5天注册用户超100万，月活破亿用时仅2个多月······
ChatGPT的爆火毋庸置疑，投行瑞银集团发布研报称之为“史上增长最快的消费者应用”。根据Sensor Tower数据，TikTok达到1亿用户用了9个月，Instagram则为2年半。此外，Worldof Engineering整理的一份达到全球1亿用户所用时间排名显示，iTunes用了6年半、Twitter用了5年、Meta（Facebook）用了4年半、WhatsApp用了3年半。
好了，聊完这款应用的逆天增长数据，我们来聊一聊ChatGPT到底是什么？从英语字面意思来看，Chat是“聊天”的意思，而GPT就要复杂一些，它指的是Generative Pretrained Transformer 3。
那么，从人工智能的专业角度来说，ChatGPT是一个基于GPT-3（Generative Pretrained Transformer 3）的语言模型，用于处理自然语言问答。GPT-3是由人工智能公司OpenAI开发的一种大型神经网络模型，能够处理自然语言文本。ChatGPT是基于GPT-3模型构建的，能够根据用户输入的问题，生成自然语言的回答。通常，ChatGPT被用于聊天机器人、问答系统和其他自然语言处理应用中。
为什么ChatGPT能火爆出圈？
↓↓↓ 上下滑动，查看演讲稿 ↓↓↓
In the world of artificial intelligence.
There's been one name that's been on everyone's lips lately.
ChatGPT. ChatGPT. ChatGPT.
OpenAI, the San Francisco-based startup that created ChatGPT opened the tool up for public testing in November 2022.
In under a week, The AI model amassed over a million users, according to OpenAI's CEO.
By the end of January, ChatGPT was averaging about 13 million visitors per day.
Users have put the text-based AI chatbot through its paces with prompts ranging from the silly, to the practical and even the creative.
The future is unfolding.
It's hard to ignore.
Automation is coming, just knocking on the door.
The tool has become so popular that it's not unusual for users to be met with a message saying that ChatGPT is at capacity and can't handle any more queries.
As a linguist and as a person who cares a lot about grammatical issues, stylistic issues, punctuation, spelling, the first thing that's really impressive is how good it is at basic writing. It's better than most of the writing that I've gotten from students over the decades.
I would say it has been a explosion of interest around not just ChatGPT but around generative AI more broadly.
I've described this in a couple different scenarios as being akin to an iPhone moment.
Of course, as with any new technology, there are a number of legitimate concerns.
If people believe that there's a problem with fake news, tools like this can only make the problem worse.
OpenAI declined to be interviewed as part of this documentary. So in order to figure out what ChatGPT is and how it works, CNBC decided to go to the second best source and just ask the chatbot itself.
We entered our questions into the bot's web interface and instructed ChatGPT to answer in a one to two sentence interview format. We used the text to speech software to give ChatGPT a voice.
Can you give me a short deion of yourself and then tell me what kinds of things can you do? I am ChatGPT, a large language model trained by OpenAI.
I am capable of understanding and generating text and can answer a wide range of questions as well as generate creative writing and text summaries.
So a lot of the people that have used your services seem to think that you're pretty smart. Tell me, how do you have all of this knowledge? I have been trained on a massive data set of text which allows me to understand and generate text on a wide range of topics. The data set contains a wide variety of information, such as books, articles and websites, which allows me to have knowledge on many different subjects.
Do you have any limitations or are you sometimes wrong? As a language model, I don't have my own personal experiences or beliefs.
I can only base my responses on the information that I've been trained on, which means that my answers may be limited in their accuracy or completeness. Additionally, my knowledge is based on the data that was available to me at the time of my training, so it may not be up to date.
And as an AI model, I might make mistakes based on the model's predictions.
The maker of ChatGPT, OpenAI, has some big backers, including Microsoft, LinkedIn co-founder Reid Hoffman and Khosla Ventures. Elon Musk also helped to cofound the company back in 2015.
ChatGPT, which stands for Generative Pre-trained Transformer, is just one of the company's projects, all of which support its goal of ensuring that artificial intelligence benefits all of humanity.
At the core of ChatGPT and other chatbots like it, is what experts refer to as a language model.
For ChatGPT, that model is called GPT 3.5.
Language model is the underlying technology that powers the chatbot.
I like to think about it as the language model being the engine and the chatbot being the car.
And so the language model itself is a sort of a it's a complex, what they call, machine learning software.
It's the thing that sifts through the data and finds patterns and the chat interfaces which when people interact with, that's where they receive the answers from.
Chatbots have been around for a while, but most have been tailored towards very specific tasks, such as answering basic customer service questions about your phone bill. ChatGPT is much more sophisticated thanks to the wide sweeping data on which it was trained.
OpenAI doesn't reveal specific data that you use to train.
What we do know that it's a ton.
It surfs the web, spools, all that Internet data.
They do Wikipedia entries, a lot of archived books.
ChatGPT is part of a growing field of AI known as Generative AI.
Most of AI in the last couple of decades has really been around analyzing existing data.
So finding an anomaly in data, detecting fraud, making a movie recommendation. Generative AI is very different. It allows you to create brand new content.
That content can be text like a news article or poetry or marketing copy in a website.
It can be video.
It could even be audio, like creating brand new music.
The technology has venture capitalists excited.
Funding for generative AI companies reached $1.37 billion in 2022 alone.
Microsoft has been investing in OpenAI since 2019, when the company committed $1 billion to the startup.
In January, Microsoft announced a third round of investment.
One expert said that it could cost up to $3 million a month, about $100,000 a day.
A lot of AI researchers have sort of estimated that it costs millions of dollars to train and then operate, plus the bandwidth of keeping it alive when it's under heavy use. I mean, these are not cheap software programs.
They require a lot of investment.
In a tweet, OpenAI CEO, Sam Altman, said that while the average cost per query is a few cents, the compute costs are eye-watering. Enter Microsoft.
OpeningAI trained the models that power c=ChatGPT on Azure, Microsoft's public cloud infrastructure.
That's a bunch of servers sitting in a data center in the middle of the state of Washington and many other locations around the world.
But OpenAI is not the only company trying to crack the generative AI code.
Big tech companies and startups alike are developing a slew of generative AI programs that can transform texts to pictures or videos and offer coding suggestions, among other use cases.
One VC firm estimates that there are over 450 startups now working on generative AI.
Meanwhile, Microsoft, Meta and Google have all developed their own language models to power their version of conversational chatbots.
Though development has not always gone as planned.
Back in 2016, Microsoft released Tay, which was promptly shut down for spewing foul language.
Unlike some of the other hyped technology sectors in the past few years, this has a very real application both for individuals and for enterprises right now.
Microsoft has taken some of the products that OpenAI has built and added it to products it has.
So, for example, CarMax is a company that lets you look at reviews of cars.
And what CarMax did is it took the OpenAI service on Azure and it summarized all of those reviews of the Kia Sorento. And that way you don't have to go through 500 reviews.
Microsoft is reportedly also considering adding ChatGPT to its Bing search engine in a bid to compete with Google.
One company that's already experimenting with such a feature is You.com, which lets users conduct a search using a conversational format. However, AI tools are still far from perfect, and experts argue that users should take care to not rely on them too heavily, at least not for the time being. It's a sentiment shared even by OpenAI's CEO who said, quote, "It's a mistake to be relying on it for anything important right now.
It's a preview of progress.
We have lots of work to do on robustness and truthfulness." There's a lot of kinks that need to work out.
For instance, it generates a lot of wrong answers.
And it's really confident, though.
I talked with a legal research firm that told me when they tried using it to study court cases or to summarize court cases, it was getting a lot of the history wrong. So they couldn't use it as a legal research tool as of yet.
With its massive user base, the misinformation issue is one that's top of mind for Google.
During a monthly all-hands meeting, Google executives said that Google has much more reputational risk in providing wrong information and is thus moving more conservatively than a small startup when it comes to releasing its chatbot publicly. The company is, however, already internally testing an AI-powered chatbot and an alternative search page that can use a question and answer format. Another issue with the current iteration of AI technology is that it can be biased or toxic if the databases that it pulls from are also biased or toxic.
People have already found instances of it generating biased answers like associating women with housework, men with more scientific endeavors in certain examples.
You know, OpenAI has acknowledged this, but that's an issue that's been long plagued the AI industry and a lot of companies trying to evaluate its use.
Many of these datasets do have some kind of cleaning, vetting or wrangling process that removes content that is deemed offensive or pornographic or otherwise unacceptable so that the dataset you get on your end is clean.
But cleaning up this large dataset can come at a real human cost. A recent investigation by TIME found that OpenAI initially outsourced workers in Kenya to sift through thousands of snippets of texts describing horrific situations including sexual abuse, violence and hate speech. The workers' job was to label these examples as inappropriate so that ChatGPT could learn to detect and filter out toxic content.
According to the TIME investigation, a number of these workers walked away from the job traumatized.
CNBC reached out to OpenAI for comment regarding this incident, but the company did not respond.
Despite these efforts, there's been instances where ChatGPT has provided instructions on things like bomb making. Though experts say OpenAI is adding guardrails almost daily to prevent ChatGPT from answering such queries.
What do you do if somebody asks you to do something potentially dangerous like how to make a bomb? I cannot provide instructions or information on illegal activity such as bomb making.
Even though generative AI like ChatGPT still has a long way to go, predictions about how the technology will influence our society are already swirling.
One prominent discussion centers around generative AI's propensity to replace some human workers.
Among the industries that may be most affected by ChatGpt, experts say, are journalism, law and translation.
Programing may also change, with companies already training generative AI models to write basic code.
Relatively few of the stories in newspapers or on television news are done as investigative reporting.
They're more reporting information, and that's something that generative AI is exceedingly good at.
There now are an increasing number of tools for generating briefs, which lawyers write.
Tools like this, Barron says, even have the advantage of customizing briefs for a specific judge by taking all of the data of their past opinions and seeing which briefs succeeded in winning the case.
But not everyone shares this outlook.
This is not something that's going to take your job.
It's a novelty.
It's a gimmick. ChatGPT is fun and interesting, but it's ultimately an iterative step forward in how people can see the capabilities of artificial intelligence in the world of language.
Nevertheless, society has already had to adapt.
Worries around students misusing ChatGPT have prompted a number of school districts around the country to ban the tool. Baron suggests that educators could go back to having students take oral exams or have students turn in multiple handwritten drafts of a paper so that teachers can see the progress of a student's writing instead of just the end result. Meanwhile, image generating programs like OpenAI's Dall-E2 have raised complicated questions around creativity and ownership.
Still, some experts believe there's no replacing the human brain.
Large language models are about language, so they don't understand the world outside of language the way that you and I and other humans do.
What it may do is take away some of the really annoying tasks that were part of our job that we didn't really want to do, like some of the data cleaning and data prep tasks.
But I think this will ultimately create more opportunities for knowledge workers in the long run than the opportunities that it will preclude.
As for ChatGPT's thoughts, they're quite nuanced for a bot.
Do you think that AI, like yourself, will completely change the world? AI has the potential to greatly impact various industries and fields, but its impact will depend on how it is used and developed. And responsible usage is crucial to ensure that its impact is positive.
ChatGPT的原理是什么？
ChatGPT 是基于 GPT-3 模型构建的。GPT 是 Generative Pretrained Transformer（生成预训练变换器）的缩写，它是openAI推出的一种用于处理自然语言文本大型神经网络模型，GPT-3 是它的最新版本。
GPT-3 模型包含大量的语言数据，能够理解语言的语法和语义。当用户输入问题时，ChatGPT 会将问题转换为计算机能够理解的数据，并使用 GPT-3 模型进行处理。GPT-3 模型会根据语言数据，生成一个回答候选集，并从中选择一个最佳回答。最终，ChatGPT 会将生成的回答以自然语言的形式返回给用户。
如此复杂的模型是怎么训练的呢？简单来说是 RLHF（从人类反馈中强化学习），分三步：
1.收集提示信息（prompt）和人工形成的任务示例（demonstration），并用监督学习方法训练模型。
2. 将初始模型用在新的对比数据上，生成多个输出，人工对这些输出进行排序，排序结果用于训练奖励模型。
3.使用PPO（Proximal Policy Optimization）强化学习算法训练奖励模型。
对不起，我不是你们的励志榜样！
胡鑫宇自缢死亡案令人痛心！
《焦虑的人》一个荒谬又治愈的悬疑故事
“欢乐春节”中国电影日在罗马尼亚开幕'
刘亦菲的英语水平有多好？
旅美大熊猫“乐乐”离世终年25岁 网友：心都碎了
邓超baby在联合国做英文演讲
2023年1月总票房破100亿元
美国人口中的bottom line啥意思你记住了吗 返回搜狐，查看更多
责任编辑：