iiinova — where your digital self keeps evolving

Ishan Khatri @ishan.khatri.io · 24/09/2026

I’ll have a few more exciting things to share as my time at Stack comes to a close, and I’m looking forward to the next big adventure 🫡

000

Ishan Khatri @ishan.khatri.io · 24/09/2026

Last but certainly not least, Kyle and Neehar, the entire reason that research has been fun is you guys and I wouldn’t be here without you.

100

Ishan Khatri @ishan.khatri.io · 24/09/2026

It’s thanks to my internship at Argo that I met and collaborated with the best in this field, and I’m eternally grateful to Deva for taking me under his wing and the entire lab at CMU (Tarasha, Jenny, Aljosa, Jono) for adopting me.

100

Ishan Khatri @ishan.khatri.io · 24/09/2026

Finally I want to thank a ton of people. All of my colleagues from Argo, Motional & Stack and my management chain who supported me the entire time. These are all incredible places to work, filled with the smartest people and if you’re looking to work in autonomous driving you should 100% apply.

100

Ishan Khatri @ishan.khatri.io · 24/09/2026

Over the past 6 years working on autonomous driving I have learned so much about deploying robotics systems into the real world. I'll be taking this experience and pivoting slightly to focus on dexterous manipulation and related topics.

100

Ishan Khatri @ishan.khatri.io · 24/09/2026

Overdue career update time! I’ve moved to Toronto and will be starting graduate school (MS/PhD) @utoronto.ca & @vectorinstitute.ai, advised by Florian Shkurti! I will remain at Stack AV until the end of the year to wrap up some projects, after which I’m open for internships!👇

100

Ishan Khatri @ishan.khatri.io · 23/09/2026

… and artificially inflates the metrics. We re-evaluate using sane metrics and find that open vocabulary models suck at classification. It turns out if you just compute SigLIP embeddings on the crops and re-classify within the taxonomy using cosine similarity, you get much better results.

000

Ishan Khatri @ishan.khatri.io · 23/09/2026

And OV3D-Bench is a re-evaluation of 3D monocular object detectors. Honestly a reckoning that’s a long time coming. I love this field but this “target aware” evaluation crap needed to go. You can’t prompt with only classes that are present in the image!!! This hides the false positives…

100

Ishan Khatri @ishan.khatri.io · 23/09/2026

Quick belated paper promotion. We had 2 papers at ECCV that I was in Malmo to present! UniFlow is a foundation model for 3D LiDAR scene flow, demonstrating zero-shot generalization across datasets and LiDAR sensors for the first time. We also showed that semantics don’t generalize the same way.

141

Reposted by Ishan Khatri

Nathan Lambert @natolambert.bsky.social · 09/09/2026

It's a common agreement among my friends not at OpenAI/Anthropic that people at the labs operate with a religious energy (Ant especially). It manifests very out of touch interactions, which carries into public comms here (e.g the recent quitting). Company cultures create this.

1283

Ishan Khatri @ishan.khatri.io · 02/09/2026

And the gall of OAI and other modeling companies to use their incompetence as a marketing exercise. Like I know u guys have to pump the stock but really??? As a reminder *taps the sign*

120

Ishan Khatri @ishan.khatri.io · 02/09/2026

Agreed. I feel like I’m in loony toons land every time I see anyone talk about this. Even people who I respect are falling into this trap and it’s leaking outside the AI bubble and into the mainstream + the government. Giving legitimacy to this nonsense is damaging our field!!!

100

Ishan Khatri @ishan.khatri.io · 07/08/2026

Smh, some of us greybeards have been using 42 since before LLMs existed, where do you think they learned the reference from?

150

Ishan Khatri @ishan.khatri.io · 07/08/2026

I’m actually currently working on a “template” repo that sets up what I think is a good development environment/workflow and has built in instructions for agents to make this whole process a bit easier. Will publish soon 🤞

110

Ishan Khatri @ishan.khatri.io · 07/08/2026

Presumably “make sure your repo conforms to…” was your coding standards doc

120

Ishan Khatri @ishan.khatri.io · 07/08/2026

Ah yes I think that was the “you don’t have access to this doc” link that you referenced but I couldn’t see :)

210

Ishan Khatri @ishan.khatri.io · 07/08/2026

Sorry it’s a bit of a disorganized brain dump but this is everything I can think of off the top of my head. May add more if remember :)

110

Ishan Khatri @ishan.khatri.io · 07/08/2026

Force any ai generated functions to contain shape hints and type hints. Use dltype (or jaxtyping) for all tensors. Use ruff, pyre, pyright or mypy for typing and linting.

110

Ishan Khatri @ishan.khatri.io · 07/08/2026

Finally LLMs are very useful for debugging especially for wrangling dependency hell. However you can quickly get into sloppy bad states very easily. Make sure to follow best practices (virtual envs like uv or pixi for python/torch) and docker containers for cuda etc. specify in agents md!

110

Ishan Khatri @ishan.khatri.io · 07/08/2026

It is critical that you check data loading code and visualize. LOOK AT THE DATA!!! If you are not double checking inputs/outputs here you will suffer.

110

Ishan Khatri @ishan.khatri.io · 07/08/2026

Any algorithmic code should be either written directly by you or prompted in enough detail (pseudo code from you) that you fully understand every single line. Data plumbing (loading a dataset, plumbing it through torch dataloaders etc.) is a good use of LLMs BUT you should provide structure…

110

Ishan Khatri @ishan.khatri.io · 07/08/2026

When using AI for code I think it’s most useful for plumbing data and creating visuals. I think asking AI to invent its own tests is a smell. You should enumerate the test cases yourself even if AI writes them. Do not let an AI agent write a bunch of worthless tests, more != better here.

140

Ishan Khatri @ishan.khatri.io · 07/08/2026

At the risk of giving you too much useless confirmation I broadly agree with everything written here. The one thing I’d add is guidance of how/where to use AI to write code. You state that this is okay as long as you maintain understanding but I think it’s important to perhaps give some examples.

110

Ishan Khatri @ishan.khatri.io · 05/08/2026

Becky chambers is the GOAT. U should read this book if you haven’t already :)

000

Ishan Khatri @ishan.khatri.io · 29/07/2026

Yeah this is a banger format

1513

Ishan Khatri @ishan.khatri.io · 19/06/2026

@beenwrekt.bsky.social and I had a discussion about this a while back! bsky.app/profile/isha...

000

Ishan Khatri @ishan.khatri.io · 14/06/2026

It’s a beautiful day to make some GPUs go brr

Picture of a MacBook outside, the terminal is open on the screen and 4x GB200 are visible.

060

Ishan Khatri @ishan.khatri.io · 14/06/2026

Okay I’m gonna poke the bear (sorry!) do you believe in the validity of these claims? Because I find them really really hard to take seriously for a variety of reasons, and I’m having a hard time understanding how folks I perceive to be intelligent could be so misguided.

240

Ishan Khatri @ishan.khatri.io · 10/04/2026

The exact number really doesn't matter here, but it's super nice to see that people are actually reading the things I have worked on :)

050

Ishan Khatri @ishan.khatri.io · 24/03/2026

You also have to be good at writing. Unfortunately this is not one of my strengths so even if I’m inspired I just put down a couple bullets before smashing my head into the table and giving up 🙃 All this is to say that writing is a skill (which I lack unfortunately)

120

Ishan Khatri @ishan.khatri.io · 09/03/2026

Ahh good paper choice. So many of my former colleagues on this one 🤩

010

Ishan Khatri @ishan.khatri.io · 08/03/2026

While I recognize the importance of scientific writing and precise terminology it would be nice if we made papers more approachable for those of us who don’t read Ramanujan’s notebooks in our free time 💀

010

Ishan Khatri @ishan.khatri.io · 08/03/2026

Haha yeah I know, this was more of a vent post than a real question. In this case it’s just too much math terminology that I’m unfamiliar with despite knowing the underlying concepts. I will get through it. But I will also curse the authors the whole time 🙃

110

Ishan Khatri @ishan.khatri.io · 07/03/2026

Am I allowed to tell an AC that I'm too stupid to review a paper?

100

Ishan Khatri @ishan.khatri.io · 28/02/2026

Guys I never expected to go from watching this guy play minecraft as a kid (this is why I learned how to code btw) to watching him train LLMs. What the actual fuck: youtu.be/aV4j5pXLP-I?...

youtu.be

I Trained My Own AI... It beat ChatGPT

YouTube video by PewDiePie

030

Ishan Khatri @ishan.khatri.io · 22/02/2026

When I was TAing regularly I was always astonished at how bad some of the cheats were at covering up… why cheat so obviously? I have come for realize that it’s just laziness If you can put in the work to cheat properly you could probably just do the work lol.

000

Ishan Khatri @ishan.khatri.io · 22/02/2026

The last time I can remember a public official who was this effortlessly cool is Obama.

020

Ishan Khatri @ishan.khatri.io · 18/02/2026

Made a dedicated account for photography. Best viewed on @flashes.blue :)

010

Ishan Khatri @ishan.khatri.io · 16/02/2026

Anecdotally it also seems that activity on bsky comes in waves. People get all excited about leaving Twitter after "the latest nazi thing" they come to bsky, and then they leave here after a week or so. I have 3k followers here but still get more engagement on Twitter where I have 300.

010

Ishan Khatri @ishan.khatri.io · 16/02/2026

I suspect that graduate students are still incentivized to "obtain a following" in order to promote their work, and I suspect they feel like Twitter is still the best place to do that, especially with a lot of Industry folk still there rather than here.

110

Reposted by Ishan Khatri

Jess Miers 🦝 @jmiers230.bsky.social · 10/02/2026

The Discord age verification discourse focuses on adults losing access to communities they have a right to access. That's important. I'm also worried, though, about the kids. They seem to be an afterthought in these conversations. They have much to lose too. 🧵 www.theverge.com/tech/875309/...

theverge.com

Discord will require a face scan or ID for full access next month

Age verification for all.

917888

Ishan Khatri @ishan.khatri.io · 07/02/2026

Inside you there is an agent swarm

020

Ishan Khatri @ishan.khatri.io · 04/02/2026

@eugenevinitsky.bsky.social this reminds me of a legendary post made by my TA (and future collaborator) Kyle. Clearly you’re a lot nicer than my instructors ever were 😂

A piazza post where the students are asking for an autograder with unit tests and the TAs have responded by savagely roasting the students for asking, emphasizing that in the real world you must write your own unit tests.

140

Ishan Khatri @ishan.khatri.io · 04/02/2026

Yeah gradescope is great! I used it when I TAd classes at UMass :)

110

Reposted by Ishan Khatri

Thorne 🌸 @ens0.me · 02/02/2026

I would be deeply disappointed if Gavin is the nominee I would still vote for him, and so should you

5403

Ishan Khatri @ishan.khatri.io · 03/02/2026

Have heard good things about CachyOS (dunno if it counts as weird enough tho). Otherwise why not try vanilla arch + a fun WM? I’ve been enjoying hyprland a lot recently. Tiling is amazing if u have a big monitor.

110

Ishan Khatri @ishan.khatri.io · 03/02/2026

Unfortunately as it turns out, those folks are more interested in whatever larp they can do to scam folks & get money. I’ve also noticed a big overlap of the crypto larpers and AI hype shills. I’m not complaining though. The fact that they aren’t here makes it a better experience for me.

030