Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

The most useful data comes from real user interactions, feedback, and traces. With self-distillation, your model can learn from production traces even when the original environment can’t be replayed.

Applied Compute

7,748 subscribers

12,254 views • 13 days ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Today we're launching agent feedback! The best way to power agent learning is with real user reactions. Feedback stores thumbs, ratings, and comments next to traces and metrics. Then query them directly or visualize:

Today we're launching agent feedback! The best way to power agent learning is with real user reactions. Feedback stores thumbs, ratings, and comments next to traces and metrics. Then query them directly or visualize:

Sam Bhagwat

10,725 views • 1 month ago

Introducing 𝙾𝚙𝚎𝚗 𝚃𝚎𝚊𝚌𝚑 Better Claude Teach that works with any agent, any model, and with your own data. Features: -> Embed Open Teach into your own custom agents -> Learn from in-browser activity and from agent interactions -> Self-hosted for full data sovereignty -> Future proof and cross-stack Get early access:

Introducing 𝙾𝚙𝚎𝚗 𝚃𝚎𝚊𝚌𝚑 Better Claude Teach that works with any agent, any model, and with your own data. Features: -> Embed Open Teach into your own custom agents -> Learn from in-browser activity and from agent interactions -> Self-hosted for full data sovereignty -> Future proof and cross-stack Get early access:

Atai Barkai

38,367 views • 1 month ago

📊We’ve rebuilt Dashboards in LangSmith to make them more useful for investigation and reporting. Put KPIs next to trends, compare metrics with different units, break traces down by model or user, add notes, and arrange the dashboard freely.

📊We’ve rebuilt Dashboards in LangSmith to make them more useful for investigation and reporting. Put KPIs next to trends, compare metrics with different units, break traces down by model or user, add notes, and arrange the dashboard freely.

LangChain

15,969 views • 26 days ago

Well, even if you don’t care about your data going to China, I think it’s worth caring about not using a censored model that the DeepSeek app serves. And that’s why it’s worth using the R1 model on Perplexity with web search grounding and reasoning traces.

Well, even if you don’t care about your data going to China, I think it’s worth caring about not using a censored model that the DeepSeek app serves. And that’s why it’s worth using the R1 model on Perplexity with web search grounding and reasoning traces.

Aravind Srinivas

225,806 views • 1 year ago

Excited to open source our world-model-harness! `wmh` makes it easy to go from agent traces -> faithful replication of your production environment Basically, an LLM pretends to be a Docker container but 5x faster Below is a comparison running 8 SWE-bench tasks

Excited to open source our world-model-harness! `wmh` makes it easy to go from agent traces -> faithful replication of your production environment Basically, an LLM pretends to be a Docker container but 5x faster Below is a comparison running 8 SWE-bench tasks

Silen Naihin

14,085 views • 2 months ago

📢 Lyra: Generative 3D Scene Reconstruction via Video Diffusion Model Self-Distillation Got only one or a few images and wondering if recovering the 3D environment is a reconstruction or generation problem? Why not do it with a generative reconstruction model! We show that a camera-conditioned video diffusion model can be transformed into a generative reconstruction model that directly outputs a high-quality 3D Gaussian Splatting representation through self-distillation, without requiring real-world training data. Check out our results in the video (wait for dynamic scenes in the second half!) : Project Page: Code and Models: Paper:

📢 Lyra: Generative 3D Scene Reconstruction via Video Diffusion Model Self-Distillation Got only one or a few images and wondering if recovering the 3D environment is a reconstruction or generation problem? Why not do it with a generative reconstruction model! We show that a camera-conditioned video diffusion model can be transformed into a generative reconstruction model that directly outputs a high-quality 3D Gaussian Splatting representation through self-distillation, without requiring real-world training data. Check out our results in the video (wait for dynamic scenes in the second half!) : Project Page: Code and Models: Paper:

Sherwin Bahmani

66,804 views • 11 months ago

I had an awesome time interviewing idan shenfeld and Jonas Hübotter from MIT and ETH Zurich about self-distillation. this very promising post-training paradigm where the model acts as its own teacher by conditioning on environment feedback or demonstrations. we cover the SDPO algo for reinforcement learning with rich feedback and SDFT for continual learning without forgetting along with many applications. we dig into how it works, why it's simpler and faster than GRPO, and where this is already showing up in production systems. table of content: 0:00 - what is self distillation 2:50 - idan (MIT) and jonas (ETH Zurich) introduction and motivation 18:40 - different perspective of on-policy self-distillation (presentation) 36:00 - metacognition and specificity in self-distillation 37:24 - very long hard task and self-distillation 42:00 - continual learning with self-distillation (presentation) 1:16:50 - what is next in this research direction? 1:20:00 - is there any experience with subjective feedbacks? 1:22:50 - quality vs number of feedbacks? 1:26:40 - what setting would self-distillation struggle vs GRPO? the slides were super crisp really cool of them to share! enjoy my guys 🌹

I had an awesome time interviewing idan shenfeld and Jonas Hübotter from MIT and ETH Zurich about self-distillation. this very promising post-training paradigm where the model acts as its own teacher by conditioning on environment feedback or demonstrations. we cover the SDPO algo for reinforcement learning with rich feedback and SDFT for continual learning without forgetting along with many applications. we dig into how it works, why it's simpler and faster than GRPO, and where this is already showing up in production systems. table of content: 0:00 - what is self distillation 2:50 - idan (MIT) and jonas (ETH Zurich) introduction and motivation 18:40 - different perspective of on-policy self-distillation (presentation) 36:00 - metacognition and specificity in self-distillation 37:24 - very long hard task and self-distillation 42:00 - continual learning with self-distillation (presentation) 1:16:50 - what is next in this research direction? 1:20:00 - is there any experience with subjective feedbacks? 1:22:50 - quality vs number of feedbacks? 1:26:40 - what setting would self-distillation struggle vs GRPO? the slides were super crisp really cool of them to share! enjoy my guys 🌹

Yacine Mahdid

13,091 views • 4 months ago

"You don't learn the standards from the staff. You learn 'the expected standards' from the staff. You learn the real standards from each other & the way you interact. The standards show up in your interactions & the experience you have with each other-that's the real culture."

"You don't learn the standards from the staff. You learn 'the expected standards' from the staff. You learn the real standards from each other & the way you interact. The standards show up in your interactions & the experience you have with each other-that's the real culture."

The Winning Difference

234,256 views • 3 years ago

Over the last few months, we’ve been thinking about how to learn from “off-domain” data - data from non-robot sources like video or simulation. These data sources are not quite good enough to learn policies (even monolithic VLA models) directly, but they still contain lots of information that can be useful for generalizable robot control. How can we develop robot learning models that are able to make use of this type of data for generalizable control? In new work, that we call HAMSTER, we show that VLMs can be useful for enabling robotic learning from off-domain data, but specifically when used through hierarchical VLA architectures. We show that this class of models can learn generalizable robot policies for the real world from large-scale, off-domain data. A 🧵 (1/10)

Over the last few months, we’ve been thinking about how to learn from “off-domain” data - data from non-robot sources like video or simulation. These data sources are not quite good enough to learn policies (even monolithic VLA models) directly, but they still contain lots of information that can be useful for generalizable robot control. How can we develop robot learning models that are able to make use of this type of data for generalizable control? In new work, that we call HAMSTER, we show that VLMs can be useful for enabling robotic learning from off-domain data, but specifically when used through hierarchical VLA architectures. We show that this class of models can learn generalizable robot policies for the real world from large-scale, off-domain data. A 🧵 (1/10)

Abhishek Gupta

11,994 views • 1 year ago

Introducing the DIA Oracle Playground A developer testing environment where you can get familiar with DIA oracles, experiment with integration code, and test real-time price feeds before production deployment. Learn by doing, with live oracle data. 🧵

Introducing the DIA Oracle Playground A developer testing environment where you can get familiar with DIA oracles, experiment with integration code, and test real-time price feeds before production deployment. Learn by doing, with live oracle data. 🧵

DIA Oracles

20,264 views • 1 year ago

I think about this idea from MrBeast a lot. Community can be an unfair advantage when you find people to share with and learn from. When you’re operating solo, you just can’t compete with someone who is learning from the experiences of several others in real time.

I think about this idea from MrBeast a lot. Community can be an unfair advantage when you find people to share with and learn from. When you’re operating solo, you just can’t compete with someone who is learning from the experiences of several others in real time.

Jay Clouse

636,505 views • 3 years ago

You probably have thousands of agent traces sitting in production. And most of them are doing nothing for you. Today we’re launching the new Kitaru to change that. Kitaru ingests traces from Langfuse, Braintrust, or any OTel source, investigates what went right and wrong, groups recurring failures into cohorts, and builds evaluators around them. Then you can replay those same production cases against a different model, prompt, context, or agent setup and ask: What would have happened if we changed this? Instead of traces being something you inspect after a failure, they become part of a continuous improvement loop. Open source. Free to use. Try it out today:

You probably have thousands of agent traces sitting in production. And most of them are doing nothing for you. Today we’re launching the new Kitaru to change that. Kitaru ingests traces from Langfuse, Braintrust, or any OTel source, investigates what went right and wrong, groups recurring failures into cohorts, and builds evaluators around them. Then you can replay those same production cases against a different model, prompt, context, or agent setup and ask: What would have happened if we changed this? Instead of traces being something you inspect after a failure, they become part of a continuous improvement loop. Open source. Free to use. Try it out today:

Hamza Tahir

22,264 views • 20 days ago

PlaqueBoyMax lists the most iconic moments that helped him blow up and says the “plant” narrative comes from self doubtful people who project hate when you can see his evolution from 2022

PlaqueBoyMax lists the most iconic moments that helped him blow up and says the “plant” narrative comes from self doubtful people who project hate when you can see his evolution from 2022

PlaqueBoyMaxUpdates

28,551 views • 9 months ago

Introducing the OpenRouter MCP, live model intelligence right inside your agent Your agent builds and ships, but when it comes to choosing the right model for the right job, it guesses from 6 month old training data Watch it pick, price, and test the right model:

Introducing the OpenRouter MCP, live model intelligence right inside your agent Your agent builds and ships, but when it comes to choosing the right model for the right job, it guesses from 6 month old training data Watch it pick, price, and test the right model:

OpenRouter

264,567 views • 2 months ago

Introducing W&B Skills for coding agents! Watch us install the skill, point it at a live fine-tuning project with thousands of traces and RL training runs, then query it all from the terminal in seconds. Also works with W&B Weave to pull in your agent traces and spot failures.

Introducing W&B Skills for coding agents! Watch us install the skill, point it at a live fine-tuning project with thousands of traces and RL training runs, then query it all from the terminal in seconds. Also works with W&B Weave to pull in your agent traces and spot failures.

Weights & Biases

264,792 views • 6 months ago

User feedback is the most important thing. With that motto, Pixquare just added 2 new enhancements based on user feedback (available soon): - Custom objects can be long-pressed + dragged to add - Hany joystick can toggle/select tools Artwork by @ArtNature_it and 𝐜𝐡𝐢𝐥𝐢 ❀»

User feedback is the most important thing. With that motto, Pixquare just added 2 new enhancements based on user feedback (available soon): - Custom objects can be long-pressed + dragged to add - Hany joystick can toggle/select tools Artwork by @ArtNature_it and 𝐜𝐡𝐢𝐥𝐢 ❀»

Pixquare - Pixel Art Editor

10,159 views • 2 years ago

Im done with Qwen3.8 27B 😎 Atlas covers it completely now: all 1,199 tensors, the full architecture, and what happens inside the model while it runs 🚀 The final part is the Living Model. I ran English, code and agent traces through Qwen and mapped signal flow across all 64 layers, activation quantization, every full attention layer, linear memory, 1.1M neurons, Vision, and the damage from swapping each layer to INT4. Everything comes from the actual weights or real runs. No placeholder data anywhere. Dark mode and a proper mobile view are in too. It finally feels like a product.

Im done with Qwen3.8 27B 😎 Atlas covers it completely now: all 1,199 tensors, the full architecture, and what happens inside the model while it runs 🚀 The final part is the Living Model. I ran English, code and agent traces through Qwen and mapped signal flow across all 64 layers, activation quantization, every full attention layer, linear memory, 1.1M neurons, Vision, and the damage from swapping each layer to INT4. Everything comes from the actual weights or real runs. No placeholder data anywhere. Dark mode and a proper mobile view are in too. It finally feels like a product.

Alexey Fateev

31,908 views • 6 days ago

Inferact CEO Simon Mo says open and closed models are converging. The moat is environment: "What really differentiates open-weight models from closed-weight models? In the end, there's not much differentiation. It's more about the distribution strategy and go-to-market strategy. Capability-wise, I don't really see a big gap, not even today." "One of the most important parts is just the data. It's about who gets what data, and then what are the environments you are building to let the model improve on itself." "For Moonshot, they have built some of the best environments for front-end coding... the ability for this model to code and then see what the rendered code is, and then continue looping in this iterative process." "Now, this is about their environment to improve the model. It's not about just source data. It's not about where they get the data from. Rather, it's who can build the best environment and who can make the most optimization and algorithmic choices to leverage all this learning from this environment." "The next year is all gonna be about that: how open-weight model labs are differentiating and really getting the model to meet the real world." Simon Mo

Inferact CEO Simon Mo says open and closed models are converging. The moat is environment: "What really differentiates open-weight models from closed-weight models? In the end, there's not much differentiation. It's more about the distribution strategy and go-to-market strategy. Capability-wise, I don't really see a big gap, not even today." "One of the most important parts is just the data. It's about who gets what data, and then what are the environments you are building to let the model improve on itself." "For Moonshot, they have built some of the best environments for front-end coding... the ability for this model to code and then see what the rendered code is, and then continue looping in this iterative process." "Now, this is about their environment to improve the model. It's not about just source data. It's not about where they get the data from. Rather, it's who can build the best environment and who can make the most optimization and algorithmic choices to leverage all this learning from this environment." "The next year is all gonna be about that: how open-weight model labs are differentiating and really getting the model to meet the real world." Simon Mo

a16z

39,424 views • 27 days ago

Sharing an OS tool I have been building for a few weeks: - git for agent traces. With OT, you have a local client that helps you santise, review and securely share traces from your coding sessions, privately or publicly, as a Hugging Face dataset. 0.3 adds experimental session blame which I am very excited about - attributing commit/PR changes back to the traces that produced them (all local) unlocks a new layer in observability and issue tracking. A few short demos of the local TUI client and the upcoming opentraces hub platform. Try it out! keen on your feedback🙏

Sharing an OS tool I have been building for a few weeks: - git for agent traces. With OT, you have a local client that helps you santise, review and securely share traces from your coding sessions, privately or publicly, as a Hugging Face dataset. 0.3 adds experimental session blame which I am very excited about - attributing commit/PR changes back to the traces that produced them (all local) unlocks a new layer in observability and issue tracking. A few short demos of the local TUI client and the upcoming opentraces hub platform. Try it out! keen on your feedback🙏

Gabriele Farei

10,565 views • 4 months ago

When I get a new device, I don't restore from backup. A new device means a fresh start: better privacy, stronger security, and no digital clutter. You’d be shocked at the digital traces that follow when you transfer your setup.

When I get a new device, I don't restore from backup. A new device means a fresh start: better privacy, stronger security, and no digital clutter. You’d be shocked at the digital traces that follow when you transfer your setup.

Naomi Brockwell priv/acc

26,730 views • 4 months ago