Anthropic: The driving force behind Claude 3.5, the most powerful rival of GPT-4, how to create a more "compliant" AI large-scale language model?

0710 Anthropic
Anthropic3

With a large amount of technology industry funds flowing into the AI field, by understanding the fundraising status and concepts behind various AI startups, we can also have a clearer understanding of future artificial intelligence trends.

Taking the creation of generative AI and large-scale language models as an example, OpenAI is not the only one that excels. It is backed by huge investments from technology giants Amazon and Google. Anthropic has undoubtedly become the most high-profile unicorn this year after OpenAI.

What kind of company is Anthropic? Why did the founders separate from OpenAI? What is the difference between the strategic positions of the two companies?

Today I’m going to share the Anthropic story with you!

Anthropic4

3 Takeaways if you only have 1 minute

  1. The difference between Anthropic and OpenAI:
    Anthropic focuses on developing safe and transparent AI technology and is committed to the explainability and controllability of AI systems. At the same time, the training method is more "constitutional". The training process will determine the model through human feedback (RLHF) to consider non-Western, non-rich or The values of industrialized culture are suitable for applications that require a high degree of trustworthiness and ethics. In contrast, OpenAI focuses on powerful generation capabilities and natural language processing flexibility, and is suitable for diverse scenarios such as entertainment, education, and general business applications.

  2. Differences in shopping mall strategic positions:
    Anthropic is popular in high-risk industries such as finance, medicine, and law because its products can provide detailed explanations of data sources and decision-making processes, making the interaction process with humans more transparent and trustworthy.
    OpenAI has wider visibility and coverage in the market, especially in industries such as entertainment and education that require flexible and natural generation capabilities.

  3. Anthropic prospects:
    In terms of expanding its business territory, shortly after OpenAI released GPT 4o, Anthropic also announced that Claude has been launched in Europe and has begun to support multiple languages such as French, German, Italian and Spanish, aiming to achieve more diversification and higher coverage. Make rapid progress.
    Because Europe has always been one of the loudest regions on the issue of artificial intelligence safety and regulation, especially the Artificial Intelligence Act passed earlier this year. Co-founder Amodei said that Anthropic has been committed to complying with public intelligence regulations such as the EU GDPR and is ready to operate within the European framework.

    In terms of user experience, Anthropic also hopes that users can interact with Claude in the most intuitive way - whether through mobile phones, personal computers or the API itself. Currently, they are also actively creating a very smooth experience so that Claude users can switch freely between different accounts and different devices.

 

Founding background

Anthropic1

Anthropic was founded in 2021 by brother and sister Dario and Daniela Amodei, former OpenAI executives. The other five members are also from OpenAI. Anthropic was founded out of doubts about the OpenAI development process, hoping to find another way to develop a humane and safe AI system through more rigorous training methods to ensure that the AI application process can remain transparent and safe.

Anthropic’s core philosophy: create CAI (Constitutional Artificial Intelligence)

CAI (Constitutional AI) what is?

CAI is the abbreviation of "Constitutional AI".

Explain CAI in one sentence: AI trained with a set of standards that meets current regulations.

You can imagine that there is a "constitution" in the process of training AI. This constitution tells AI what to do in various situations is correct and safe, so that the output content can be more in line with today's moral and legal norms. , while also making the training process more transparent.
The "AI Constitution" created by Anthropic draws its regulations from documents such as the United Nations Universal Declaration of Human Rights, as well as the internal team's experience in interacting with AI.

Anthropic7

How was the AI constitution behind CAI formulated?

Anthropic developed this AI constitution mainly through public opinion surveys:

They invited about 1,000 Americans of different ages, genders, incomes and regions to participate in the formulation of an AI constitution on the open source voting platform Polis, allowing them to vote on some existing AI rules according to their own ethnic groups and positions. make a suggestion.
After collecting approximately 1,200 diverse suggestions and 38,200 votes, the Anthropic team further filtered out the rules that most people agreed with. Finally, these rules are organized into an AI constitution, which is used as a guideline for training new AI systems.

Anthropic2

AI alignment: Aligning AI with human values?

AI alignment refers to making AI "consistent" with human values, and Anthropic's CAI method is designed to achieve this goal.

But how?
By studying the core technology Mechanistic interpretability (mechanical interpretability/model explanation ability), Anthropic is similar to cutting open the brain of a robot, allowing developers to clearly understand how the internal AI system operates, how it produces each sentence, and how it thinks. Strictly supervise the thinking process of artificial intelligence, rather than just focusing on the quality of the results.

Anthropic5

Anthropic vs. OpenAI 

 

Both Anthropic and OpenAI are committed to building large-scale language models, but Anthropic's Claude is more focused on developing AI systems that can "clearly explain the thinking process" and are more compliant and controllable.

In contrast, OpenAI's ChatGPT focuses more on model flexibility and develops more powerful general AI technologies, such as more approachable speech, image recognition, image output, etc.
Because the training purposes are different, although OpenAI has firmly established its position in user coverage, Anthropic will be a more suitable choice for companies that strictly require AI security and explainability.

What is the difference between the two methods of training AI?

Anthropic: Reinforcement Learning with Human Feedback (RLHF)
OpenAI: Large-Scale Data and Deep Learning

Anthropic looks forward to developing a more transparent AI system, so it will place more emphasis on the controllability and explanation capabilities of the AI model during the training process. Anthropic mainly uses reinforcement learning and human feedback (RLHF), and the team will focus on ensuring the safety and reliability of the AI system in various scenarios.
In contrast, OpenAI hopes that the output can be diverse and flexible, so it focuses more on large-scale data and deep learning to maximize the performance and generation capabilities of the AI model. For example, GPT-4 uses a large amount of text data for pre-training to improve the model's natural language processing and generation diversity.

 

What are the differences between the two language models?

Taking the output text as an example, the most obvious differences between the two models are:

Anthropic (Claude)

  1. High content transparency:
    Claude provides a detailed explanation of the process and data sources, explaining which scientific models he uses to predict and analyze.
  2. Compliance:
    Claude's responses are generally more compliant, suitable for areas where strict legal and ethical standards need to be followed.

OpenAI (ChatGPT)

  1. More flexible and natural:
    ChatGPT's response focuses more on generating natural and fluent text, providing clear and concise explanations, suitable for a wide range of interactions and diverse application scenarios, such as entertainment, education and general business applications.
  1. Be more creative:
    ChatGPT's generation capabilities are powerful, able to handle a wider range of topics, and provide a rich interactive experience.

 

Let’s have a more concrete experience. How will the two models respond to the same prompt?

Take “Explain the causes and effects of climate change” as an example

Claude's response:

“Climate change refers to changes in Earth’s long-term weather patterns, primarily caused by increases in greenhouse gas emissions caused by human activities. Burning fossil fuels (such as coal, oil and natural gas) and deforestation are both major causes of increases in greenhouse gases.
These gases (such as carbon dioxide and methane) accumulate in the atmosphere, forming a 'greenhouse' that increases the temperature of the Earth.
This greenhouse effect leads to an increase in extreme weather events, rising sea levels and changes in ecosystems, which in turn affects human health, agricultural production and water supply. "

“We use multiple scientific models to simulate and predict the impacts of climate change, which are tuned based on historical data and current trends.
Our analysis considers greenhouse gas emissions under different scenarios and assesses their potential impact on global temperatures and weather patterns. These models and data come from authoritative organizations such as the IPCC (Intergovernmental Panel on Climate Change). "

ChatGPT's response

"Climate change is caused by the combination of human activities and natural factors. During human activities, burning fossil fuels releases large amounts of carbon dioxide, a major greenhouse gas that will warm the earth.
In addition, deforestation reduces the absorption of carbon dioxide, exacerbating climate warming. Natural factors such as volcanic eruptions and changes in solar radiation also affect climate, but under the influence of human activities, these natural factors play a relatively small role.
The impacts of climate change include an increase in extreme weather events, rising sea levels, melting glaciers, ecosystem damage, and human health problems. These changes could lead to reduced agricultural production, water shortages and more climate refugees.

 

It is obvious that Claude will explain the process with clear reasoning and attach reference sources;
ChatGPT is more user-friendly and uses simpler and easier-to-understand language to help users understand.

To compare the two models in a more anthropomorphic way:

can put Claude imagines a gentle but somewhat serious teacher who is very attentive to the emotions and safety of his students. Ensure that your words and deeds will not harm students through a strict code of ethics (CAI). At the same time, we frequently communicate with students and parents (regular open source voting), and adjust teaching methods after adopting students’ opinions to better meet everyone’s expectations and needs.

In comparison, GPT is more interesting but less controlled. You can imagine that GPT is a very smart teacher who can teach students according to their aptitude. He has a wealth of knowledge and can answer various questions and provide detailed explanations. But sometimes the diversity of students' backgrounds may be overlooked, or sensitive topics may not be considered that may be harmful to some students.

Anthropic and OpenAI, a comparison of their strategic positions on the AI battlefield

 

In today's AI industry, Anthropic and OpenAI each occupy an important strategic position.

Because Anthropic focuses more on the security and transparency of AI technology, it is favored by companies in industries with higher risk and compliance requirements, such as finance, medical and legal. AI products that emphasize interpretability can help these industries better understand and apply AI technology.

OpenAI targets a wider range of users. Because of ChatGPT's flexible content output and diversified product portfolio, it occupies a wide share of the current market. Especially in entertainment, education and general business applications, it has higher visibility and usage.

 

Anthropic investment status

Anthropic6

 

 Anthropic has raised nearly US$8 billion so far, with a post-investment valuation of approximately US$18.4 billion, of which more than US$7 billion came from Amazon’s US$4 billion investment to collaborate on building smart chatbots on AWS; and Google’s subsequent US$2 billion increase, we are looking forward to Collaborate with its DeepMind AI research center to create safer AI applications.

Although these investments are still not equivalent to the US$10 billion Microsoft invested in OpenAI, they still show that there will be more diverse faces on the AI battlefield in the future.



The future of Anthropic

In terms of expanding its business territory, shortly after OpenAI released GPT 4o, Anthropic also announced that Claude has been launched in Europe and has begun to support multiple languages such as French, German, Italian and Spanish, aiming to achieve more diversification and higher coverage. Make rapid progress.
Because Europe has always been one of the loudest regions on the issue of artificial intelligence safety and regulation, especially the Artificial Intelligence Act passed earlier this year. Co-founder Amodei said that Anthropic has been committed to complying with public intelligence regulations such as the EU GDPR and is ready to operate within the European framework.

In terms of user experience, Anthropic also hopes that users can interact with Claude in the most intuitive way - whether through mobile phones, personal computers or the API itself. Currently, they are also actively creating a very smooth experience so that Claude users can switch freely between different accounts and different devices.

3 Takeaways

 

  1. The difference between Anthropic and OpenAI:
    Anthropic focuses on developing safe and transparent AI technology and is committed to the explainability and controllability of AI systems. At the same time, the training method is more "constitutional". The training process will determine the model through human feedback (RLHF) to consider non-Western, non-rich or The values of industrialized culture are suitable for applications that require a high degree of trustworthiness and ethics. In contrast, OpenAI focuses on powerful generation capabilities and natural language processing flexibility, and is suitable for diverse scenarios such as entertainment, education, and general business applications.

  2. Differences in shopping mall strategic positions:
    Anthropic is popular in high-risk industries such as finance, medicine, and law because its products can provide detailed explanations of data sources and decision-making processes, making the interaction process with humans more transparent and trustworthy.
    OpenAI has wider visibility and coverage in the market, especially in industries such as entertainment and education that require flexible and natural generation capabilities.

  3. Anthropic prospects:
    In terms of expanding its business territory, shortly after OpenAI released GPT 4o, Anthropic also announced that Claude has been launched in Europe and has begun to support multiple languages such as French, German, Italian and Spanish, aiming to achieve more diversification and higher coverage. Make rapid progress.
    Because Europe has always been one of the loudest regions on the issue of artificial intelligence safety and regulation, especially the Artificial Intelligence Act passed earlier this year. Co-founder Amodei said that Anthropic has been committed to complying with public intelligence regulations such as the EU GDPR and is ready to operate within the European framework.

    In terms of user experience, Anthropic also hopes that users can interact with Claude in the most intuitive way - whether through mobile phones, personal computers or the API itself. Currently, they are also actively creating a very smooth experience so that Claude users can switch freely between different accounts and different devices.

 

Conclusion

Anthropic CEO atInterview last weekThe report mentioned that the cost of artificial intelligence training is still horribly high. A low-end and basic model costs about US$100 million, and the Cluade in hand costs more than US$1 billion to train.

It follows that the sky-high cost of developing artificial intelligence has set a high bar for startups that want to train their own models to compete with OpenAI or Anthropic, and the battlefield for large language models will remain in hand in the short term. Dominated by tech giants with huge amounts of money.

The emergence of Anthropic allows people to have more diverse imaginations about the appearance of AI large-scale language models, and also deploys solutions in advance for the ethical problems that AI may cause in the future.

Thank you for reading this article!
If you feel that you still have more to learn, please subscribe to my newsletter [Roxanne's Tech Talk]
I will share more interesting science and technology stories on it! 🥳

Join 500 people and learn the latest technology and knowledge together Subscribe Roxanne's Tech Talk
You are also welcome to communicate through Linkedin👩🏻‍💻 Roxanne Chen



en_USEnglish