2024-10-13-PitchBook-新兴空间简报_AGI研究（英）-10份_10页_460kb

Emerging Space Brief: AGI Research Summary

Core Content

Artificial General Intelligence (AGI) is a field of research aimed at creating software that can reason, solve problems, and adapt to new challenges without task-specific programming. Unlike narrow AI, which is designed for specific tasks, AGI seeks to replicate human-like intelligence across all domains.

Main Points

1. Historical Context and Development

• AGI research began in the mid-20th century with the invention of digital computers and the conceptualization of the Turing test in 1950.
• The term "artificial intelligence" was coined in 1955 by John McCarthy.
• Early optimism gave way to an "AI winter" in the 1970s due to research setbacks.
• The 21st century has seen renewed progress, with milestones such as IBM's Deep Blue defeating Garry Kasparov in 1997, and DeepMind's AlphaGo beating Lee Sedol in 2016.
• The release of OpenAI's ChatGPT in 2022 marked a significant leap toward AGI.

2. Current Technologies and Limitations

• Transformer models (e.g., GPT-4): Excellent at sequential data processing and natural language generation, but data-intensive and lack causal reasoning.
• Reinforcement learning: Models learn through trial and error in structured environments, but struggle with real-world complexity and sparse rewards.
• Self-play: AI systems that experiment and conduct research, but face regulatory and dead-end issues.
• Joint-Embedding Predictive Architecture (JEPA): Predicts future states based on abstract features, but is currently limited to image completion.

3. Emerging Approaches for AGI

• Neurosymbolic AI: Combines neural networks with symbolic reasoning, but is immature and difficult to scale.
• Cognitive architectures: Simulate human cognitive functions like memory and decision-making, but modeling dynamic thought processes remains challenging.
• Embodied AI: AI that interacts with the physical world, but scaling is complex.
• Long-term memory systems: Enable cumulative learning across time, but efficient implementation is difficult.
• Large world models (LWMs): Develop 3D reasoning and learn like children, but spatial reasoning needs custom datasets.

4. Challenges in AGI Development

• Technological: High computational costs, hardware constraints, energy consumption, and limitations of deep learning.
• Conceptual: Defining intelligence, knowledge representation, transfer learning, common-sense reasoning, and consciousness.
• Ethical: Risks of labor displacement, existential threats from superintelligence, and the need for ethical frameworks to ensure alignment with human values.

Applications and Milestones

Market Activity and Investment Trends

Recent Deal Activity (As of September 27, 2024)

• 74 companies have raised $255.6B in AGI-related investments.
• 321 deals were made, with a 56.4% YoY increase.
• $101.8M was the median deal size, a 158.3% YoY increase.
• $25.8B was invested in the trailing twelve months, a 21.5% YoY increase.

Top VC-backed AGI Research Companies

Top AGI Research Companies by Active Patents

Top AGI Research Investors

Recommended Reading

• "LLMs Are a Dead End to AGI, Says François Chollet," Freethink, Kristin Houser, August 3, 2024.
• "How Much Does It Cost to Train Frontier AI Models?" Epochoch AI, Ben Cottier, et al., June 3, 2024.
• "Why Artificial General Intelligence Lies Beyond Deep Learning," RAND, Swaptik Chowdhury and Steven W. Popper, February 20, 2024.
• The Geopolitics of Artificial Intelligence, Lazard, October 17, 2023.