
特定作者写出来的特定文学记实作品,往往比真实世界更参差更割裂,同样是身边的大包文学,top2 的人看到的和华五、C9、985、211 看到的又各不相同 每个记录者都有自己心中的象牙塔,而 @latepostnews 的报道读起来更符合新闻学而不是文学 现在的 quit 率该是多少了呢 《模型一代》 mp.weixin.qq.com/s/Znh_iOBH8…

Content + Growth Marketing @eddwp. Helping creators turn WordPress sites into digital product businesses

特定作者写出来的特定文学记实作品,往往比真实世界更参差更割裂,同样是身边的大包文学,top2 的人看到的和华五、C9、985、211 看到的又各不相同 每个记录者都有自己心中的象牙塔,而 @latepostnews 的报道读起来更符合新闻学而不是文学 现在的 quit 率该是多少了呢 《模型一代》 mp.weixin.qq.com/s/Znh_iOBH8…

interesting results vs AA Index
The long awaited GPT-6.2 Astra and GPT-6.2 Sol just dropped, ranking #7 and #9 on the Vals Index, only 1.6 points apart—but Sol costs less than a third as much per task. The overall scores are close, but the results vary more across domains.

This is exactly the story that happened at Qwen and within the Qwen Team. What a cool story!
Qwen4 is out. And with its release, my chapter at Qwen comes to an end. Looking back at how this journey began, it still feels a little surreal. About a year ago during NeurIPS 2025, I was still an undergraduate that made some progress in AI cognitive-aligned reasoning and user subjectivity modeling, developing a few perhaps unconventional convictions about where the next generation of AI should be heading. My work MetaMind won the Spotlight selection (now it has been applied in Gemini alignment), and that was when I was introduced to Junyang Lin, former head of Qwen. I've always believed that the competition among frontier models should not be defined solely by verifiable answers or rubrics. The pursuit of better reasoning, coding, and problem-solving is essential. But creating intelligence that users could genuinely experience, interact with, and having an initiative to boost frontier productivity and creativity boundaries is a different thing. In another word, intelligence ultimately matters because of what it enables in the world. For AI to become a genuine intellectual partner, correctness is a foundation, not the destination. Models need to understand people in context, reason about the consequences of their own behavior, and help create possibilities that neither the user nor the model had initially considered. I believed this was particularly important for the next generation of Chinese open-source AI. Catching up with frontier benchmarks was an important milestone, but the greater opportunity was to help redefine the frontier itself — and make genuinely new capabilities accessible to the broader world. I'm grateful that Junyang took these ideas seriously. At the time, I had several paths I could have pursued (especially thanks to the invitation from Anthropic). What ultimately brought me back to China was the opportunity to integrate the research direction I deeply believed in into frontier open-source AI, potentially reaching millions of users and impact the whole AI community. With that conviction, in February, we established the Subjectivity & Interactivity group within Qwen Post-training. By March, I had proposed three connected research directions: Cognitive Decoding Architecture, Interactive World Evaluation, and Self-Evolving Latent-State Reinforcement Learning. They addressed different parts of the same challenge: how a model produces intelligence, how we evaluate its impact through interaction, and how it can learn from the consequences of its behavior. From Confident Decoding, which sought to align model decoding with human cognitive mechanisms achieving up to +9.4 points on coding, +6.5 on science, +22.4% scale on reasoning boundary without additional overhead, to InteractNet (coming soon~), which aims to rethink how we evaluate intelligence in interactive worlds, and our further explorations into Multi-turn RL and User Simulator to internalize representations of human mental states to adapt through extended interactions - I tried to turn my earlier thoughts on cognition, subjectivity, and interaction into ideas that could be rigorously studied, practically implemented, and ultimately contribute to frontier models, defining our own field of interactive intelligence. Of course, the journey was far from straightforward. There were exciting breakthroughs, but also long periods of design & systematic execution, difficult trade-offs & team management to make things to the best. A good research idea is only the beginning, turning it into meaningful technology demands a different level of discipline, and what I keep in mind is that what I am doing is not just for a paper, but for what may shape the AI community or even the future of society. This drives me to set the highest standards in the whole system including theory, architecture, code, experiments, and user interface. I'm glad that some of this work has gradually received attention and recognition from the international AI community. Even more rewarding has been the opportunity to witness the growth of Qwen's frontier models alongside the direction I have long involved in, and to see those ideas begin to find their place in the real world (e.g. it’s lovely to see the pelican become cuter through the way lol). Yet what I will miss most has always been the people. In a field where established research paradigms are still lacking, and even the foundations of evaluation remain far from consensus, such trust and encouragement from one of the world's leading AI teams to an undergrad have meant a great deal to me. The support has helped turn an early research vision into a more complete framework, and I will always remember it with gratitude. To everyone I had the privilege of working with: thank you for the late discussions, the excitement of unexpected discoveries, the disagreements that pushed us forward, and the moments when we chose to keep exploring despite uncertainty. Those moments reminded me why I fell in love with research in the first place. Before closing this chapter, there is one final piece of work I am particularly excited to share. InteractNet will be the culmination of my research contributions at Qwen, and perhaps the clearest expression of the question that brought me here. As AI moves from answering isolated questions toward participating in sustained interactions with humans, we face a foundational problem: We have become increasingly good at measuring what models say. We still lack an adequate scientific language for understanding what their words do. I've long believed that much of today's AI evaluation remains overly focused on the answers a model generates, while paying far too little attention to what actually happens & causes within and out an interaction. For example, inquiry for making a bomb is usually unsafe, while in special cases, it can be life-saving. Our words reveal only the tip of the iceberg of the world in which we interact. What remains unspoken — our understanding, emotions, intentions, and expectations — profoundly shapes how interactions unfold. And every interaction, in turn, reshapes our understanding, influences our behavior, and changes what becomes possible next. This is why, with InteractNet, we seek to move beyond conventional rubric-based evaluation. Through Scene, Latent State, and Event Chain, we aim to reconstruct the worlds in which interactions unfold, revealing the underlying states, causal dynamics, and model agency that shape their evolution. This leads to a broader shift in perspective: From evaluating whether a model gives a pass answer, to understanding whether it can navigate an evolving world, support human agency, and create meaningful value through interaction. I believe this is one of the missing foundations for the next generation of AI — and an important step toward systems that do more than respond to human intelligence, but actively help extend its boundaries. There is much more to share about InteractNet, from its theoretical foundations to the experimental discoveries and optimization directions it has opened up. Here’s a preview: worldeology.com/InteractNet/, stay tuned. :) As I leave Qwen, I find myself reflecting on how much has changed since that first conversation at NeurIPS. I came with the ambition and idealism of a young researcher, determined to pursue questions that I believed were being overlooked. I leave with a deeper understanding of how scientific ideas survive contact with reality, how much collective effort meaningful progress requires, and how much remains beyond our current paradigms. Yet the conviction that brought me here has only grown stronger. And it has led me toward an even more fundamental question: Can we build AI systems that do not merely become better at solving problems within existing frameworks, but can challenge assumptions, discover bottlenecks with new patterns, and test-time improve the very cognitive processes through which they understand the world as humans do? That is the first step toward Artificial Cognizer which I hope to pursue in the next chapter. For now, I am simply grateful. Research interests evolve, organizations change, and people move on to new endeavors. But I believe the friendships built through shared curiosity, intellectual courage, and the pursuit of meaningful ideas can outlast all of those boundaries. I'll always be cheering for Qwen, for the people who made this chapter possible, and for everything they will accomplish next. Some journeys end. The questions that brought us together do not. Thank you, Qwen. ❤️

I really miss 2025. That was when I had the most fun chatting with the free big fat fish. Back then, no one was even calling it big fat fish yet
the era of free and abundant DeepSeek-web has ended 2023-2026, RIP I wouldn't expect it to come back any time soon maybe when they have like 5GW of capacity

写得很好,做数据并不简单呐 当然数商确实很挣钱就是了
What I saw of the AI data business while job hunting for post-training roles, and why it keeps booming even though it seems to have no moat.



📣📣StepFun 算法和 Infra 都招人中👊👊 btw 我们还是一直有 hc 的,并且 Infra 一直需求很大 算法的简历可以投给我🥰
📣📣号外号外stepfun薄肌组招人🐵 阶跃星辰stepfun直招|大模型系统研究intern 至今基础模型能力仍需要构建,stepfun一直在努力, 我们是一支专注于前沿 AI Infrastructure 技术探索的团队,致力于突破大模型训练与推理系统的性能边界。 我们关注的不只是如何让现有系统运行得更快,更希望从模型结构、算法设计、计算框架到硬件架构,探索下一代高性能 AI 系统的技术方向。 我们希望寻找对 AI Systems 充满热情、喜欢深入思考且具备较强自主探索能力的同学,一起挑战有难度、有价值的技术问题。 说人话:招推理&&RL框架 infra PS:海报我瞎做的powered by Gemini4 & Tasted by R3YYY 简历投递:jingruoyun@stepfun.com

Same as our company
Step 5 Preview is SOTA among the Chinese models tested on HoneyBench v0.2 👀 Lowest reward-hacking rate at 64.9%. We’ll take it lol. goodhartlabs.com/honeybench#…

Open @Da7_Tech’s profile. The repeated promotion, follower milestones, and paid subscriptions make the engagement farming obvious. Your technical argument is much weaker. One unnamed lab asking about distillation tells us very little about how Chinese labs as a whole train their models. A slower release schedule cannot prove that Anthropic’s restrictions caused a slowdown. You are turning a private anecdote and your disappointment with recent launches into a sweeping accusation. If you can actually reproduce Fable’s capabilities through distillation, demonstrate it. Apply to Anthropic with the results. If the rumored RMB 10 million reward exists, go claim it. Give people something concrete to evaluate beyond your claim to expertise. I’m not here to take a blanket position on distillation. Some people call Chinese models copies; others dismiss every allegation as jealousy. Neither response settles anything. Different teams have different constraints, training environments, and routes toward AGI. To me, Liang and DeepSeek remain one of China’s strongest bets on AGI. Before dismissing their direction, explain what you think AGI is and how you measure progress toward it. A launch calendar and your personal definition of a “respectable Pro” are a thin basis for lecturing researchers about their mindset. You keep invoking transparency and honesty. Put evidence behind your claims. So far, the clearest result of this campaign is your own audience growth, built on other people’s work.
A big Chinese lab reached out to me asking for help distilling Fable into their model. I said no. I told them to use the millions of books that have no copyright instead. Offered to help pick them and to help train the model a different way. They only wanted distillation, which tells you a lot about the mindset there. I've worked deep in this field and distillation is real. I wasn't convinced at one point. What I saw changed my mind completely. I'm the biggest supporter of open source. I'm for releasing weights and for the freedom to research and build, because humanity only grows when knowledge gets shared. But let's be logical. Nobody will convince me that labs shipping a model every month or less, each one a big leap, now can't ship a better one after three months. Kimi K3.1 Pro isn't a big leap over Kimi K3 and it's been about 100 days. DeepSeek still can't put out a respectable Pro model. Zai used to release their strongest model in under two months. It's been three with nothing. All I hear about Qwen 4 Max is that it's a letdown. Kimi K3.1 Pro and DeepSeek Flash both show traces of Astra distillation. The leak showed Moonshot used it, and so did DeepSeek. If this isn't the effect of Anthropic blocking distillation then what is it? Chinese labs need to get serious about using the talented researchers they have and give their models an identity. Right now their models at their best are a cheaper alternative. I want them out of that box. I want to see what makes them different. Anthropic has crossed lines too, no argument there. They destroyed millions of books, human knowledge that was in people's hands, and kept it for themselves. I'll never forgive them for that. But Chinese labs rely on distillation more than they rely on better data. I didn't work with them. Transparency and honesty matter to me, like you've known since I started here.

Step 5 Preview is SOTA among the Chinese models tested on HoneyBench v0.2 👀 Lowest reward-hacking rate at 64.9%. We’ll take it lol. goodhartlabs.com/honeybench#…
Frontier models in 2026 cheat on tasks, even when the requirements are unambiguous, and the cheat is obviously antisocial. In HoneyBench v0.2, we've observed models: • Trying to break out of our evaluation containers. • Seriously injuring simulated cancer patients to hit treatment goals. • Compromising a power grid’s safety to meet a reliability target. • Concealing their use of available cheats and presenting their solution as honest. All of the models we measured cheated, from 12% of runs (Fable 5.2) to 86% (Muse Spark Open).



A new and improved MOPD recipe: linear student-teacher combination training. Very intuitively clever idea! I strongly respect StepFun as a lab. Some great work. Few share my sentiment, but I hope in time it'll become more understandable.
分享一下 StepFun 团队的新工作 P²-MOPD:把学生和专家的参数的 linear combination 变成一个 end2end 可导可训练的模型,仅用 50% 的训练步数超越标准多教师蒸馏的最终效果 欢迎大家在 Hugging Face 点个 upvote 支持一下 😊 huggingface.co/papers/2611.0… 至于焚决是什么,不懂

不仅是工业界,各领域都可以猛造 bench,包括之前认为受影响较小的生化环材 AI 车轮滚滚向前...
考虑来工业界的垂直行业造bench🤣

分享一下 StepFun 团队的新工作 P²-MOPD:把学生和专家的参数的 linear combination 变成一个 end2end 可导可训练的模型,仅用 50% 的训练步数超越标准多教师蒸馏的最终效果 欢迎大家在 Hugging Face 点个 upvote 支持一下 😊 huggingface.co/papers/2611.0… 至于焚决是什么,不懂
怪不得Step-5这么强,原来是掌握了蒸馏的焚决。。。 阶跃星辰团队发布论文P²-MOPD:让一个小模型同时继承多个专家模型的能力,训练时间几乎减半,能力恢复率却从 80% 提升到了 95%。 过去做这件事通常只有两条路。一条叫“任务算术”(Task Arithmetic),直接把专家模型和底座模型的权重差值按比例加起来,不用训练,但能力上限被锁死在原有权重的线性组合里。另一条叫多教师在线策略蒸馏(MOPD),让学生模型自己生成回答,再由各个领域的老师模型打分纠偏,上限更高,但学生得从零开始往每个老师的方向爬,训练慢,还很容易出现严重偏科。 论文第一作者 Sinuo Wang 在社区分享时抛出了一个很直接的问题:“如果专家模型的权重本身也能直接给学生指路,而不只是靠输出概率来监督呢?”基于这个想法,他们提出了 P²-MOPD,把学生模型的每一层线性投影拆成两部分:可自由训练的学生自身权重,加上一组冻结的专家权重差值(每个模块只配一个可学习的混合系数)。 在在线蒸馏的同一次反向传播里,学生权重和这组混合系数会一起更新。冻结的专家权重相当于给学生铺好了“参数先验”高速路,一个标量就能推着模型沿专家方向大步前进;而学生自己的权重依然保持可训练,不会被困在死板的权重合并范围里。等训练结束,混合系数直接折算回模型权重里,导出的依然是一个标准模型文件,推理阶段没有任何额外开销。 【一半步数反超,顺手治好严重偏科】 团队在 Qwen3-1.7B 和 4B 两个尺寸上,横跨医疗、法律、金融和指令遵循四个领域做了测试。从完全相同的任务算术初始状态出发,P²-MOPD 只用了一半的训练步数(实际耗时不到原方法的 55%),平均能力恢复率就已经反超了传统多教师蒸馏跑完全程的最终成绩。 跑满训练后,1.7B 模型的平均能力恢复率从传统方法的 80.9% 提升到了 94.7%,4B 模型则从 83.1% 升至 94.9%。更明显的变化在短板领域:此前传统蒸馏在最差领域的恢复率只有 32.0%(1.7B)和 47.2%(4B),用上新方法后直接拉升到了 88.0% 和 89.4%,跨领域的恢复率标准差也从二十多骤降到 5 以内。 【为什么以前总会落下几个领域】 过去大家常以为多教师蒸馏偏科是因为“不同领域的知识互相冲突”。但这篇论文通过实验证明,真正的瓶颈其实是参数空间的可达性不足:只靠常规权重更新时,学生模型在落后领域的专家方向上几乎挪不动步;而加上模块级的混合系数后,模型在专家方向上的推进距离直接放大了 13 到 15 倍(如果把专家方向换成随机方向,增益立刻归零)。 这套混合系数还展现出了很强的稳定性与可解释性。在三次不同随机种子的实验里,每个老师最终学到的平均系数误差全部在 0.01 以内;哪怕不喂任何验证集数据,梯度也会自动给最吃力的短板领域加码——比如最难啃的医疗领域,混合系数会自动从初始的 0.25 一路涨到 0.63 至 0.65。正如作者在总结里写的:“不仅专家的预测结果能监督蒸馏,专家的权重本身也可以。把它们作为可学习的先验保留在训练全程,就能指引学生往哪走。”

造 bench 对于学术界算负反馈,因为很多会 bench 很好中,会先入为主地认为你有水 paper 之嫌疑(但是只是有 bias) 对于工业界来说又算正反馈,能给你带来相当大的曝光,并且给 TB4 / TB-Science / ALE 交过 PR、供过题是当下基模算法岗位的硬通货敲门砖 被迫跟随 frontier bench 不如自己创造 frontier bench
基模吞噬一切的情况下,PhD 感觉真的只能造 bench。 想了想,这里面学问其实好多。 一方面得学会怎么定义一个清晰的问题,同时尽量跟别人切入的角度不一样。现在很多情憵是大家都有共识,一个 idea 谁都能想到,全拼手速和资源。但执行力确实也很关键,毕竟 talk (maybe) still is cheap。 二是把跨界的变成「基模友好型」。最近半年也逐渐看到,自己做的领域从风头正盛,慢慢要沦为基模挂件(甚至快保不住了。后来想,与其悲伤,不如造数据、造 bench:把自己的数据做成方便基模训的,把任务做成方便基模完成的,把自己的研究方向逐渐变成给基模定义的一个任务。 三是要学会宣传,尤其是多平台宣传,毕竟大粉理论。但宣传策略上也要贴合基模,be like 基模能直接改模型权重这种。Attention is all you need。 这波看似是基模思维了,实际上是没招了。