How ideas spread
How a capability travels from the lab that shipped it first to everyone else, and how fast.
Reasoning models
Models trained with RL to think at length before answering, with accuracy rising with thinking time.
- o1-preview OpenAI · 12 September 2024 · first
- Open-O1 Open-O1 project · 5 October 2024 · 23 days after
- O1 Replication Journey GAIR · 8 October 2024 · 26 days after
- Marco-o1 Alibaba · 13 November 2024 · 62 days after
- R1-Lite-Preview DeepSeek · 20 November 2024 · 69 days after
- QwQ-32B-Preview Alibaba (Qwen) · 28 November 2024 · 77 days after
- Gemini 2.0 Flash Thinking Google DeepMind · 19 December 2024 · 98 days after
- R1 DeepSeek · 20 January 2025 · 130 days after
- Kimi k1.5 Moonshot AI · 20 January 2025 · 130 days after
- Grok 3 (Think) xAI · 17 February 2025 · 158 days after
- Claude 3.7 Sonnet Anthropic · 24 February 2025 · 165 days after
- Gemini 2.5 Pro Google DeepMind · 25 March 2025 · 194 days after
- Magistral Mistral · 10 June 2025 · 271 days after
- MobileLLM-R1 Meta · 12 September 2025 · 365 days after
- Muse Spark Meta · 8 April 2026 · 573 days after
Existence proof plus visible API behaviour; the missing recipe slowed copies until DeepSeek published one. Source is the chapter B08 diffusion table.
Terminal coding agents
Command-line agents that read a repo, run tools and edit code autonomously.
- Claude Code Anthropic · 24 February 2025 · first
- Codex CLI OpenAI · 16 April 2025 · 51 days after
- Gemini CLI Google · 25 June 2025 · 121 days after
Parallel product convergence on the same agent-in-terminal pattern.