How Does OpenAI's Internal Operation Actually Work

The internal operation mechanism of OpenAI has long attracted much attention, especially after it shifted from a non-profit organization to a "capped-profit" model, its structure has become more complex. The following is a collation of OpenAI's internal operation mode based on public information, employee disclosures and industry analysis:

1. Organizational Structure: A "Dual Entity" of Non-profit and For-profit

The core structure of OpenAI is divided into two parts:

OpenAI Nonprofit (Non-profit Parent Company)

Mission: To ensure that AGI (Artificial General Intelligence) benefits all humanity, rather than being monopolized by a small number of companies or countries.

Control Right: Dominated by the board of directors. It was initially composed of technology leaders such as Elon Musk and Sam Altman, and has undergone multiple changes later (such as the board "coup" and Altman's return in 2023).

Key Power: The parent company holds overwhelming control over the for-profit subsidiary OpenAI LP, including the technological direction and capital utilization.

OpenAI LP (Capped-profit Subsidiary)

Function: Obtain funds through commercialized products (such as ChatGPT, API services) to feed back research.

Profit Cap: The returns of early investors are limited to 100 times the initial investment (may be adjusted later), and excess profits belong to the non-profit parent company.

Partner: Microsoft is the largest financial backer (with an investment of more than 13 billion US dollars), but the contract stipulates that Microsoft has no right to access the code of underlying models (such as GPT-4).

2. Decision-making Mechanism: Board of Directors vs. Management

Board of Directors

Composed of scientists, ethicists and entrepreneurs, it is responsible for supervising the safety of AGI.

In November 2023, the board suddenly dismissed CEO Sam Altman, and later reversed the decision under pressure from employees and investors, which exposed the internal conflict between "commercialization speed" and "safety priority".

Management

The CEO (Sam Altman) dominates daily operations, and the CTO (Mira Murati) is responsible for technology implementation.

The commercialization teams (such as those in charge of ChatGPT Plus subscription and enterprise API) and the research institutes (such as the GPT and DALL·E teams) are separated but collaborate with each other.

3. R&D Process: How to Build ChatGPT and GPT-4?

Research Direction

Led by scientist Ilya Sutskever (who faded out of daily management after 2023), it focuses on large language models (LLM), multimodality (such as GPT-4V) and reinforcement learning from human feedback (RLHF).

The Superalignment safety team is responsible for researching how to control AI that is smarter than humans.

Datap>

It uses mixed data sources: public web pages, co-licensed content (such as the agreement with News Corp), and synthetic data.

Training Cost: It is rumored that the training of GPT-4 cost more than 100 million US dollars, relying on the supercomputing cluster of Microsoft Azure (tens of thousands of NVIDIA GPUs).

Productization

After the research team releases a prototype (such as GPT-3), the product team optimizes the user experience (such as the interaction design of ChatGPT).

It adopts a "gradual deployment" strategy: GPT-4 is first opened to paid users, and then its functions are expanded.