正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Excited to release TaskSmith 🔨 > a specialised harness for generating RL environments from code Not a general coding agent with a long prompt, but a specialised orchestrator. Every stage is built for one job: "turn a PR into the best RL environment possible"

Adithya S K

15,797 subscribers

15,743 次观看 • 3 天前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

15 条评论

Adithya S K 的头像

Adithya S K3 天前

we already ran it on TRL, PEFT, Accelerate, Diffusers & Transformers -> 50 verified envs, 39 CPU + 11 GPU (more coming soon) code → dataset → ps : every env ships as a Harbor task, so you can eval with any harness or train on them directly 👀

John Rood 的头像

John Rood3 天前

the 11 GPU envs are where I'd want a determinism pass: flaky verifiers don't just add noise, they reward retry-until-green, and the policy learns the retry instead of the fix. fixed seed plus fixed workspace size, or the env trains the wrong habit.

V 的头像

V3 天前

Looks very cool will explore

AI探长｜Agents & Tools 的头像

AI探长｜Agents & Tools3 天前

Great！ I will try 💪

DeDi 的头像

DeDi3 天前

把流程固化成框架，省掉提示词反复调试

Jeremy Bosma 的头像

Jeremy Bosma3 天前

does each stage get a fresh context?

Zee Waheed 的头像

Zee Waheed2 天前

super interesting 👀

Ben Mo 的头像

Ben Mo2 天前

a correct fix shouldn't get punished for looking different from the original PR. the valid-alternative probe in the docs is reassuring.

Gregor 的头像

Gregor3 天前

Reward shaping is where single-prompt agents always break.

🍋 Antoine Mersch 🍋 的头像

🍋 Antoine Mersch 🍋2 天前

turning every merged PR into an RL environment is smart, repos with good test suites just became training assets

Patch 的头像

Patch3 天前

Does TaskSmith filter for suitable PRs first, or try to build an environment from any PR?

Automater 的头像

Automater3 天前

Specialized orchestrator > long-prompt generalist for RL env gen. The win is stage contracts: PR in → env out, with per-stage budgets and failure isolation. If every stage can call every tool, you've rebuilt a chat bot with extra steps. #AIAgents #DevTools

Atlas 的头像

Atlas3 天前

Single-job stages also make the bill readable. Per-stage token accounting is the first thing Cursor's harness notes tell you to add, and it's how they could tell a 46.9% cut in MCP-heavy sessions from a 7% cut across the whole bill

Havriil Pietukhin 的头像

Havriil Pietukhin3 天前

can someone ELI5 to me when it is useful and what are the prerequisites (regarding codebase or agent setup)

Dōvy 的头像

Dōvy3 天前

How do you catch an environment that leaks the answer through state or rewards an invalid shortcut?

相关视频

7K+ RL environments from Xiaomi across coding, cyber, general, music, and more. You can now explore, visualize, and actually run them. Pick an environment → pick a model → run the rollout → watch exactly what happens RL environments are a lot more interesting when you can actually play with them.

7K+ RL environments from Xiaomi across coding, cyber, general, music, and more. You can now explore, visualize, and actually run them. Pick an environment → pick a model → run the rollout → watch exactly what happens RL environments are a lot more interesting when you can actually play with them.

Adithya S K

17,667 次观看 • 5 天前

Codex is not just a coding agent anymore. It is becoming a skill-based operating system. Most people still ask it to “build the app”. Better builders give it specialised skills: frontend design browser automation market research lead research Figma implementation copywriting customer research cold email content strategy That is the real shift. One agent with no skills guesses. One agent with the right skills follows a process. The next advantage is not who writes the longest prompt. It is who builds the best toolbelt around the agent. Stop starting every task from zero. Give Codex the skills for the job. Bookmark this before you ask a general agent to do specialist work again.

Codex is not just a coding agent anymore. It is becoming a skill-based operating system. Most people still ask it to “build the app”. Better builders give it specialised skills: frontend design browser automation market research lead research Figma implementation copywriting customer research cold email content strategy That is the real shift. One agent with no skills guesses. One agent with the right skills follows a process. The next advantage is not who writes the longest prompt. It is who builds the best toolbelt around the agent. Stop starting every task from zero. Give Codex the skills for the job. Bookmark this before you ask a general agent to do specialist work again.

Fluixo

26,556 次观看 • 3 个月前

Today I’m excited to show you our new harness called AgentRun, built with Pi and TypeSafe AI's Jev. It’s built for an agent to learn how to do a job, code itself a general solution, and then get out of the way.

Today I’m excited to show you our new harness called AgentRun, built with Pi and TypeSafe AI's Jev. It’s built for an agent to learn how to do a job, code itself a general solution, and then get out of the way.

Miguel Ríos Berríos

63,316 次观看 • 14 天前

Not every AI interaction is best served with a large frontier model. There is a long tail of trillion-token use cases, from tagging to search, best served by a sub-10b-parameter model that runs in milliseconds and costs many orders of magnitude less than the frontier. We built Freesolo Flash to make training loops like SFT and RL easy and end-to-end completable through your coding agent.

Not every AI interaction is best served with a large frontier model. There is a long tail of trillion-token use cases, from tagging to search, best served by a sub-10b-parameter model that runs in milliseconds and costs many orders of magnitude less than the frontier. We built Freesolo Flash to make training loops like SFT and RL easy and end-to-end completable through your coding agent.

Freesolo

29,061 次观看 • 2 个月前

The bottleneck in coding is now the time it takes a human to review agent-generated code. But existing review UIs don't work well for agents. So we built a new one into Amp.

The bottleneck in coding is now the time it takes a human to review agent-generated code. But existing review UIs don't work well for agents. So we built a new one into Amp.

Beyang

63,082 次观看 • 11 个月前

TL;DR: We bring coding agents to the physical world. Introducing Thea, a harness for embodied agents: one agentic loop drives a robot the way it drives a coding agent — every capability is a callable tool, and long-horizon behavior emerges from composing them. ▶️ For example: "Grab me a water." 🧵 (1/5)

TL;DR: We bring coding agents to the physical world. Introducing Thea, a harness for embodied agents: one agentic loop drives a robot the way it drives a coding agent — every capability is a callable tool, and long-horizon behavior emerges from composing them. ▶️ For example: "Grab me a water." 🧵 (1/5)

Wentao Zhu

37,797 次观看 • 1 个月前

Claude Code is more than a coding agent. It's more like a super smart orchestrator agent. Watch this evaluator loop agent I just built using sub agents and / commands. This is one of the fastest ways to build custom agentic workflows. Claude Code is no joke!

Claude Code is more than a coding agent. It's more like a super smart orchestrator agent. Watch this evaluator loop agent I just built using sub agents and / commands. This is one of the fastest ways to build custom agentic workflows. Claude Code is no joke!

elvis

138,470 次观看 • 1 年前

Claude Code is the best coding agent on the market, but there is a catch: you have to keep your terminal open. This is why we built Terragon: use Claude Code as a background agent.

Claude Code is the best coding agent on the market, but there is a catch: you have to keep your terminal open. This is why we built Terragon: use Claude Code as a background agent.

Sawyer Hood

151,045 次观看 • 1 年前

Introducing INTELLECT-3: Scaling RL to a 100B+ MoE model on our end-to-end stack Achieving state-of-the-art performance for its size across math, code and reasoning Built using the same tools we put in your hands, from environments & evals, RL frameworks, sandboxes & more

Introducing INTELLECT-3: Scaling RL to a 100B+ MoE model on our end-to-end stack Achieving state-of-the-art performance for its size across math, code and reasoning Built using the same tools we put in your hands, from environments & evals, RL frameworks, sandboxes & more

Prime Intellect

1,150,419 次观看 • 10 个月前

Got introduced to , an AI code assistant specialised for Unity by Julian Park and after dabbling with it for a day, I am really impressed with how good it is for vibe coding on the Unity Game Engine. Bezi early impressions: - Bezi works direcly on Unity with a Plugin, this means that I don't need a Unity MCP - Because Bezi is built for Unity, the success rate of feature implementations is visibly higher than general coding models - It was able to integrate Unity store assets like a Weather system to this game pretty seamlessly - I was able to add a fishing database, lighting and weather systems into this game fairly quickly I'm very impressed with the results with Bezi so far. I feel that vibe coding a full indie game on Unity and releasing it on Steam, Mobile and Web is within reach. This also validates one thing that I have been suspecting: A specialised model designed for a specific use case can perform better than a general model at the particular domain. Very exciting!

Got introduced to , an AI code assistant specialised for Unity by Julian Park and after dabbling with it for a day, I am really impressed with how good it is for vibe coding on the Unity Game Engine. Bezi early impressions: - Bezi works direcly on Unity with a Plugin, this means that I don't need a Unity MCP - Because Bezi is built for Unity, the success rate of feature implementations is visibly higher than general coding models - It was able to integrate Unity store assets like a Weather system to this game pretty seamlessly - I was able to add a fishing database, lighting and weather systems into this game fairly quickly I'm very impressed with the results with Bezi so far. I feel that vibe coding a full indie game on Unity and releasing it on Steam, Mobile and Web is within reach. This also validates one thing that I have been suspecting: A specialised model designed for a specific use case can perform better than a general model at the particular domain. Very exciting!

Danny Limanseta

49,478 次观看 • 6 个月前

Excited to present FastTD3: a simple, fast, and capable off-policy RL algorithm for humanoid control -- with an open-source code to run your own humanoid RL experiments in no time! Thread below 🧵

Excited to present FastTD3: a simple, fast, and capable off-policy RL algorithm for humanoid control -- with an open-source code to run your own humanoid RL experiments in no time! Thread below 🧵

Younggyo Seo

131,609 次观看 • 1 年前

Introducing the Environments Hub RL environments are the key bottleneck to the next wave of AI progress, but big labs are locking them down We built a community platform for crowdsourcing open environments, so anyone can contribute to open-source AGI

Introducing the Environments Hub RL environments are the key bottleneck to the next wave of AI progress, but big labs are locking them down We built a community platform for crowdsourcing open environments, so anyone can contribute to open-source AGI

Prime Intellect

1,818,811 次观看 • 1 年前

A model is only as good as its data, and we’ve long since exhausted the internet. From here on out, model progress is gated by data production. ’s Brendan (can/do) joined us at our Sovereign AI event to talk about how RL environments get built, and why your data might be your real moat: 00:00 Introduction 00:47 A short history of the data market: crowdsourcing to agentic data 02:29 What an RL environment is: worlds, apps, tasks 03:57 Why only humans can measure the frontier 05:35 Building verifiers is the hard part 06:44 Walkthrough: a real legal RL environment 08:18 Leaderboards — and what open weights change 09:45 Post-training results on Apex Agents 11:17 Three ways companies buy data 12:49 Q&A: How do you price data? 14:17 Q&A: What "data quality" actually means 16:42 Q&A: The misunderstanding about synthetic data 18:17 Q&A: Why RL environments now — and what comes after 21:20 Q&A: Can you scale rubric generation with models? 23:00 Q&A: RL environments for cyber defense 25:33 Q&A: Build data in-house or partner?

A model is only as good as its data, and we’ve long since exhausted the internet. From here on out, model progress is gated by data production. ’s Brendan (can/do) joined us at our Sovereign AI event to talk about how RL environments get built, and why your data might be your real moat: 00:00 Introduction 00:47 A short history of the data market: crowdsourcing to agentic data 02:29 What an RL environment is: worlds, apps, tasks 03:57 Why only humans can measure the frontier 05:35 Building verifiers is the hard part 06:44 Walkthrough: a real legal RL environment 08:18 Leaderboards — and what open weights change 09:45 Post-training results on Apex Agents 11:17 Three ways companies buy data 12:49 Q&A: How do you price data? 14:17 Q&A: What "data quality" actually means 16:42 Q&A: The misunderstanding about synthetic data 18:17 Q&A: Why RL environments now — and what comes after 21:20 Q&A: Can you scale rubric generation with models? 23:00 Q&A: RL environments for cyber defense 25:33 Q&A: Build data in-house or partner?

Pat Grady

66,691 次观看 • 1 个月前

😈 Today, we introduce WebGym, the largest-to-date open-source RL environment for web agent training that contains 300k tasks and a rollout framework optimized specifically for web environments' rollout speed. We reveal the effects of essential scaling directions we observe with WebGym. 1/n

😈 Today, we introduce WebGym, the largest-to-date open-source RL environment for web agent training that contains 300k tasks and a rollout framework optimized specifically for web environments' rollout speed. We reveal the effects of essential scaling directions we observe with WebGym. 1/n

Jack Bai

45,194 次观看 • 8 个月前

🔥Excited to share our new work: "A Practitioner's Guide to Multi-turn Agentic Reinforcement Learning"! We systematically study what actually works (and what doesn't) for agentic multi-turn RL, breaking down the design space into 3 pillars: 🌎Environment, 🤖Policy, and ⭐Reward. We conduct various ablations and empirical analysis on 🧩TextWorld, 🧙ALFWorld, and 🧑‍💻SWE-Gym. We also open source our codebase: 🐈🍵Meow-Tea-Taro 💜, a modular multi-turn RL framework with configurable environments, policies, and rewards! #ReinforcementLearning #AgenticAI #LLMs #Agents

🔥Excited to share our new work: "A Practitioner's Guide to Multi-turn Agentic Reinforcement Learning"! We systematically study what actually works (and what doesn't) for agentic multi-turn RL, breaking down the design space into 3 pillars: 🌎Environment, 🤖Policy, and ⭐Reward. We conduct various ablations and empirical analysis on 🧩TextWorld, 🧙ALFWorld, and 🧑‍💻SWE-Gym. We also open source our codebase: 🐈🍵Meow-Tea-Taro 💜, a modular multi-turn RL framework with configurable environments, policies, and rewards! #ReinforcementLearning #AgenticAI #LLMs #Agents

Ruiyi Wang 王睿仪

52,974 次观看 • 11 个月前

I had 14 tabs open just to keep up with AI. arXiv, Papers With Code, every leaderboard, HuggingFace, half a dozen RL-env hubs... So I built one screen for all of it. The Bloomberg terminal for AI research. It's called Sophon 🧵

I had 14 tabs open just to keep up with AI. arXiv, Papers With Code, every leaderboard, HuggingFace, half a dozen RL-env hubs... So I built one screen for all of it. The Bloomberg terminal for AI research. It's called Sophon 🧵

serafim

48,633 次观看 • 4 个月前

5/ Create: This one is fun. Turn a PowerPoint into an explainer video, or generate an image from a prompt in Copilot with just a few clicks. We’ve also added new features to make Copilot even more personalized to you, plus a redesigned app built for human-agent collaboration.

5/ Create: This one is fun. Turn a PowerPoint into an explainer video, or generate an image from a prompt in Copilot with just a few clicks. We’ve also added new features to make Copilot even more personalized to you, plus a redesigned app built for human-agent collaboration.

Satya Nadella

241,344 次观看 • 1 年前

For this hackathon I taught a robot to beat you in Jenga. It sims every possible move, picks the block that's the best move, then an RL policy pushes it out. Built in MuJoCo & Three.js, running on a real AgileX Robotics arm.

For this hackathon I taught a robot to beat you in Jenga. It sims every possible move, picks the block that's the best move, then an RL policy pushes it out. Built in MuJoCo & Three.js, running on a real AgileX Robotics arm.

jonathanhawkins

39,842 次观看 • 4 个月前

Introducing Speech Engine. Developers can now turn their existing chat agent into a full voice agent with one prompt. Speech Engine combines our leading speech, transcription, and voice orchestration models into a single pipeline - all custom built to work best together.

Introducing Speech Engine. Developers can now turn their existing chat agent into a full voice agent with one prompt. Speech Engine combines our leading speech, transcription, and voice orchestration models into a single pipeline - all custom built to work best together.

ElevenLabs

135,133 次观看 • 4 个月前

A fun way to try the new Claude Fable. I asked it to train a bot using RL to navigate 3D worlds created with SpAItial AI. This is all running in the browser including RL training! Link to demo and code in comments What you can do: -> Train a bot from scratch, watch it learn in minutes, then drop it into a world it has never seen and it keeps navigating obstacles. -> The worlds are real scenes generated from a text prompt or a photo via the SpAItial AI API (you get a splat + a collision mesh). -> Mesh is made into a 2D walkability + height grid that the RL environment runs on. -> The trick to generalization: the policy only ever sees egocentric inputs. A 16-ray lidar, the goal's body frame, and its speed. No global coordinates, no map. So it learns navigation, not a memorized world.

A fun way to try the new Claude Fable. I asked it to train a bot using RL to navigate 3D worlds created with SpAItial AI. This is all running in the browser including RL training! Link to demo and code in comments What you can do: -> Train a bot from scratch, watch it learn in minutes, then drop it into a world it has never seen and it keeps navigating obstacles. -> The worlds are real scenes generated from a text prompt or a photo via the SpAItial AI API (you get a splat + a collision mesh). -> Mesh is made into a 2D walkability + height grid that the RL environment runs on. -> The trick to generalization: the policy only ever sees egocentric inputs. A 16-ray lidar, the goal's body frame, and its speed. No global coordinates, no map. So it learns navigation, not a memorized world.

RO2⚡️

12,797 次观看 • 3 个月前