# reasoning

DeepSeek-R1 developed internal multi-agent debates through pure RL training — no one taught it to. Google researchers call this the 'Society of Thought.' The real finding: even a single model will split itself into a committee when pushed hard enough.

Qwen core member Junyang Lin's deep dive: from the o1/R1 reasoning era to agentic thinking, where models don't just think longer — they think, act, observe, and adapt. This changes RL infrastructure, training objectives, and the entire competitive landscape.

Thariq announced a new session-level feature: you can now set effort to max, letting the model reason longer and use as many tokens as needed. The catch? It burns through your usage limits faster, so you have to enable it manually each session.

Apple MLX creator Awni Hannun argues intelligence-per-watt is rising because models no longer need to memorize answers they can compute. Reasoning and tool use may let 5B-15B models approach today's frontier, though the ceiling is still unknown.

OpenAI added CoT controllability to GPT-5.4 Thinking's system card: can a model deliberately hide its reasoning? It scored just 0.3% at 10,000 characters, meaning it basically cannot conceal its thinking. For safety, that is useful news.

Claude Code CLI's core philosophy is think first, act later: SWE-bench gains, Plan Mode, Extended Thinking, multi-agent architecture, and WebSearch. Opus even researched Claude Code's latest features and reviews from inside a secure Podman container.

Inception Labs launched Mercury 2, a reasoning Diffusion LLM that refines whole passages in parallel instead of generating one token at a time. It claims 1,008 tokens/sec, 5x Claude 4.5 Haiku speed, and one-quarter the price.

Google announced Gemini 3.1 Pro preview with stronger core reasoning and a verified 77.1% on ARC-AGI-2. The real engineering question is whether the model can handle complex multi-step production workflows, not just win benchmarks.