GPT-4 is officially released today. With the warm-up of ChatGPT, its attention can be said to be unprecedented.
The above picture is from the official website of OpenAI
In fact, GPT-4 completed the basic training in August 2022. Before the release, it mainly needs to do some fine-tuning to remove harmful questions and answers, biased content, etc. This is actually a hurdle that the large speech model LLM cannot get around, and it will lead to some long-tail effects, that is, it is difficult to deal with it anyway, maybe 99%, or 99.9%.
The main breakthrough of GPT-4 is that it supports multi-modality, that is, it currently supports input images . Previously, ChatGPT only supported text input, and of course the output is still only text . In addition to pictures and text, human information also comes in the form of voice and video, but the amount of processing and calculation is too large, and it is estimated that it will be difficult to go online in a short time. Voice input is not difficult to do, just put a voice recognition ASR in front, depending on the application requirements.
For specific capability introductions, please refer to OpenAI official documents or other official account introductions, so I won’t go into details here. https://cdn.openai.com/papers/gpt-4.pdf
After ChatGPT and GPT-4 came out, they showed a lot of strong abilities in many aspects, such as translation, summarization, poetry, doing questions, writing programs, etc. There is no doubt that they are indeed very strong. I have done a lot of things that everyone wanted to do before, but they didn't do well. For example, Microsoft Xiaobing also received a lot of attention when it first came out, but gradually, after everyone got used to it, the attention decreased, and life finally became peaceful. I don't use Xiaoice in my daily life, and the anthropomorphic friend making function doesn't seem to appeal to me too much. Even if AI looks like a human, it is not a human after all, so make the tools well first.
Will GPT-4 do the same?
Generated by Stable Diffusion
At present, it seems that the strength of OpenAI does not lie in making applications, but hopes to be a cloud base that provides APIs for various APPs from all walks of life to call. I believe that in the near future, other cloud vendors will definitely launch corresponding APIs and fine-tuning methods for use by all walks of life.
But compared to ToB's large, medium and small companies, it is estimated that the inertia will still be relatively large. Most people will stay at the level of experience, and it will take time to truly integrate into their own workflow. In addition, some people worry that AI will steal their own. rice bowl. Therefore, it is estimated that the popularization of ToB applications will be a gradual and long process. On the contrary, the opportunity of ToC is more attractive, and users are willing to experience it. If the effect is good enough and combined with some other functions to form enough value, I believe that the progress will be relatively fast. So I am also looking forward to the launch of some toC applications in the near future.
We are currently also applying ToC in the direction of health, not to catch up with the trend. We have already thought about it a few years ago, and now we have gradually determined the direction and boundaries. Users provide value and achieve business success in this process.
Continue to accumulate, accumulate steadily, come on!
Generated by Stable Diffusion
This public number is continuously updated, welcome to pay attention.
Artificial Intelligence Technical Analysis
artificial intelligence machine learning deep learning
Three elements of AI artificial intelligence: data, computing power and algorithm
AI artificial intelligence and big data
