Steve Yu @siyryu - Twitter Profile

Steve Yu

@siyryu

2 days ago

没有人发现这张图很有趣吗？山左边的蒲公英是anthropic logo，有人抓着蒲公英走agi的捷径，代表的意思不言而喻。山的右边是小米，脚踏实地，一步一步的往上爬。

Fuli Luo

@_LuoFuli

2 days ago

MiMo-V2.6: The Hard Road to Scaling Up RL MiMo-V2.6 is very likely one of the largest single RL runs, by compute, that any open-source model team has undertaken to date. In an era when compute is brutally scarce, we still chose to dedicate a team of several dozen people to one goal over an extended period: scaling up RL. That takes more than research conviction. It takes a vision for AGI, respect for the unknown, and the nerve to walk straight into the hardest problems. The result is a model whose potential was built through mid-training and unlocked through heavy RL. Today, it is the number one open-source model. I strongly recommend reading the technical report. I believe it will become one of those papers that Agent RL practitioners keep reopening and discovering something new in each time. In my view, the research innovations and engineering challenges behind it surpass those of DeepSeek R1, which I was partly involved in. Some will ask: why MixRL instead of MOPD? First, they are not competing choices. We ran MixRL on verifiable tasks of moderate difficulty, including code and related agentic tasks, and found that the resulting models generalize remarkably well. Second, tasks that are difficult to verify, extremely long-horizon, or simply too challenging to include in a joint RL run are trained separately. Including them would substantially reduce rollout efficiency or introduce significant rollout staleness. We then merge the resulting capabilities through MOPD. Games, 3D tasks, and tasks with subjective evaluation signals all fall into this category. There is also a third, slightly cheeky answer. Our team is flat enough and free enough of organizational silos that MixRL simply is not difficult for us. More importantly, everyone enjoys working this way. People from different domains come together every day, driven by the pursuit of AGI and intelligence that can continuously improve itself, to confront and resolve the RL bottlenecks in each field. I will always remember the RL daily update meetings from this period. They were intense and dense, with intelligence emerging in real time. To help the open-source community focus on solving real Agentic RL problems, we have released a Qwen model distilled from MiMo RL trajectories as a stronger starting point for RL, along with 7K diverse environments and a complete RL training framework. We hope these resources will help move Agentic RL research forward. MiMo-V2.6 is only the beginning. In an era when intelligence is easy to replicate, we still choose the hard road toward self-improvement and AGI. Much of what lies ahead remains unknown. But we are willing to keep investing the time, compute, and passion required to take on one hard problem after another and work each of them all the way through, until intelligence crosses into a new regime.

_LuoFuli's tweet photo. MiMo-V2.6: The Hard Road to Scaling Up RL

MiMo-V2.6 is very likely one of the largest single RL runs, by compute, that any open-source model team has undertaken to date. In an era when compute is brutally scarce, we still chose to dedicate a team of several dozen people to one goal over an extended period: scaling up RL. That takes more than research conviction. It takes a vision for AGI, respect for the unknown, and the nerve to walk straight into the hardest problems.

The result is a model whose potential was built through mid-training and unlocked through heavy RL. Today, it is the number one open-source model. I strongly recommend reading the technical report. I believe it will become one of those papers that Agent RL practitioners keep reopening and discovering something new in each time. In my view, the research innovations and engineering challenges behind it surpass those of DeepSeek R1, which I was partly involved in.

Some will ask: why MixRL instead of MOPD? First, they are not competing choices. We ran MixRL on verifiable tasks of moderate difficulty, including code and related agentic tasks, and found that the resulting models generalize remarkably well. Second, tasks that are difficult to verify, extremely long-horizon, or simply too challenging to include in a joint RL run are trained separately. Including them would substantially reduce rollout efficiency or introduce significant rollout staleness. We then merge the resulting capabilities through MOPD. Games, 3D tasks, and tasks with subjective evaluation signals all fall into this category.

There is also a third, slightly cheeky answer. Our team is flat enough and free enough of organizational silos that MixRL simply is not difficult for us. More importantly, everyone enjoys working this way. People from different domains come together every day, driven by the pursuit of AGI and intelligence that can continuously improve itself, to confront and resolve the RL bottlenecks in each field. I will always remember the RL daily update meetings from this period. They were intense and dense, with intelligence emerging in real time.

To help the open-source community focus on solving real Agentic RL problems, we have released a Qwen model distilled from MiMo RL trajectories as a stronger starting point for RL, along with 7K diverse environments and a complete RL training framework. We hope these resources will help move Agentic RL research forward.

MiMo-V2.6 is only the beginning. In an era when intelligence is easy to replicate, we still choose the hard road toward self-improvement and AGI. Much of what lies ahead remains unknown. But we are willing to keep investing the time, compute, and passion required to take on one hard problem after another and work each of them all the way through, until intelligence crosses into a new regime.

226

4K

356

1K

634K

1

0

47

Steve Yu

@siyryu

6 days ago

@LuvLetter_moe 这是很正常的，用来测流量，毕竟封标对b站视频很重要

1

6

0

2K

Steve Yu

@siyryu

7 days ago

@realchendahuang 用远程环境就好了，amp orbs之类的。worktree只解决了代码并行更改，没有解决环境并行，Sandbox终会替代worktree

1

0

82

siyryu retweeted

Steve Yu

@siyryu

8 days ago

@CompleteSkeptic Basically a giant reward model for everything🤔

0

1

0

173

Steve Yu

@siyryu

8 days ago

@CompleteSkeptic Basically a giant reward model for everything🤔

0

1

0

173

Steve Yu

@siyryu

12 days ago

@KyxBtc 腾讯混元啊？？

0

2

0

2K

Steve Yu

@siyryu

13 days ago

Meet Shan — a Shared Human–Agent Notebook. https://t.co/TSNvj7xvog

0

55

Steve Yu

@siyryu

13 days ago

After spending so much time working with agents, I often feel a little disoriented. I’ve built so much. But how much have I actually learned? Over the past three years, LLMs and agents have gone from something like an eight-year-old to an Olympiad gold medalist. Agent harnesses are everywhere. But much less attention seems to go toward human growth. How can AI help us understand the world better? So I built Shan, a Shared Human–Agent Notebook. It started with a thought: what if I could put all my ideas into a box, and the box could organize them for me? With one rule: leave what I wrote intact. Be quiet. Help through comments, or other ways that give me room to think. Let it handle the things my brain struggles with. Remind me what I was working on two years ago. Research, in parallel, the ten ideas that popped into my head while I was on the toilet. And a lot more. And I want to own my notes. GitHub works. I want every note to be a Git commit, and I want to be able to git clone the whole notebook. As it happens, Tibo gave me more Astra tokens than I know what to do with. So naturally, Shan also connects to my Codex subscription and helps me burn through all those unused resets. All of this is built. Shan is now in early beta. Link in the reply. Hope you like it.

siyryu's tweet photo. After spending so much time working with agents, I often feel a little disoriented. I’ve built so much. But how much have I actually learned?

Over the past three years, LLMs and agents have gone from something like an eight-year-old to an Olympiad gold medalist. Agent harnesses are everywhere. But much less attention seems to go toward human growth. How can AI help us understand the world better?

So I built Shan, a Shared Human–Agent Notebook.
It started with a thought: what if I could put all my ideas into a box, and the box could organize them for me?
With one rule: leave what I wrote intact. Be quiet. Help through comments, or other ways that give me room to think.

Let it handle the things my brain struggles with. Remind me what I was working on two years ago. Research, in parallel, the ten ideas that popped into my head while I was on the toilet. And a lot more.

And I want to own my notes. GitHub works. I want every note to be a Git commit, and I want to be able to git clone the whole notebook.

As it happens, Tibo gave me more Astra tokens than I know what to do with. So naturally, Shan also connects to my Codex subscription and helps me burn through all those unused resets.

All of this is built. Shan is now in early beta.

Link in the reply. Hope you like it.

2

0

377

Steve Yu

@siyryu

14 days ago

@Tsj_estwld 你也孙学上了？

1

0

14

Steve Yu

@siyryu

27 days ago

I’d love to see Amp add a feature that automatically renames a thread every three turns. The goal of a task can evolve as the conversation progresses, which can make it quite difficult to find previous tasks later on. Alternatively, it would be great if Amp could provide a rename thread Plugin API so we could implement small features like this ourselves. Thanks! @amp

0

52

Steve Yu

@siyryu

about 1 month ago

理解了一下。总的来说，DeepSeek 证明了这套 Harness 的插件系统可以做到时空可组合，却没有证明我们为什么Coding Agent 需要遵循这样的限制，于是被喷成嘉豪了。但如果抛开一个预设——DeepSeek Harness 只是一个要和 Codex / Claude Code 竞争的 Coding Agent——答案可能会有所不同。我们都知道，Coding 不是 AI 的全部。在 Coding 之外，LLM 的下一波高增长场景仍然不明朗。这个时候，对外开放一片肥沃的土地，招呼大家一起来种地，看看上面能不能长出一些有趣的下一代 Agent 产品，或许才是 DeepSeek 当下更感兴趣的事情。这让我想起 2023 年，肖弘邀请 Peak 加入蝴蝶效应时，那会儿还没有 Manus。真正吸引 Peak 的，是他们正在做 Monica：一款可以承载各种 AI 功能的浏览器插件，也是一块能够从 data-driven 的视角观察用户究竟会拿 AI 做什么的“空白画布”。这些对真实需求和使用方式的积累，后来也成为了 Manus 诞生的重要前提。 DSH 大概就是梁文锋的 Monica 吧。

0

2

0

149