Pinned
Mengdi Chen
1,376 posts
Mengdi Chen
@mengdi_en
🧵1/6: Today was my last day on the React team, because I was laid off by Meta. As I reflect on my 3.5 years with the company, I feel immense gratitude for the talented, friendly, and warm-hearted colleagues I had the privilege of working alongside. I've learned so much from them- Replying to @dai_shiThis is said but totally expected. After Meta laid off Douglas Armstrong I became the only maintainer of Recoil. Then they laid off me too. Although tons of internal projects still use it, no one wanted to take over the responsibility (that doesn't count in perf review)
- I'm really grateful for people who try upgrading to React 18 ⚛️ in the first week of release and notify the library maintainers about compatibility issues (some even submit PRs!). Y'all open source heroes! ❤️
- Thrilled to open-source Fuji-Web, the vision-LLM-based web agent project I started in Nov 2023 when GPT-4V was just released. I'm so proud the interesting idea has been turned into a state-of-the-art Web Agent. (You can find benchmarks in the blog post.) We started this researchAI agents are transforming how we interact with computers and the digital world. Powered by the latest development of Large Language Models (LLMs), we believe in the near future, AI agents will become highly capable personal assistants that can streamline our daily tasks,
00:00 - Fully automate your browser! Introducing WebWand, the first ever AI web agent powered by multi-modal model (and it actually works!) #GPT4V
00:00 - Replying to @mengdi_en🧵4/6: My biggest regret is not being able to complete the DevTools backend refactoring project, which aimed to eliminate painful blockers for the React core team. I was close to getting it done, and I hope to find a way to help get it done even after leaving Meta.
- Replying to @mengdi_en🧵2/6: My journey at Meta began with an internal data visualization tool, followed by becoming a maintainer of the state management library @recoiljs. I'm grateful for the opportunity to have contributed to its development and growth.
- Replying to @mengdi_en🧵6/6: Thank you to everyone who has been part of my journey at Meta. Your support and camaraderie have meant a lot to me. As I look to the future, I'm grateful for the experiences I've had and the opportunities that lie ahead in the tech industry.
- Replying to @mengdi_en🧵5/6: As I embark on a short break to recharge and plan my next adventure, I will explore new engineering opportunities (preferably in NYC). If you know of any positions or projects that align with my skills and experience, please don't hesitate to reach out.
- Replying to @mengdi_en and @recoiljs🧵3/6: I was then given a dream opportunity to join the React team (x.com/mengdi_en/stat…). Diving deep into React's internals and DevTools complexity has been an incredibly rewarding experience. I'm proud to have tackled long-standing problems, some deemed impossible.Dreams do come true 🥳 beta.reactjs.org/community/team…
- Per Thomas's suggestion, I ran this experiment with @google Gemini Pro 1.5, the LLM with longest context as of today. I grabbed a lengthy gov document, made a copy with 10 random text edits. The result is... 100% hallucination, unfortunately.
Has anyone tried this "haystack" test for long context LLMs? 1. Take a large (50k+ tokens) document. 2. Make 100 random small edits. 3. Append the documents and give put them in the context. 4. Ask the LLM to find the differences. In contrast with typical "lost in the middle"
