Agentic AI
created: Sun, 12 Oct 2025 19:45:23 GMT, modified: Mon, 22 Dec 2025 03:36:13 GMT
State of AI Report 2025
Model Economics and Architecture
- Capability per dollar doubles in 3-6 months (5x faster than Moore's law)
- Model routing is a competitive advantage
- Smaller/dumber models to solve simple/specific tasks
- Bigger/smarter models for more complex tasks
- Model release roadmap is tied to fundraising
- Multi-model architecture
Browser as AI Operating System
- Browser is an AI operating system by default to which agents are plugged in
- Answering engines based on search
- Purchase intent with high conversion rate
- Dependency on search engine(s): e.g., Google
Compute Infrastructure
- Datacenters, sovereignty, power consumption/constraint per token
- Scaling demand for tokens, to which models
- Hardware constraints are impacting model availability and evolution
- Anthropic and OpenAI outages
- Custom chips and hardware makers
Model Evolution and Measurement
- How we measure AI success, intelligence and reasoning gains
- e.g., Claude performance degradation in Aug, 2025
- Models bigger, smarter, thinker
- Closed/open models
- China is the leader in open weight models
- Moonshot AI built a 1T-param MoE with 32B active trained using MuonClip
- Concentration of talents in ecosystem
- Qwen models are used as a base for derivative work and fine-tuning
- Hybrid ecosystem (closed frontier models/open models for compliance/volume)
- China is the leader in open weight models
AI Sovereignty
- Sovereignty in AI
- Model access is evenly distributed, but very few are gaining from it
Emerging Paradigms
- World models
- Scaling paradigm shift from static pre-training to dynamic, on-the-fly adaptation
- Continuous learning
- Superhuman AI systems could become "teachers" rather than just "tools"
- Extracted novel chess techniques and teach human grandmasters
- AI is moving from answering questions to generating, testing, and validating new scientific knowledge
- AlphaEvolve: a coding agent for algorithm discovery and engineering impact
- Computer Use Agents (CUA) have improved by leaps and bounds, and still fall short
- Usage costs: Some users are costing upwards of $50k/month for a single seat of Claude Code