这集关于多模态和下一个“GPT-4时刻”，讨论了一个有意思的问题：随着模型规模扩大，对话能力、知识量和情商都在变强；但推理能力（尤其数学）表现是先上升后平缓，再扩大反而是下降。更大的模型做数学题倾向于跳步，不老实——这可能是next token prediction的本质缺陷🗺️ xiaoyuzhoufm.com/episode/683d2c…

Kimi K2 just hit 65.8% on SWE-bench. That's higher than GPT-4.1 (54.6%). And it's open source. You can use it right now in Cline. 🧵

Need to learn infra.

Thinking Machines Lab exists to empower humanity through advancing collaborative general intelligence. We're building multimodal AI that works with how you naturally interact with the world - through conversation, through sight, through the messy way we collaborate. We're

「 Data Contamination，Qwen2.5 」 Qwen2.5 系列的 Data Contamination 问题被证实，模型在预训练阶段已经见过评测题目。 前几个月，数篇 LLM Reasoning + RL 的论文发现，用极弱或随机奖励即可显著提升 Qwen 系列数学推理能力。 这引发出 Qwen 模型在 pretraining 阶段已经见过评测题目的疑问。

Any vertical agents are fossil fuel (a.k.a., data) generators for the general agents. Verticle agents will vanish in the long run. Only general agents will prevail. This is just getting started.

Kimi K2 is hot. This is my interview from last year with Kimi's founder, Yang Zhilin. Back then, many people said they really liked the title — 'Towards the Endless and Unknown Snow Mountains.' Hope it inspires you as well. Chinese version & podcast link at the end:)

📢📢📢 Releasing OpenThinker3-1.5B, the top-performing SFT-only model at the 1B scale! 🚀 OpenThinker3-1.5B is a smaller version of our previous 7B model, trained on the same OpenThoughts3-1.2M dataset.

1/N I’m excited to share that our latest OpenAI experimental reasoning LLM has achieved a longstanding grand challenge in AI: gold medal-level performance on the world’s most prestigious math competition—the International Math Olympiad (IMO).

It's coming.🥹🥹

i’m much more inclined to say that the RL *system* inside OpenAI is AGI rather than than any fixed model checkpoint which comes out of it

🚀 Call for Papers — NeurIPS Conference 2025 Workshop Multi-Turn Interactions in LLMs 📅 December 6/7 · 📍 San Diego Convention Center Join us to shape the future of interactive AI. Topics include but are not limited to: 🧠 Multi-Turn RL for Agentic Tasks (e.g., web & GUI agents,

This year was a major paradigm shift, where we can solve problems end to end in natural language. With novel reinforcement learning techniques, we are able to train an advanced Gemini model on multi-step reasoning proof data, which advances the model's capabilities in terms of

just finished the tech report and pushed to github. good night.

Days after the Deepmind/OpenAI results, the Kimi K2 paper already details how to do RL with non-verifiable rewards! So much impressive stuff in this paper

🚨 Olympiad math + AI: We ran Google’s Gemini 2.5 Pro on the fresh IMO 2025 problems. With careful prompting and pipeline design, it solved 5 out of 6 — remarkable for tasks demanding deep insight and creativity. The model could win gold! 🥇 #AI #Math #LLMs #IMO2025

"We design a general reinforcement learning framework that combines verifiable rewards (RLVR) with a self critique rubric reward mechanism. The model learns not only from externally defined tasks but also from evaluating its own outputs." Interesting🤔

Rubric reward via llm-as-a-judge is the way🔥

Three stages of cat-titude: playful derp, yarn entanglement, and existential teacup crisis. Which stage are you in today? #CatLife #CuteCats #FelineFun #PetHumor #KittenLove

Our paper won an outstanding paper on ACL 2025. Try our best open-source multimodal reasoning model Intern-S1 at huggingface.co/internlm/Inter…. This 241B MoE model combines strong general-task capabilities with state-of-the-art performance on a wide range of scientific tasks,