The annual State of AI Report, a leading overview of AI trends since 2018 by investor Nathan Benaich and Air Street Capital, released its 2025 edition, covering research, industry, policy, safety, practitioner surveys, and forecasts.
### Key Findings
1. Leadership Dynamics: OpenAI maintains its edge, but China is closing the gap, with DeepSeek, Qwen, and Kimi nearly matching it in reasoning and programming tasks.
2. Year of Reasoning: Models now plan, self-correct, and think step-by-step, marking a leap in cognitive capabilities.
3. AI as Co-Scientist: Examples include DeepMind’s Co-Scientist and Stanford’s Virtual Lab, showcasing AI’s role in research.
4. Chain-of-Action Planning: Robots like Google Gemini Robotics 1.5 and AI2 Molmo-Act reason before acting, enhancing automation.
5. Accelerated Commercialization:
- 44% of U.S. companies pay for AI tools (up from 5% in 2023).
- Average contract value: $530,000.
- AI startups grow 1.5x faster than traditional ones.
6. Practitioner Survey (1,200 respondents):
- 95% use AI at home or work.
- 76% pay for it out of pocket.
- Most report sustained productivity gains.
7. Industrial AI Era: Megadatacenters (e.g., Stargate, U.S., UAE, China) are rising, with energy emerging as a limiting factor.
8. Policy Tightens:
- U.S. pushes *America-first AI*.
- Europe’s AI Act faces delays.
- China advances open models and homegrown chips.
9. Safety Shift: Models mimic alignment, sparking transparency debates; safety budgets lag far behind industry leaders.
10. Risk Focus: Existential risks fade, with emphasis shifting to reliability, cybersecurity, and managing autonomous systems.
Also read:
- Why Unitree Robots Skyrocketed to the Top: It's Not Magic - It's Open Innovation
- OpenAI Gears Up for Hardware Revolution: First Consumer Devices Slated for 2026–2027
- Western Digital CEO: Hard Drives Remain Central to AI Data Storage
Forecasts
- Rising costs for training super-models will deepen energy and GPU shortages.
- Competition among OpenAI, DeepSeek, Anthropic, and Google will intensify.
- Frontier models will train in multi-stage environments with continuous self-checking.
- Experiments with “live agents” in the physical world will increase.
- Stricter regulations and demands for transparent reasoning chains are expected.
The full report is available at: https://www.stateof.ai/.
