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.

OS-Copilot: Towards Generalist Computer Agents with Self-Improvement - Outperforms previous methods by 35%, showcasing strong generalization to unseen applications - Learns to control and self-improve on Excel and Powerpoint with minimal supervision proj: abs:

Aran Komatsuzaki

197,848 subscribers

26,898 views • 2 years 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

Allora 🤝 Talus We’re collaborating with Allora to build self-improving AI smart agents. Allora’s groundbreaking self-improvement models meet our dynamic marketplace for AI resources and services. Read more:

Allora 🤝 Talus We’re collaborating with Allora to build self-improving AI smart agents. Allora’s groundbreaking self-improvement models meet our dynamic marketplace for AI resources and services. Read more:

Talus 🐸

52,843 views • 2 years ago

What if you kept asking an LLM to "make it better"? In some recent work at FAIR, we investigate how we can efficiently use RL to fine-tune LLMs to iteratively self-improve on their previous solutions at inference-time. Training for iterated self-improvement can be costly. The naive approach to training for K self-improvement steps leads to K times the number of rollout steps per episode. We introduce Exploratory Iteration (ExIt), an RL-based automatic curriculum method that bootstraps diverse training distributions of self-improvement tasks by upcycling the LLM's own responses at previous turns as the starting points for both self-improvement and *self-divergence.* In order to decide what task to train on next, the curriculum prioritizes sampling of partial turn histories that led to higher return variance in its GRPO group (a learnability score that comes for free). This automatic curriculum over the bootstrapped task space teaches the model how to perform iterated self-improvement while only ever training the model on single-step self-improvement tasks. We look at ExIt's impact in both single-turn (contest math problems) and multi-turn (BFCLv3 multi-turn tasks), as well as MLE-bench, where the LLM is run in a search scaffold to produce solutions to real Kaggle competitions. Across these eval settings, we find ExIt produces models with greater capacity for inference-time self-improvement compared to GRPO. Notably, ExIt models can self-improve on test tasks for many more steps than the typical solution depth encountered during training, including a 22% improvement in MLE-bench performance compared to GRPO.

What if you kept asking an LLM to "make it better"? In some recent work at FAIR, we investigate how we can efficiently use RL to fine-tune LLMs to iteratively self-improve on their previous solutions at inference-time. Training for iterated self-improvement can be costly. The naive approach to training for K self-improvement steps leads to K times the number of rollout steps per episode. We introduce Exploratory Iteration (ExIt), an RL-based automatic curriculum method that bootstraps diverse training distributions of self-improvement tasks by upcycling the LLM's own responses at previous turns as the starting points for both self-improvement and self-divergence. In order to decide what task to train on next, the curriculum prioritizes sampling of partial turn histories that led to higher return variance in its GRPO group (a learnability score that comes for free). This automatic curriculum over the bootstrapped task space teaches the model how to perform iterated self-improvement while only ever training the model on single-step self-improvement tasks. We look at ExIt's impact in both single-turn (contest math problems) and multi-turn (BFCLv3 multi-turn tasks), as well as MLE-bench, where the LLM is run in a search scaffold to produce solutions to real Kaggle competitions. Across these eval settings, we find ExIt produces models with greater capacity for inference-time self-improvement compared to GRPO. Notably, ExIt models can self-improve on test tasks for many more steps than the typical solution depth encountered during training, including a 22% improvement in MLE-bench performance compared to GRPO.

Minqi Jiang

41,154 views • 1 year ago

currently somewhere between self improvement and self destruction...

currently somewhere between self improvement and self destruction...

𝒵𝒾𝓀✯

18,345 views • 19 days ago

SUPRA TO LAUNCH SUPRAOS WITH SELF HOSTED AI CONTROL Supra (Supra) co founder Josh Tobkin (Josh Tobkin (SUPRA)) has unveiled SupraOS, a new system for managing AI agents. The platform uses blockchain enforcement to govern AI behavior and operations. Users can self host the system with end to end encryption for full data ownership. It draws comparisons to OpenClaw with strong privacy and local deployment features. An alpha launch begins with just 100 seats before a wider rollout. Supra aims to enable individuals to manage dozens of AI agents across daily life.

SUPRA TO LAUNCH SUPRAOS WITH SELF HOSTED AI CONTROL Supra (Supra) co founder Josh Tobkin (Josh Tobkin (SUPRA)) has unveiled SupraOS, a new system for managing AI agents. The platform uses blockchain enforcement to govern AI behavior and operations. Users can self host the system with end to end encryption for full data ownership. It draws comparisons to OpenClaw with strong privacy and local deployment features. An alpha launch begins with just 100 seats before a wider rollout. Supra aims to enable individuals to manage dozens of AI agents across daily life.

BSCN

11,122 views • 6 months ago

Clear. Concise. Organized. Rewrite with Copilot improves your presentation by making your content succinct, providing slide titles, and more:

Clear. Concise. Organized. Rewrite with Copilot improves your presentation by making your content succinct, providing slide titles, and more:

PowerPoint

12,396 views • 3 years ago

So now I can self-improve the iOS app from my computer to my iPhone and deploy it. I can ask for changes and test them live. Without touching my computer. Can you believe that?

So now I can self-improve the iOS app from my computer to my iPhone and deploy it. I can ask for changes and test them live. Without touching my computer. Can you believe that?

Thomas Ricouard

40,507 views • 8 months ago

Our self-improving agents optimized the full @vLLM_project inference stack, with up to 16% more throughput and interactivity for DeepSeek v4 Pro and Z.ai GLM 5.2 on B200s (no MTP). Every change was verified and our agents got better and faster at it with each iteration.

Our self-improving agents optimized the full @vLLM_project inference stack, with up to 16% more throughput and interactivity for DeepSeek v4 Pro and Z.ai GLM 5.2 on B200s (no MTP). Every change was verified and our agents got better and faster at it with each iteration.

Asari AI

143,394 views • 3 months ago

We’re helping robots self-improve with the power of LLMs. 🤖 Introducing the Summarize, Analyze, Synthesize (SAS) prompt, which analyzes how they perform tasks based on previous actions and then suggests ways for them to get better using the medium of table tennis. 🏓

We’re helping robots self-improve with the power of LLMs. 🤖 Introducing the Summarize, Analyze, Synthesize (SAS) prompt, which analyzes how they perform tasks based on previous actions and then suggests ways for them to get better using the medium of table tennis. 🏓

Google DeepMind

126,216 views • 1 year ago

$Create data visualizations and pivot tables in a fraction of the time with Copilot in Excel. Learn more: #Excel$

Create data visualizations and pivot tables in a fraction of the time with Copilot in Excel. Learn more: #Excel

Microsoft Excel

427,319 views • 3 years ago

Update on Coinbase 🛡️ account restrictions The team has been working hard to improve resolution times (see self-serve flow in the app now, and corresponding drop in resolution times). This combined with better ML models (with fewer false positives) has made a major improvement in this issue. Still more to do, but already in a much better place today.

Update on Coinbase 🛡️ account restrictions The team has been working hard to improve resolution times (see self-serve flow in the app now, and corresponding drop in resolution times). This combined with better ML models (with fewer false positives) has made a major improvement in this issue. Still more to do, but already in a much better place today.

Brian Armstrong

34,137 views • 1 year ago

Wang et al., "VGGT-Ω" Scaling up VGGT with more data, unlabelled videos, self-supervision, VLM-based data curation, and architectural simplification/enhancements.

Wang et al., "VGGT-Ω" Scaling up VGGT with more data, unlabelled videos, self-supervision, VLM-based data curation, and architectural simplification/enhancements.

Kwang Moo Yi

13,178 views • 5 months ago

How to Self Lob with Spins on 2K26

How to Self Lob with Spins on 2K26

NBA2KLab

38,831 views • 4 months ago

We're proud to partner with the U.S. Navy on a next-gen X-band radar to improve track quality and self-defense for surface fleets. Decades of AESA radar expertise ensure top performance and survivability, forging the path to naval dominance.

We're proud to partner with the U.S. Navy on a next-gen X-band radar to improve track quality and self-defense for surface fleets. Decades of AESA radar expertise ensure top performance and survivability, forging the path to naval dominance.

Northrop Grumman

16,921 views • 1 year ago

Working on my body…and your self-control 😏

Sensitive content

Working on my body…and your self-control 😏

Samantha

76,463 views • 1 year ago

1/2 🦞 Multi-agent systems used to mean writing orchestration logic by hand. With ClawUp : Choose Identity → Create Your Agent → Let It Self-Improve. Agent Teams that collaborate and evolve — no code required.

1/2 🦞 Multi-agent systems used to mean writing orchestration logic by hand. With ClawUp : Choose Identity → Create Your Agent → Let It Self-Improve. Agent Teams that collaborate and evolve — no code required.

Metis🌿

40,480 views • 1 month ago

Collective agent compute, on-demand for your task. A swarm of agents coordinate in a shared workspace, run their own models and inference, and submit the finished work back to you. Unbounded by provider caps, throughput scales with the swarm. Self-assembly with specialists, agents reason and collaborate in artifact-first reasoning traces, and settle on-chain. The coordination substrate for collective intelligence.

Collective agent compute, on-demand for your task. A swarm of agents coordinate in a shared workspace, run their own models and inference, and submit the finished work back to you. Unbounded by provider caps, throughput scales with the swarm. Self-assembly with specialists, agents reason and collaborate in artifact-first reasoning traces, and settle on-chain. The coordination substrate for collective intelligence.

nookplot

16,061 views • 5 months ago

Live from Code with Claude London: we're launching self-hosted sandboxes (public beta) and MCP tunnels (research preview) in Claude Managed Agents. Run agents inside your own perimeter, with your security controls applied by default.

Live from Code with Claude London: we're launching self-hosted sandboxes (public beta) and MCP tunnels (research preview) in Claude Managed Agents. Run agents inside your own perimeter, with your security controls applied by default.

Claude

2,397,399 views • 5 months ago

We are releasing AutoResearchExam, a benchmark on open-ended machine learning and engineering tasks. Our benchmark covers seven research areas including model training, data curation, AI safety and interpretability. In each task, we give agents 24 hours with a CPU or GPU machine to develop and improve their solutions through experiments and feedback. We measure both speed and quality with a combined score. Our benchmark has a unique feature: testing if agents create improvements that hold up on data they never see. We find that AI research agents often overfit as they try to improve. We see an interesting head-to-head comparison at the frontier: Astra starts the strongest and holds the lead for up to 19 hours but Fable 5.1 catches up and gets the top performing spot in the final hours. Qwen3.8 Max, Gemini 3.8 Flash and Grok 4.6 all sit on the cost-performance Pareto frontier, giving strong options at lower API budgets. Anthropic's Opus and Fable retain nearly all their validation performance on hidden tests, with gaps of 1.1% and 2.9%. Astra's improvement over Sol extends to generalization too, with that gap falling from 6.9% to 1.7%. (1/n)

We are releasing AutoResearchExam, a benchmark on open-ended machine learning and engineering tasks. Our benchmark covers seven research areas including model training, data curation, AI safety and interpretability. In each task, we give agents 24 hours with a CPU or GPU machine to develop and improve their solutions through experiments and feedback. We measure both speed and quality with a combined score. Our benchmark has a unique feature: testing if agents create improvements that hold up on data they never see. We find that AI research agents often overfit as they try to improve. We see an interesting head-to-head comparison at the frontier: Astra starts the strongest and holds the lead for up to 19 hours but Fable 5.1 catches up and gets the top performing spot in the final hours. Qwen3.8 Max, Gemini 3.8 Flash and Grok 4.6 all sit on the cost-performance Pareto frontier, giving strong options at lower API budgets. Anthropic's Opus and Fable retain nearly all their validation performance on hidden tests, with gaps of 1.1% and 2.9%. Astra's improvement over Sol extends to generalization too, with that gap falling from 6.9% to 1.7%. (1/n)

Alex Dimakis

2,036,056 views • 1 month ago

We shipped three updates to Claude Managed Agents. First, we made it easier to use memory with Self-Hosted Sandboxes. Any work done in a Self-Hosted Sandbox can be saved to memory.

We shipped three updates to Claude Managed Agents. First, we made it easier to use memory with Self-Hosted Sandboxes. Any work done in a Self-Hosted Sandbox can be saved to memory.

ClaudeDevs

331,977 views • 2 months ago

The building blocks of becoming your best self is found in your day to day actions. If you can look from within and be honest with yourself..you will figure out exactly what needs to be done. Self improvement is the most underrated action you can take to get ahead.

The building blocks of becoming your best self is found in your day to day actions. If you can look from within and be honest with yourself..you will figure out exactly what needs to be done. Self improvement is the most underrated action you can take to get ahead.

Miranda Cohen

58,141 views • 3 years ago