My first SkyPilot PR, in Sept 2024, fixed syntax highlighting in one docs code block. 51 merged PRs later it's a company. Congrats @zongheng_yang @Michaelvll1 and the team. @tian_xia_ reviewed most of those 51, thank you!

Built by @YichuanM and @andylizf at StarTrail, out of UC Berkeley's Sky Computing Lab @BerkeleySky, BAIR @berkeley_ai, and Berkeley NLP @BerkeleyNLP. Same team behind LEANN (RAG on everything, on-device, with a 97% smaller index). Thanks to our advisors @matei_zaharia, @profjoeyg, and @sewon__min! If LEANN gave your models memory of your private data, PixelRAG gives them eyes on the open web. Code: Paper: Playground: 🐍 𝚙𝚒𝚙 𝚒𝚗𝚜𝚝𝚊𝚕𝚕 𝚙𝚒𝚡𝚎𝚕𝚛𝚊𝚐

Your coding agent reads every web page wrong. And pays 2.5x extra to do it. pixelbrowse fixes it with one screenshot. −74% tokens, 4x faster. give Claude Code eyes. 👀 New: pixelbrowse. 𝚙𝚒𝚙 𝚒𝚗𝚜𝚝𝚊𝚕𝚕 𝚙𝚒𝚡𝚎𝚕𝚛𝚊𝚐 ·

Really amazing results analyzing what's creative/novel vs. what's copied from Internet data, enabled by the amazing @liujc1998's Infini-gram! This is also enabled in @allen_ai's OlmoTrace where anyone can find matching n-grams between LLM-generated text and its training data.

Codex basically replaced OpenClaw for me at this point but the part I like most here is “memory as markdown files” pretty refreshing to hear that from someone on the Codex team. model companies usually have every reason to want memory lock-in 🙄

this customer list makes so much sense AI apps create so much stuff. it all still needs to be searched and governed. so much AI slop😂 keeping it all hot gets expensive fast serverless search on S3 feels like the right shape

this one means a lot 🥹 LEANN has been one of the most fun and intense projects I've worked on. I spent a lot of time building the system, running experiments, shaping the paper, and shipping the open source release to the community. seeing both the research recognition and community response has been incredibly special. so grateful to @YichuanM, our advisors, collaborators, and everyone who tried LEANN, opened issues, contributed PRs, or gave feedback. excited for what's next 👀

Today I shared some personal practice use cases (for research and maintaining open-source) of Claude Code at @BerkeleySky, UC Berkeley. Here is the link to the presentation: Some personal hot takes/ discussion conclusions: 1. IDEs might not be the eventual best way to interact with code. CLI-based agents require less human supervision and scale better (compared to tools like Cursor). People may eventually only review the PR rather than the intermediate coding steps (even more aggressive!). 2. Whether a problem can be solved (the boundary!) by coding agents (today) largely depends on whether there is a clear signal for the task. For example, @karpathy’s autosearch is relatively easy because the signal is clean, a very good example. (Good news for researchers: coding agents still struggle with open-ended questions. So… our jobs are probably safe for a little while.) 3. Human supervision is the bottleneck. Designing agent products that minimize human effort while leveraging humans’ advantage in context switching is a very interesting direction.

Glad to have contributed to this with @shulynnliu at Sky. By cleanly separating the evaluator and evolution mechanism, the LLM can actually optimize the search process itself. Scaling across 200+ tasks proves it works.