Claude 3 is surpassing ChatGPT, how to use the official website? Read the report in 3 seconds!
It’s from the same school as ChatGPT , but its ability is almost surpassing it! Claude 3, an Anthropic company co-founded by former OpenAI employees, launched the latest version of a large language model. Claude 3, which the Anthropic team proudly claims to be the most powerful model, is not afraid of GPT-4 and Gemini. Where does their confidence come from, and even the founders of Google, Amazon and Facebook invested in it? How to open an account on the official website? Can you use Chinese? The article “Foresight” will take you through it in its entirety.
Claude, launched by Anthropic, is shockingly announcing its latest version, Claude 3. Faster speed, longer context tokens, and more amazing memory capabilities mean that it has excellent memory and can handle extremely long articles. It no longer has to be used as before. Like ChatGPT, lengthy copy and paste is an extraordinary treat for users. What are the characteristics of Claude? Who is Anthropic, the developer behind it? How did the founder’s beliefs affect Claude’s training process? The article “Foresight” is organized for you to read.
What is Claude AI?
Claude is an AI model launched by Anthropic, and it is also an AI interactive service that can talk to humans. We can interact with Claude directly from the website just like using ChatGPT, or we can call Claude API and invest in internal development of the enterprise.
According to Anthropic’s official website , Claude is good at processing text and can generate a large amount of content including documents, letters, questions and answers, etc. It can also edit, rewrite, summarize, and classify these contents; it can also naturally talk to people and play different roles, like It can communicate with real people; because of the huge training data, it is proficient in different languages and is familiar with programming, and can answer specialized knowledge in many cultures and fields; finally, it automates the workflow, and it can perform certain tasks according to the instructions given by the user. Solve tasks logically.
However, he has no way of accessing the web page, which allows users to interact with it by posting information from outside. Claude also keeps in mind the Anthropic philosophy and hopes to generate helpful, honest, and harmless content. This principle is called “HHH” (Helpful, Honest, and Harmless). Therefore, through special training techniques, he hopes to meet the developers’ needs. the behavior exhibited.
How good is Claude 3?
This time, Anthropic has launched a total of three poetically named models. From weak to strong, they are Claude 3 Haiku (haiku), Claude 3 Sonnet (sonnet) and Claude 3 Opus (numbered works of classical music).
According to Anthropic’s own testing, in terms of reasoning ability, mathematical ability, and college-level knowledge, the best-performing Claude 3 Opus comprehensively outperforms opponents such as Google and OpenAI, whether it is the GPT3.5 behind the free version of ChatGPT or the GPT behind the paid version. -4, neither Google’s Gemini 1.0 Ultra nor Gemini 1.0 Pro can compete with Claude 3 Opus.
If we not only look at text capabilities but also include visual capabilities, Opus still outperforms GPT-4V and is ahead of Gemini 1.0 Ultra.
Previously, Claude did not open the multi-modal function. With this update, enterprise users will be able to upload slides, photos, text files and charts just like using ChatGPT, and ask Claude to help identify and interpret them.
For users who are familiar with the generative AI services of various companies, this is not the first time that they have seen the self-promotion of the “strongest model”, but the actual effectiveness still needs to be judged by the market.
Observe the Anthropic update, which improves the correctness of answers, more accurately determines user intentions, adds multi-modal capabilities, provides structured output formats and function calling functions, etc. Overall, it is similar to what OpenAI and Google have done in the past. , everyone is on the road to gradually improving model performance, but it can hardly be called fresh.
Regardless of the evaluation results and regardless of how many percentage points the quality has improved, the three features that really make Claude 3 stand out are its ability to process extremely long articles at ultra-fast speeds without forgetting them.
Anthropic mentioned in its introduction that the Claude 3 model can support tasks such as real-time user responses, automatic completion, and data extraction, emphasizing that “responses must be real-time.” For customer service and e-commerce enterprise users, if the conversation robot is calling API It takes too long, making customers wait and feel impatient. Such services are worthless. Therefore, Anthropic believes that knowledge retrieval and automated sales will be important battlefields in the future.
Anthropic pointed out that currently, different models of Claude 3 can handle at least 200,000 tokens, but if we look at the model capabilities alone, it can actually accept more than 1 million tokens. If you actually use the paid version of ChatGPT and Claude 3, you will find that the latter can process long documents extremely quickly and will not omit the middle and last paragraphs of long documents. It is very suitable for processing long documents such as academic research and industry reports.
Anthropic also emphasized that Claude 3 has strong recall capabilities, like finding a needle in a haystack, able to accurately find the correct information from a large amount of content input to the model.
ModelNumber of API processing tokens (10,000)
Claude 3 (additional fee) :100
Claude 3 (Standard Edition): 20
Claude 2 :10 or 20
GPT-4:8 or 12.8
GPT-3.5: 0.4 or 1.6
Gemini 1.5 Pro (additional cost): 100
Gemini 1.5 Pro (Standard Edition): 12.8
Gemini 1.0: 3.2
Note: The number of tokens that the model can handle may change as the company launches new versions; Claude is mainly called context window.
How to use Claude? How to open an account on the official website?
After opening Claude’s page , just enter your email address or use a Google account to start the registration process.
In addition to entering your name, you must also provide your mobile phone number to receive the verification SMS. Only after passing the verification can you start using Claude.
Before use, Anthropic will remind users of situations they may encounter while using Claude, such as generating misleading information, offensive content, etc.
You can then start using Claude directly. Now open Claude, which is the latest version, Claude 3. In addition to general Q&A, Claude will also remind you that it is now possible to process image formats and extract effective information from them.
Just like ChatGPT, we can start Q&A, but what’s amazing about it is that it can receive a longer context window (context window) than ChatGPT. That is to say, we can provide longer input, and Claude can “ Remember”.
Who is Anthropic, the company behind Claude? It’s actually related to ChatGPT!
Anthropic, the company that launched Claude, has a long history. Before co-founding Anthropic, the founders, brother and sister Dario Amodei and Daniela Amodei, also served as senior executives at OpenAI, which developed ChatGPT. Their elder brother Dario served as vice president of research and development, and their sister Daniela Amodei served as vice president of research and development. He serves as Vice President of Security and Policy.
According to Venture Beat , the brother and sister decided to leave the company in 2021, taking 9 employees with them, because they could not agree with OpenAI’s development path of accepting a US$1 billion investment from Microsoft and pivoting to an industry. They then established Anthropic, intending to create a more transparent, An artificial intelligence system that is more trustworthy. Similar to the two, there is Musk, the founder of Tesla, who also chose to withdraw from OpenAI because he did not agree with OpenAI’s route of moving closer to capital. because he did not agree with OpenAI’s route of moving closer to capital .
Security and trust have always been Anthropic’s core tenets. Together, they were selected into Time magazine’s “Top 100 Most Influential People in AI” list. . For Dalio and Daniella, the alignment of artificial intelligence systems with human values is a top priority, which makes them and other Compared with companies that develop AI technology and applications, they are still outstanding.
For researchers in the field of machine learning or general artificial intelligence, they often encounter an unpleasant metaphor: the AI decision-making process is like a black box. Therefore, many people have invested in research to open the black box and dismantle the reasons for AI decision-making, hoping to launch explainable AI and increase the machine’s interpretability (mechanistic interpretability). Dalio and Daniela Exactly two of them.
Observing the development process of Anthropic, it is somewhat similar to OpenAI. They position themselves as an AI safety-research lab, but if they want to create advanced models, they need computing power. Just like OpenAI introduces funds to focus on model development, Anthropic also raises funds and Authorize the right to use the model to investors and customers.
However, they are not like OpenAI’s corporate structure. Anthropic is a public benefit corporation (PCB), which prioritizes social and public interests. Unlike ordinary for-profit companies, it does not have to worry about investors’ demands for financial returns. A look, or pressure from a powerful partner.
Who invests in Anthropic? Silicon Valley rich dads Google and Amazon are here!
According to Crunchbase data , including Google, Amazon, FTX founder SBF (Sam Bankman-Fried), former Google CEO Eric Schmidt, and Facebook co-founder Dustin Moskovitz have all participated in investments in Anthropic .
In May 2021, Jaan Tallinn, an engineer involved in the development of Skype, led the investment and invested approximately US$124 million in Anthropic in the A round. Tallinn has invested in DeepMind and is very concerned about the existential risks caused by AI to humans. He participated in the founding of the Existential Risk Research Center at the University of Cambridge in the UK and the Future of Life Institute in the United States. The latter is where Musk recently participated in signing the “Stop” Organization of the initiative “6 Months for Advanced AI System Development”.
In April 2022, SBF, the now infamous founder and CEO of cryptocurrency exchange FTX , led the Series B investment of approximately US$580 million. In February 2023, Google announced an investment of US$300 million in Anthropic to obtain about 10% of the shares. In addition to developing its own Bard, it also hopes that Claude can be used as a weapon to fight ChatGPT.
In May 2023, a total of US$450 million was invested in Series C, led by Spark Capital and also participated by Google and Zoom. In August 2023, South Korean telecom company SK Telecom also invested US$100 million, hoping to create a large language model suitable for telecom companies.
In September 2023, Amazon announced that it would invest up to US$4 billion in Anthropic. For Anthropic, AWS will become the cloud service provider that the company relies on. Just like OpenAI uses Microsoft’s computing power and funding, Anthropic can invest new resources in improving In terms of model stability and performance, Amazon follows Microsoft’s lead and opens Anthropic’s models to customers through AWS, just like it allows customers to use OpenAI models on Azure, such as calling Claude 2 to generate content and provide conversation services.
What are the characteristics of Claude AI training process?
Because of the belief of the company’s founder, Claude’s design appears to be very principled.
During Claude’s training process, the developers first formulated several principles called “Constitutional AI” so that the machine could abide by them. In the first stage, Anthropic first lets the model generate content, and then lets the model self-criticize and modify the generated content in response to the criticism, thereby adjusting the direction of the model-generated content. In this stage, the model will evaluate its own based on the principles set by humans. Answer, so it is supervised learning. The second stage is to generate content from the fine-tuned model, and then use other models to “choose one”, that is, use other models to judge the quality of the generated content, thereby training a preference model.
If you are familiar with the process of OpenAI training ChatGPT, you will definitely remember the “reinforcement learning from human feedback” (RLHF) stage, which is to ask human markers to evaluate the quality of the generated content. Anthropic just asks AI to do it. Human evaluation is therefore based on “reinforcement learning from AI feedback (RLAIF)”.
In fact, Claude’s training process is very similar to ChatGPT. However, although it also learns and improves from labeled data and feedback from humans or machines, Claude formulates principles at the beginning to give the AI something to follow. Generating content that reflects the values of its creators (humans) allows Claude to reduce the hidden biases of human taggers.
Through this training process, Anthropic has trained an AI assistant that will not cause harm, but will not avoid questions, even if the user deliberately asks potentially dangerous questions. For example, ChatGPT explicitly prohibits content related to crime and violence, Claude You can still object. Anthropic emphasized that they have improved the transparency of AI’s decision-making and allowed humans to more precisely control AI’s behavior without increasing human intervention.
Thanks for your reading. Share your thoughts, and suggestions, and help shape a better experience. If you find it inspiring, share it with your friends give it a ‘clap’ and follow. Let’s build something great together — drop your comments below!