“exam maxxing kumon monsters” is sending me lmfao 🤣

This is a real rebuttal from the papers I am reviewing at NeurIPS... the reviewer who wrote this gave the paper a 2 (reject) with a confidence of 5 (absolutely certain & familiar with all related work, the highest confidence available). What the hell are we even doing here...?

My team at Stack just open sourced our runtime shape checking library for PyTorch & Numpy. It's hugely inspired by @patrickkidger.bsky.social's Jaxtyping but fixes a few key issues we ran into (namely, compatibility with torch.jit.trace & torch.compile). Try it with `pip install dltype`

JL Lemma mentioned! 😍

On my way to @cvprconference.bsky.social! Catch our presentation on the 2025 Argoverse challenges at the Workshop on Autonomous Driving tomorrow at 11:30am in Grand C1!

Once again I am pleading with researchers. DO NOT HOST YOUR DATASETS ON GOOGLE DRIVE. Downloading 100s of GB from a web browser with rate limits is a recipe for disaster. Please use large file storage buckets like S3 that have a CLI 😭

Just over a month left to submit to this year's Argoverse 2 challenges! Returning from previous years, are our motion forecasting and lidar scene flow challenges. And NEW for this year with a $10k prize pool is our Scenario Mining challenge!🧵👇

On my way to Singapore for ICLR! If you want to chat about vision or robots, hit me up!

I apologize if you didn't follow to watch me slowly lose my mind over 3D CV research code, but holy shit guys... a coordinate transformation is GODDAMN useless with a reference frame. You can't just specify something as the "pose", it has to be "pose w.r.t. origin" which has to be explicit!

Was cleaning out some random papers from my desk and found this shipping receipt from @bhphoto.com from last year when they sent me my SD card that I left in a camera I returned. I didn’t even realize I left it in there until it showed up at my doorstep. This is why I’ll be a B&H customer for life.

People who work in 3D vision. PLEASE write your code with UNITS in mind. METRIC depth means you need to have METRIC units. So instead of writing `depth` just post-fix with the unit. For example `depth_m` or `depth_cm`. Think about how much time we collectively save, and type those extra characters.

One of those "productivity rut" days today. Sitting at my desk. code in front of me, and I just keep opening random tabs of other crap to distract myself. But alas... deadlines don't wait for motivation.

Freezing out the NIH's funding means freezing out American innovation. Research into curing cancer, Alzheimer's, and Parkinson's all delayed. And ultimately it won't just cost us advancements in life-saving research, it will cost us lives.

I think I will write a long form blog post on my website about this... but the distilled Deepseek R1 models are blowing my mind when it comes to coding tasks. I'm using the 4-bit quantized 32B distilled model which is based on Qwen 2.5 32B & it's been on par with the free ChatGPT offering (o3 mini?)

If you want to lead the way, then we have to restore funding for our academic institutions via NSF and NIH. You can’t cultivate top talent domestically if you don’t find science!

Domain verification works and is super sick. Just sayin’ 🤷‍♀️

This is such an important thing to understand and unfortunately the vast majority of crypto bros I interact with have no idea how market cap works. I think in large part because the term is typically used to talk about companies which do have some tangible backing value/assets.

Legend ❤️

Does anyone know how I can message the CVPR ACs? I spent the weekend dying of COVID and need an extra day or two to wrap up my reviews. Would like to message the ACs and give a heads up that I will finish but I cannot find where to do so in OpenReview. help me @csprofkgd.bsky.social ur my only hope

Lots of anti-Indian, anti-immigrant sentiment in the zeitgeist recently. Will use this opportunity to post my 🔥 outfit from a wedding this past week. Plenty of my non-Indian friends wearing traditional Indian clothing as well and absolutely killing it. Why be racist when you could look this good?

@merve.bsky.social said it best👇

A short list of tips for keeping a clean, organized ML codebase for new researchers: eugenevinitsky.com/posts/quick-...

The best one could be to speed up progress in AI at this moment of history is to shape whatever is the community's reward function so that a fraction of the compute and brain power currently knob-turning GPTs to gain 0.25% on whatever unrealistic benchmark try SOMETHING NEW.

As I haven't found it out there yet, I made the Women in computer vision started pack. Many more missing, please let me know how is already in bsky to add them! go.bsky.app/BowzivT

why do some people treat LLM-driven agents as POMDPs, and not as PODPs? are there works that treat them as PODPs?

Wasn't going to wade into this but... as long as the internet exists in an open form, scraping will exist, and that's kind of the whole point... APIs exist because interacting with data is valuable and scraping (the old fashioned way) is annoying.

And the perpetuating result is that we have keynote talks at major conferences about “research vs engineering” 🫠

Arxiv sharing reminder pdf ❌ abs ✅

Just a heads up to everyone: @deep-mind.bsky.social is unfortunately a fake account and has been reported. Please do not follow it nor repost anything from it.

Is this a thing we’re doing now? “Members of technical staff” really? Why not “machine learning engineer” or “research scientist”. is that not sexy enough anymore? *sigh* not trying to call out Matthew btw, this is definitely a “don’t hate the player, hate the game” kinda comment.

Wonderful blog post by Matt Mason: mtmason.com/the-beetle-r... Quote from it: “... When humans have to handle things not designed for their hands, it’s all tools! Look at their dentists and surgeons. Look at their watchmakers. ...”

Posting a call for help: does anyone know of a good way to simultaneously treat both POTS and Ménière’s disease? Please contact me if you’re either a clinician with experience doing this or a patient who has found a good solution. Context in thread

I'm hiring research interns for next summer to work on 4D reconstruction, 3D perception, & multi-modal diffusion modeling! To apply you must be a PhD student with work authorization in the USA. You will be expected to submit your work for publication at a top conference. 👇

This is the take! Nobody wants to stand up for iterating on the flawed status quo, but we've BEEN seeing what happens when scientific releases forgo peer review and it STINKS.

Also not to toot my own horn too much but we did this in ZeroFlow and we have a whole section about the significance of errors compared with the training variance over 3 training runs. It does cost compute but more people should do this!

Never doubt @csprofkgd.bsky.social 🫡

Training variance is a thing and no one measures it because research models get trained once to beat the benchmark by 0.2 AP or whatever and then never trained again. In prod one of the first things we do is train (the same model) a ton over different shuffled splits of the data in order to… 1/3

📢 Ultimate test of #NLP bluesky: I need emergency reviewers for NAACL submissions on encoders (one multilingual, one for sentence embeddings). Help a desperate editor abandoned by the ACs! Author response starts tomorrow, so that's a true emergency. If you're my hero, lmk your openreview profile.

I feel like there has been a vibe shift over the past decade. It used to be that everyone had a cool API & people were encouraged to build. VCs wanted to see "ecosystems". Now it appears that keeping a walled garden on data is the norm (perhaps b/c of the "threat" of models trained on scraped data).

One does not judge a friendship by the years it has endured, but by the contents of the soup during times of sickness. - MLK probably

I’ve said it before and I’ll say it again. Your depth estimation demo is meaningless unless you show me the point cloud. Preferably animated :) I gotta make robots work irl. Do you know how hard that is already? Ain’t nobody got time to deal with shoddy point clouds 💁‍♀️

Karpathy just posted a new tokenization tutorial on Bluesky. What does this mean for GPT-5? 👇🧵

Crazy that more of y’all aren’t following Chris. He posts some of the prettiest pictures if ur into CV 🥹

My Bluesky account currently has 10% of the followers I have on X. Considering that 90% of my followers over there are probably porn bots, I'd say that's encouraging. :)

This is your daily reminder that robotics is really fucking hard. It’s not just an ML problem, but also operations, hardware (sensors, cleaning, edge compute), software engineering (distributed compute, real time systems for middleware, coding standards & safety)…

R2 when I release a new real world dataset / benchmark: “how does this compare to FlyingChairs 2?”

Worked on a paper, the results are good, the math is solid, and the idea is neat. And somehow it gets good reviews on the first submission?!? Today is a good day.

I am trying to create a robotics and ai starter pack on bluesky: go.bsky.app/DfAoaJ1 Very incomplete please comment with suggestions (or just if you're missing and want to be added!)

Here's my attempt at assembling a starter pack of the still nascent computer vision community on Bluesky. Feel free to recommend other accounts that should be in this starter pack. :) #computervision (Do we use hashtags here? 😅) go.bsky.app/PkAKJu5

Continuing yesterdays rant about ML research code, I am back with another plea. Can we stop passing dictionaries around everywhere? Dictionaries are not self-documenting because I have to RUN the code in order to see what keys/values are in it. Dataclasses are almost always better. Use them pls 🙏