思维怪怪 (@0xLogicrw)

Why can Kimi ship K3? Let me tell my story. Earlier this year, I left academia for industry. I talked to a lot of companies along the way. Here's what I saw: 1⃣Arrogance. They believe the AI war is over, and they won. No hunger for the future, and no hunger for talent. 2⃣Restlessness. Young labs short on foundation, either rushing to catch the frontier or pivoting away from the competition. 3⃣Fear. Strong teams with real experience, but from the second tier, they can't quite bring themselves to aim for #1#. 4⃣Misalignment. Everyone is optimizing for their own credit, but nobody really cares whether the company can reach AGI. Kimi was different. Over many conversations with the founders, the same thing came through every time: a raw, genuine hunger for AGI. I joined. The hunger was real. We shipped K3. This is only the beginning.

Personal Update: I've joined @StepFun_ai as a Researcher, working on LLM post-training. Many thanks to @CyouSakura and @MikaStars39 for their support. Together, we'll take another step toward the frontier.

There’s a big misconception about how GLM 5.2 was trained. Yes, they distilled Claude and GPT 5.5 — but distillation is not how they matched Opus quality. Distillation only fixed the cold start problem in RL. RLing an agentic coding model isn’t rocket science. In simplified terms: 1. RL needs trajectories — rollouts where the model actually completed a task in some env 2. No successful trajectory on a task = zero gradient = you can’t RL it. This is the cold start problem 3. Distillation solves it. You seed your model with knowledge from a smarter one (Claude, GPT) on tasks it can’t do yet 4. Now it produces positive trajectories on those tasks 5. RL on those trajectories and hill climb agentic coding 6. At that point you no longer need to distill and can solely hill climb RL to better models This is an interesting curve. I’d argue it’s harder to get to Opus 4.8 from scratch than to go from Opus 4.8 → Fable/Mythos tier. GLM 5.2 is already producing positive trajectories, so they have plenty to RL on — they’ll keep climbing to Mythos quality without distilling any further. They no longer need American models.

🧵 Deli AutoResearch SKILL is now officially open source! 🎉 Alongside it, we’re dropping our 4th survey paper — this time on Self-play. Inspired by AlphaZero, we got a powerful insight: prior knowledge doesn’t always lift the ceiling. Models can discover more globally optimal solutions just by playing against themselves. The biggest change in this paper? For the first time, the AutoResearch Agent autonomously planned GPU experiments — and submitted actual RL runs on the DeepSeek 285B model. The entire RL pipeline — experiment design, code writing, running, debugging, and conclusion summarization — was 100% automated, with zero human intervention from me. This was incredibly difficult, but an incredibly important step. GRPO is the tool being called by the AutoResearch Agent here. We see this as the beginning of our Continual Learning research journey. 🚀 As always, this is my personal research project, unaffiliated with any organization. All views are my own. #AI# #ReinforcementLearning# #SelfPlay# #OpenSource# #AutoML# #ContinualLearning# #DeepSeek#

Today I'm publishing a new essay, Policy on the AI Exponential. AI is progressing extremely fast—much faster than the policy process was built to handle. The essay lays out where I think the technology is now, and the action needed to close the gap:

Very exciting to see the cool result! Same pattern, different physics: Codex-grown heuristics matching or beating DRL agents in fluid dynamics, while staying readable, maintainable, and transferable. Heuristics were not dead. They were under-maintained.

The next step toward automating AI is automating RL environments Introducing General-Agent: A fully synthetic environment whose task corpus self-evolves and grows harder over time 4,504 tool-use tasks · 1,040 domains · 8,159 unique tools

1/5 TLDR; We used Codex to discover and maintain heuristic learning for hard fluid dynamics control cases. I’ve been applying DRL and GNN to physics since 2019,, and over the past 3 months I’ve been toying with the idea of using agents in our processes. Inspired by the blog post from @Trinkle23897, I decided to use the same strategy and have agents find readable control strategies. This means a lot to our field, where interpretability can be key for industry.

llama.cpp adds MTP for the Qwen3.6 family This is a significant milestone for the local AI ecosystem. The performance jump with these changes is massive and elevates local inference on commodity hardware further. Special thanks to Aman Gupta for leading this development!

Unsloth 给 Qwen3.6-27B 做了一套支持 MTP 的 GGUF 量化包，让这个 27B 级别的模型可以更方便地在本地通过 llama.cpp 运行。 官方说 MTP 能让 Qwen3.6 的生成速度提升约 1.4到 2.2 倍，并称 Qwen3.6-27B MTP 版本可在约 18GB 内存下运行。

ByteDance releases Lance, a 3B unified multimodal model One framework for image/video understanding, generation, and editing. With only 3B active parameters, it beats 7B+ models on benchmarks thanks to multi-task synergy and specialized MoE pathways.

Qwen3.6 now runs 2x faster with MTP GGUFs! Run locally on just 18GB RAM. ⚡️ MTP enables Qwen3.6 to generate ~1.4–2.2× faster with no accuracy change. Qwen3.6-27B MTP runs at 160 tokens/s. 35B-A3B reaches 240 t/s. GGUFs: Guide:

We're publishing our first end-to-end benchmarks for Zyphra Inference on @AMD Instinct MI355X. Our inference optimizations strongly outperform the AMD baseline and narrows the gap between MI355X and B200 for serving Kimi K2.6, GLM 5.1, and DeepSeek V3.2 🧵