AIDE has stood the test of time as the leading ML engineering agent, showing strong potential to automate data science modeling, deep learning, and AI R&D. Today, we’re sharing more details to help the community better understand its design, and build on top of it:🧵 (1/N)

🔗 Papers & code. Speedrunning Benchmark → AIRA agents → AIDE ML, our reference implementation of AIDE (MIT) →

Frontier labs keep building on the same foundation: our open source work, AIDE. Last week along, two papers from @AIatMeta related to AIDE:. 1️⃣ Automated LLM Speedrunning Benchmark.2️⃣ Thorough ablations and improvements to AIDE on MLE-Bench. Links below: 👇

RT @YuxiangJWu: Thrilled to see @WecoAI's AIDE used in Meta's work and big congrats to @MinqiJiang @BingchenZhao @MarlaMagka. It's a truly….

RT @zhengyaojiang: Solid work from @AIatMeta on ablating and improving AIDE on MLE-Bench!. The rigor of empirical evaluation has reached a….

To dive deeper, check out our paper: or explore the code on GitHub: .Excited to see what the community builds with AIDE! 🚀.

OpenAI's MLE-Bench shows that o1-preview + AIDE excels at ML engineering, but how much does AIDE contribute?. Our tests confirm that AIDE boosts performance 3.5x over o1-preview alone. (4/N).

We’re sharing more details on our internal Kaggle benchmark. Unlike MLE-Bench, we made actual Kaggle submissions whenever possible. Here, we provide insights into the benchmark setup, key results, and the limitations of our evaluation protocol. (3/N)

We break down AIDE's algorithm to highlight its design philosophy that enables interaction scaling. AIDE uses a systematic tree search, iteratively refining solutions with improvements or bug fixes while evaluating performance at each step. (2/N)

RT @junshernchan: 🏃‍♀️ MLE-bench Lite 🪶 The most common request we get on MLE-bench is to have a “Lite” version that is cheaper to run, and….

Handwriting Calculator.Typing math symbols is a pain- why not just draw them? Our AI can handle advanced math, from basic arithmetic to integrals. Try it here: (3/N)

Easy Deployment & Spreadsheet Integration.Use our single-line deploy approach or integrate Weco AI Functions directly in Google Sheets—batch-process spreadsheet data with ease. Get the add-on: (2/N)

🚀 AIDE just got better! Our ML engineering agent now has a local Web UI:.✨ Visual interface for seamless ML experiments.📊 Better tracking & progress monitoring.🔒 Fully local—no data uploads required. Try it now:

We’re looking for a frontend engineer to join us in building interfaces that deliver AI features with a prompt!. Time zone preference: US East or UK.Location: Flexibility to relocate to the Bay Area is preferred. DM this account or @zhengyaojiang if you're interested.

Excited to see OpenAI's recent project, MLE-bench, is based on our open-source effort, AIDE. In their independent evaluation, AIDE surpasses other MLE agents by a large margin!

Excited to try it out? We're starting our closed alpha test. Join the waitlist for early access!.(5/N).

Our solution abstracts away these complexities. Instead of writing intricate code, you interact with AI capabilities as if calling a strongly-typed remote function. For the same news sentiment feature, you only have to write:.(4/N)

Why did we create this? Current LLM APIs often present challenges:. • Very complex to enforce strongly-typed outputs.• Often requires domain knowledge of AI.• Heavy-weighted, chatbot-based interface. To implement a news sentiment analysis function you'll have to write:.(3/N)

We've written a blog post explaining the concept in detail. Check it out here:.(2/N)

Introducing the AI Function Builder: A developer-friendly way to integrate LLM capabilities into your software. Describe your needs in natural language and get a dedicated endpoint in seconds. Call it with our python client as if calling a strongly-typed function. (1/N)