AGI (Artificial General Intelligence)
AGI (Artificial General Intelligence) refers to artificial intelligence that matches or exceeds human capability on any cognitive task a human can do. Contrasts with current "narrow AI" (each task requires a separately trained model).
Different definitions of AGI
| Source | Definition |
|---|---|
| Academia | System achieving human-level on broad range of tasks (no specific capability list) |
| OpenAI Charter | "Highly autonomous systems that outperform humans at most economically valuable work" |
| DeepMind | Systems that can learn across domains, transfer knowledge, solve unseen problems |
| Anthropic | Needs to be robust on long-tail tasks, not just high benchmark scores |
Current progress debate
- Optimists (Hinton, Altman): GPT-5 + Agents already show early AGI signs, achieve by 2030.
- Cautious (Yann LeCun): current LLMs lack world models, persistent memory, several paradigms away from AGI.
- Skeptics: AGI itself is marketing talk; "good enough AI" is the pragmatic goal.
Key technical gaps (as of 2025)
- Continual learning: models can't continue learning from experience post-training (need full retrain each time).
- World model: LLMs have no internal physics / causal model, rely purely on statistical patterns.
- Long-tail tasks: poor on tasks with little training data.
- Reliability: can't guarantee 99.99% accuracy in critical scenarios (medical, autonomous driving).
- Self-awareness / goal stability: whether AGI needs these is philosophically debated.
Difference from ASI
- AGI = Human-level AI (most tasks)
- ASI (Artificial Superintelligence) = Far-surpassing-human-level AI
OpenAI / Anthropic both treat AGI as the prerequisite for ASI — must align AGI first before ASI. Otherwise ASI may cause uncontrollable consequences.
Practical advice
- Don't worry about AGI timeline: next 5-10 years are about "which LLM can ship in which domain".
- Follow alignment research: alignment techniques matter more than model scale before AGI arrives.
- Pragmatic deployment: rather than chasing AGI concept, deploy current LLMs well in your own business.