Can AI actually code for us? 🧵. MIT research reveals there’s a "long way to go" due to bottlenecks like assessment, codebase scale, & incorrect retrievals. The work reflects a vision to let humans focus on high-level design while routine work is automated:

RT @MIT_CSAIL: How X looked when it launched, 19 years ago this month. It was originally branded as "twttr" after co-founder Noah Glass ca….

RT @MITOCW: We are so excited to share that 6.1200J Mathematics for Computer Science is now ready and waiting for you! Watch the videos on….

How X looked when it launched, 19 years ago this month. It was originally branded as "twttr" after co-founder Noah Glass came across the word in a dictionary:

Authors: Alex Gu, Naman Jain, Wen-Ding Li, Manish Shetty, Yijia Shao, Ziyang Li, Diyi Yang, Kevin Ellis, Koushik Sen, & Armando Solar-Lezama. Paper:

Finally, we lay out a vision for ways we can move forward as a field. We discuss the necessity of large-scale community-driven efforts to collect open datasets & train open models for code. We also discuss techniques such as designing RL environments, training LLMs to program.

Effective interaction w/programmers is also important for coding agents, especially when user prompts can be vague. For example, a GitHub issue, astropy-#14181, requests support for a new file format. Models successfully added a read method, but they omitted a write method.

Dealing w/scale is also difficult for today’s coding models. For example, it often fails when scopes are large, such as debugging by wading through large JSON dumps from observability tools like Datadog. It also requires long-horizon planning abilities: Designing a database for.

MIT PhD student Alex Gu’s perspective on his team’s findings:. In this paper, we explore what the future of AI for code looks like. We show that today’s AI systems fail in 4 ways: data, scale, interaction, & measurement. When it comes to data, we find LLMs can fail dramatically

A new handheld interface from MIT gives anyone the ability to train a robot for tasks in fields like manufacturing. The versatile tool can teach a robot new skills using one of three approaches: natural teaching, kinesthetic training, & teleoperation:

56 years ago today, humans walked on the moon for the first time. Original @TIME report of the monumental event, from 1969:

"Programming is not about typing, it's about thinking." — Rich Hickey. v/@CodeWisdom.

@goyal__pramod Authors: @PreetumNakkiran, @ArwenBradley, @oh_that_hat, & @advani_madhu.

A step-by-step guide to diffusion models: v/@goyal__pramod

The Fibonacci sequence in Pascal's Triangle, v/@TheMathFlow.

MIT CSAIL's "PhysicsGen" system helps robots handle items efficiently by customizing & multiplying the data they train on. It turns a few VR demonstrations into thousands of simulations, potentially helping build huge datasets for dexterous robots:

Hello world! This is @nishanthkumar23 and I'll be taking over @MIT_CSAIL 's X/Twitter & Instagram for 24 hours! I'm a 4th year PhD @MITEECS working on AI/ML for Robotics and Computer Agents. Drop any and all questions about research, AI, MIT, or dogs (esp. robot dogs!) below 👇

A new approach from CSAIL & Google marks a shift toward teaching models to orchestrate their own parallel decoding strategy. The team's "Parallel Structure Annotation" (PASTA) enables LLMs to generate text in parallel, accelerating their response times:

A guide to competitive programming:

RT @RussTedrake: TRI's latest Large Behavior Model (LBM) paper landed on arxiv last night! Check out our project website: .