// Continual Learning from Experience and Skills // Skills are so good when you combine them properly with MCP & CLIs. I have found that Skills can significantly improve tool usage of my coding agents. The best way to improve them is to regularly document improvements, patterns, and things to avoid. Self-improving skills don't work that well (yet). Check out this related paper on the topic: It introduces XSkill, a dual-stream continual learning framework. Agents distill two types of reusable knowledge from past trajectories: experiences for action-level tool selection, and skills for task-level planning and workflows. Both are grounded in visual observations. During accumulation, agents compare successful and failed rollouts via cross-rollout critique to extract high-quality knowledge. During inference, they retrieve and adapt relevant experiences and skills to the current visual context. Evaluated across five benchmarks with four backbone models, XSkill consistently outperforms baselines. On Gemini-3-Flash, the average success rate jumps from 33.6% to 40.3%. Skills reduce overall tool errors from 29.9% to 16.3%. Agents that accumulate and reuse knowledge from their own trajectories get better over time without parameter updates. I have now seen two papers this week with similar ideas. Paper: https://t.co/YXrHcJ6Zim Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX

decent sunset last night, maybe like 7/10 https://t.co/6wgRobX3wt

You wake up and in the first 5 minutes you read that Trump: ▪️Obviously raped children ▪️ Started a war to distract from this rape ▪️ Is openly taking billions in bribes ▪️ Is a hard-core racist ▪️ Is destroying the world economy ▪️ Is openly colluding with Russia

Biden always has a nasty orange Gatorade in his inventory 😭😭 even Jesse Jackson’s funeral dawg give it a rest, a blue Gatorade for respect at least https://t.co/IweSYZfiIS

what whitaker sees when he opens the fridge at robby’s house for the first time

@andrewztan Scientists making more than professional athletes? A good start 😄

A small thank you to everyone using Claude: We’re doubling usage outside our peak hours for the next two weeks. https://t.co/W7TEBPditq

perfect timing to push on my research automations. let's go!

I need more movies like after hours where some guy just runs around having the worst day of his life https://t.co/IqWArXtuQd

didn’t really care for this movie when i was struggling as a twentysomething in New York but recently rewatched as a guy in his forties who lives in LA and let me just say: lol, lmao

This is such a funny thing to have a ghost do to you https://t.co/z5U9KHqFXF

@PhDPersuasion @3blue1brown I’m getting π_1 = 4, so maybe I’m using a different definition of things than the OP, but you can do this entirely in the native L^p geometry by defining a circle of radius r as the set of points an L^p distance of r from the origin,

@PhDPersuasion @3blue1brown the circumference via a path integral in the L^p geometry, and the diameter as 2r.

@3blue1brown How exactly are you defining things? I’m getting π_1 = 4, with the unit circle being what we would colloquially call a rhombus with side length 2.

@paulbz Oh I so want!

https://t.co/X0CHBm19Bc

https://t.co/X0CHBm19Bc

The new upgrade of Grok Imagine is absolutely 🔥🔥 Download @Grok and try it now. https://t.co/j309HBhSEe

Re: “empowering enough people”: I think this is already well underway!!! When we started @aixventureshq back in 2021 (before ChatGPT!), I remember feeling that the supply of people with deep knowledge of modern neural networks—and hence good founder CTOs—was very restricted. But 5 years later, through the efforts of @karpathy and many others (including myself and @jeremyphoward from the QT-ed thread) but certainly not discounting the contribution of the ever-growing-in-size frontier labs and places like the @GoogleDeepMind Gemini team, which are emitting many good people, it now feels like there is a large and robust supply of empowered deep learning experts. 🤯

I had a chance to chat with Andrej when he visited Tokyo in 2022, right after he wrapped up his 5-year work with Elon (he told me he was "recovering"). It was clear that his true passions are education and empowerment. You can see this reflected on his X feed—there is no hidden agenda, just a genuine desire to share what excites him. In a way, Andrej represents the ultimate nightmare for organizations like OpenAI. His goal is to empower developers outside of closed labs to train their own models, making AI knowledge accessible far beyond a small, elite group. If you empower enough people, closed labs will eventually run out of the capital needed to acquire all that distributed talent, and the monopoly on AI power will dissolve. This is just my speculation, but looking at it through this lens, it makes complete sense why he left OpenAI after just a year.

@darin_gordon started to use both at the same time today. probably have more to share later on. i like that this is a skill that i can easy optimize and build on top of