2024-03-12-OpenAI-OpenAI全网疯传的53页PDF文档_计划2027年前开发出通用人工智能（英）_53页_4mb

OpenAI's AGI Development Plan by 2027

Key Timeline and Milestones

• OpenAI initiated training of a 125 trillion parameter multimodal model, originally named Q* or Arrakis, in August 2022.
• Training completed in December 2023; launch canceled due to high inference costs.
• The original GPT-5, planned for 2025 release, was discontinued and renamed in response to delays.
• GPT-7 (formerly GPT-6) development paused by 2023-2024 due to Elon Musk's lawsuit.
• Q* 2027 (GPT-8) targeted for 2027 release to achieve AGI, originally planned as a follow-up.
• IQ estimates for the model: Q* 2023 (~48 IQ), Q* 2024 (~96 IQ), Q* 2025 (~145 IQ); delays attributed to Musk's legal action.

AGI Definition and Technical Approach

• AGI defined as AI capable of any intellectual task a human can perform, based on standard definitions.
• AGI predicted to emerge at ~100 trillion parameters, aligning with human brain synapse count (estimated 100-125 trillion).
• Relies on scaling laws from models like Chinchilla, which uses increased data to boost performance without more parameters.
• Microsoft investment of $10 billion supports computational feasibility for such models.

Evidence and Leaks

• Multiple sources (e.g., Sam Altman, Igor Baikov, rxpu) confirmed plans for a colossal multimodal model by 2027.
• Leaks in 2022-2023 revealed GPT-4/GPT-5 was trained on text, audio, images, and video data.
• OpenAI president Greg Brockman stated in 2019 a plan to develop human-brain-sized AGI by 2024.
• Despite denials, internal statements suggest ongoing development exceeds current public knowledge.

Delays and Challenges

• Elon Musk's lawsuit in 2023 caused delays in GPT-7 and related projects.
• External factors like the U.S. China chip ban in 2022-2023 may have accelerated strategic planning, not hindered it.
• Funding from Microsoft ensures financial support for large-scale AGI efforts.

Expert Opinions and Risks

• Leaders like Sam Altman express confidence in AGI feasibility by 2027, citing abundant internet data.
• Potential for AGI to surpass human-level intelligence implies risks, though OpenAI is focusing on scaling and multimodal training.
• Recent AI advancements (e.g., robotics from Tesla) support the idea that large models can achieve complex tasks.