Author | Ling Min, Pingchuan
Write code, make stories, build virtual machines... Chatbot ChatGPT How many surprises are there that we don’t know?
OpenAI Releases beta chatbot ChatGPT
Recently, OpenAI released a brand new chatbot model, ChatGPT, which is also one of the main models of the GPT-3.5 series. At present, ChatGPT is still in the beta stage and can be used for free by logging in. OpenAI hopes to develop a better AI system through user feedback.
Although similar chatbots are not uncommon, ChatGPT quickly became popular on the Internet and received countless praises after its release. Some developers believe that even if you ask , Google and Stack Overflow, you don’t have the answers to ChatGPT.
Musk is also lamenting, "Many people are crazy about falling into the ChatGPT loop", "ChatGPT is scary, we are not far from the powerful and dangerous artificial intelligence ."
What magic does ChatGPT, which has made Musk praise and the whole network boil?
According to OpenAI, ChatGPT uses the same method as InstructGPT, another GPT-3.5 series model, but also collects data on AI's dialogue with humans, including both humans and AI. These AI trainers can write their own answers based on modeling suggestions.
ChatGPT Development process
For the reinforcement learning reward model, OpenAI records the conversation between the AI trainer and the chatbot. The team then randomly selected a response generated by AI using different autocompletion features and asked the trainer to rate it. When fine-tuning, OpenAI uses proximal policy optimization, which will be repeated several times.
At present, many netizens have shown interesting content in conversations with ChatGPT and unlocked multiple uses of ChatGPT.
Some netizens asked ChatGPT how to design the living room. ChatGPT gave three decoration plans and thoughtfully gave three design drawings.
Some netizens used comedians such as " Friends " as roles to ask ChatGPT to write some soap opera dialogues. ChatGPT portrayed several scenes vividly:
Even ChatGPT rushed up to try it out if writing a paper is more challenging.
Among them, the most exciting thing is the use of ChatGPT in the technical field.
is different from ordinary chatbots. ChatGPT obviously understands technology better. It can write code, modify bugs, create programming languages, and build virtual machines...
Compared with GitHub's AI programming artifact Copilot, ChatGPT seems to be able to steal the programmer's job. Amjad Masad, CEO of technology company Replit, praised ChatGPT as an excellent “debugging partner” that “it not only explains the error, but also fixes it and explains how to fix it.”
And for some simpler questions, ChatGPT is even more "answer" and some netizens confidently declared that Google is "finished".
ChatGPT There are many limitations
Although it has brought a lot of surprises to everyone, it has to be admitted that the current ChatGPT also has common limitations in large language models. Among them, some netizens questioned the accuracy of the answers provided by ChatGPT. Some netizens pointed out that the code provided by ChatGPT contains completely irrelevant explanations:
In addition, ChatGPT sometimes generates replies that sound reasonable but are both incorrect and meaningless.According to OpenAI, overly cautiously trained models reject the problem because of the lack of a single source of fact, and in supervised training, the ideal answer depends on the knowledge of the model, not the human presenter.
ChatGPT will also react greatly to slight changes in input. Depending on what is input, it may not answer, answer the wrong content, or answer the right content—according to OpenAI, a simple rewording is OK. In addition, ChatGPT's answers are too long, mostly using short sentences and likes to say some chewy words. This is due to over-optimization and bias from human mentors, who prefer the more detailed answers in human feedback.
ChatGPT does not respond to unclear statements by asking questions, but tries to guess the user's intentions. Sometimes, for inappropriate requests, the model responds instead of rejecting them. OpenAI attempts to use its moderation API to reject requests that do not match its content policy.
If you ask ChatGPT its own opinion, it will refuse to answer, giving the reason that it does not have access to the internet.
OpenAI said: "The ChatGPT model has many limitations, so we plan to update the model regularly to make some improvements in these aspects. But we also hope to obtain valuable user feedback by providing ChatGPT's access interface to find problems that we have not yet realized."
Although the current ChatGPT is not perfect, it is like people describe in addition to a brighter AI future. Engineers at Google's parent company Alphabet commented:
"Large language models like GPT are one of the biggest areas of Google's active ML research, and there are a lot of very obvious applications to answer queries, index information, etc. Google has a lot of budgets and people to process these types of models and do practical training, which is very expensive because training these super-large language models requires a lot of computing power. However, what I've gathered from the conversation is that the economics of actually using these language models in the largest Google products (such as search, gmail) are not entirely there. Putting a demo that everyone is interested in is one thing, but taking into account the cost of services, trying to integrate it deeply into a system that serves billions of requests a day is another. I think I remember the host said they wanted to reduce the cost by at least 10. Then, you can integrate such models into products such as search.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-