what applications a large language model like GPT-4 can provide. For people, the guiding principle in determining whether large language models are suitable for application is their mechanics.
Like other machine learning models, large language models are predictive machines. Based on patterns in the training data, they predict the next token in the input sequence they receive, and they do this very efficiently.
Next token prediction is a good solution for certain tasks such as text generation. When a large language model is trained with instruction following techniques such as RLHF, it can perform language tasks such as writing articles, summarizing text, explaining concepts, and answering questions with astonishing results. This is one of the most accurate and useful solutions currently available for large language models.
However, large language models are still limited in their capabilities for text generation. Large language models often hallucinate, or make up things that are incorrect. Therefore, one should not trust them as a source of knowledge. This includes GPT-4. For example, in an exploration of ChatGPT by industry experts, it was found that it can sometimes generate very eloquent descriptions of complex topics, such as how deep learning works. This was helpful when he was trying to explain a concept to someone who might not understand it, but also found that ChatGPT could make some factual errors as well.
For text generation, the rule of thumb from industry experts is to only trust GPT-4 in domains you are familiar with and whose output can be verified. There are ways to improve the accuracy of the output, including fine-tuning the model with domain-specific knowledge, or providing context for the prompt by adding relevant information before it. But again, these methods require one to know enough about the field to be able to provide additional knowledge. Therefore, do not trust GPT-4 to generate text about health, legal advice, or science unless you already know these topics.
Code generation is another interesting application of GPT-4. Industry experts have reviewed GitHub Copilot, which is based on a tweaked version of GPT-3 and goes by the name Codex. Code generation becomes increasingly efficient when it is integrated into its IDE, such as Copilot, and can use existing code as scenarios to improve large language model output. However, the same rules still apply. Only use large language models to generate fully auditable code. Blindly trusting large language models can lead to non-functional and unsafe code.
For some tasks, language models like GPT-4 are not an ideal solution, even if they can solve the example. For example, one of the frequently discussed topics is the ability of large language models to perform mathematics. They have been tested on different mathematical benchmarks. GPT-4 reportedly performs very well on complex math tests.
However, it is worth noting that large language models do not calculate mathematical equations step by step like humans do. When GPT-4 is given the prompt "1 1=", people are given the correct answer. But behind the scenes, it's not doing the "add" and "move" operations. It performs the same matrix operations as all other inputs, predicting the next token in the sequence. It gives a probabilistic answer to a deterministic question. This is why the accuracy of GPT-4 and other mathematical large language models depends heavily on the training data set and works on a chance basis. One might see them achieving amazing results on very complex math problems but failing on simple elementary math problems.
This does not mean that GPT-4 is not useful for mathematics. One approach is to use model augmentation techniques, such as combining large language models with mathematical solvers. The large language model extracts the equation data from the prompt and passes it to the solver, which computes and returns the result.
Another interesting use case for GPT-4 is what Khan Academy is doing. They integrate large-scale language model courses into their online learning platform to serve as tutors for learners and assistants for teachers. Since this was one of the partners OpenAI advertised when GPT-4 was released, they may have fine-tuned GPT-4 based on their course content. The model also integrates well with the platform's content to provide scenarios and reduce errors. But it is worth noting that GPT-4 is not designed to solve mathematical problems, but to guide students and teachers to learn and teach mathematical concepts.
Once a user has decided whether GPT-4 is suitable for their application, it must be viewed from a product perspective. Each token generated by GPT-4 is the result of hundreds of billions of operations. The first question a user must ask themselves is, "Does my application need such a calculation?" In fact, in many cases, even if GPT-4 provides a perfect answer, there may be a simpler, faster, and Lower cost solution. For example, in many cases simple solutions such as regular expressions are more effective at extracting information.
In fact, industry experts had an interesting conversation with ChatGPT, in which they asked ChatGPT what “GPT-ify” meant. ChatGPT describes it as "the technology and capabilities of language models based on GPT (Generate Pretrained Transformer) architectures, such as GPT-3 or GPT-Neo, applied to specific tasks or problems."
Then Industry Experts asked if GPT-ify could be used to mock people who use GPT to do things that don't require it. ChatGPT responded this way: "Yes, GPT-ify can certainly be used sarcastically or sarcastically to suggest that someone overuses or abuses GPT for a task or problem that does not require such a complex language model. For example, if someone uses a pre-trained GPT model to generate a simple shopping list, someone might joke that they have 'GPTified their shopping list.'"
So don't GPTify a product without exploring other options . Like all other technologies, GPT-4 should be considered one of many tools in a product toolbox. As ChatGPT advises, "The key is to choose the right tool for the job, based on the specific requirements and constraints of the current task."
The above is the detailed content of What information do you need to know about GPT-4 applications?. For more information, please follow other related articles on the PHP Chinese website!
![[Web front-end] Node.js quick start](https://img.php.cn/upload/course/000/000/067/662b5d34ba7c0227.png)
