Jeff Willette

Our Agent Distillation paper is accepted at #NeurIPS2025 Spotlight! 🚀 Turn your small LM into a strong agent 💪 Code: github.com/Nardien/agent-…

We propose to first perform query-sparse attention and combine the output with a generic key-sparse attn. method. This drastically boosts the performance of sparse attention while adding minimal overhead, and its dead simple to implement on pretrained models.

Most sparse attention methods focus on key-sparse attention. We found that key-sparse attention causes a distributional shift in outputs. Therefore, even if you switch to dense attention during decode, the queries and the keys might no longer match...

Noice! Our paper "Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction" has been accepted to NeurIPS 2025! See you in San Diego (See part 2 of post for breakdown of our work) arxiv.org/pdf/2505.11254

So excited to share that five papers have been accepted to #ACL2025 🎉 Huge thanks to all my amazing collaborators. I am especially grateful that all the internship projects (which I have worked on during my PhD journey) have all found their way to publication. Its rewarding. 😊

There is a HUGE problem with sparse attention! We found it causes a misalignment of queries and keys, so even if you add a dense decode/generation phase, the sparsely encoded context can be forgotten. But wait... We can fix it with a simple correction! huggingface.co/papers/2505.11…

@netflix I applied for an internship and got an email asking me to fill out a form on lever.co. the link in the email is broken. Your support and lever.co support both said they can't help. What else can I do? There is no one else to contact.

@mhartington Ok Thanks! I've got a hard deadline of March 1st to submit my materials, so if you have some time before then I would be forever grateful.

@delta_skelta Saw the email. Was traveling last week, will reply!

@mhartington can you check your DM's? I sent an inquiry related to nvim typescript last week.

@gatsbyjs @kylemathews @kylemathews can you check your dm's?

@delta_skelta Yeah dm @kylemathews

@gatsbyjs I have a bit of an unusual (but simple) request related to my contributions to the project, any chance I can speak to someone about it via DM or email?

@gatsbyjs @kylemathews I sent him a dm about a week ago and haven't heard back yet.

@mhartington Cool. Yeah I see all the files in question being watched but still no fixes. Ever seen anything like that before?

@delta_skelta You _can_ use `echo TSGetProjectInfoFunc()` or something similar to output all the files tsserver watches, which in theory would provide the symbols.

@mhartington I am debugging nvim_typescript and why :TSImport is giving me no candidates. I'm down to sending raw requests to the TSServer. Is there a better way to debug this? TSServer isn't doc'd well...

@mhartington Cool. Yeah I see the files in question are being watched but still no imports. Ever seen anything like that before?

@mhartington Yeah im already logging it all from that wiki. Dobyou know if any of the commands to the server can dump all exported symbols?

@delta_skelta TSImport isn't always reliable, but `TSGetCodeFix` is a bit better. I'd checkout github.com/Microsoft/Type… for some details on how to log it. Most of the time it means that the project doesnt export proper types, or that tsserver cant resolve a symbol. This rarely happens though

I just published “Tern: Why it Breaks and How To Fix it” medium.com/p/tern-why-it-…

@flowtype can the autocomplete CLI complete methods and identifiers from non flowtyped files? (without type information of course) stackoverflow.com/questions/4757…

Monitor and get simple alerts on docker container metrics, now from within docker as well! github.com/deltaskelta/do… #Docker #dockercon

@mwitkow does grpc-proxy allow regular gRPC calls as well as grpc-web? or do they have to be on different servers?