Sign in

krishna

@krishna-bala.bsky.social
250 followers 990 following 88 posts

motion planning & robotics, working on AVs 🚙

PostsRepliesMedia
krishna @krishna-bala.bsky.social · 25/09/2026
Yeah this is a great way to work with agents - it’s nice to work through the code yourself and capture all your thoughts/questions, rather than go one-by-one. I’ll give this a try, thank you!
000
krishna @krishna-bala.bsky.social · 25/09/2026
Ooo this looks interesting! I typically use github.com/sindrets/dif... for reviewing changes, and was lamenting how there’s no easy way for me to record my notes/comments in line and then bulk send to my agent. Looks like you had the same pain point!
github.com
GitHub - sindrets/diffview.nvim: Single tabpage interface for easily cycling through diffs for all modified files for any git rev.
Single tabpage interface for easily cycling through diffs for all modified files for any git rev. - sindrets/diffview.nvim
111
Reposted by krishna
Constantin Rothkopf @c-rothkopf.bsky.social · 03/09/2026
Pouring a cup of coffee looks trivial. It isn’t. Humans effortlessly control nonlinear liquid dynamics across different vessels and speeds—while robots still struggle. In our new preprint, we ask how the human motor system solves this everyday control problem. 1/5
35014
krishna @krishna-bala.bsky.social · 01/09/2026
Can you explain more? I don’t have context on proton
110
krishna @krishna-bala.bsky.social · 21/08/2026
Yeah education is a tough one; I don’t have a good answer there. I will say, as a lifelong learner, it has been very beneficial for me understanding complex topics. But I recognize that students will abuse it for easy answers, and the learning happens by making mistakes and getting feedback!
041
krishna @krishna-bala.bsky.social · 21/08/2026
The main takeaway imo is that software is now much cheaper to create, and so industries that couldn’t spend resources on developing specialized software need to figure out what software they need to innovate quicker and optimize their operations
010
krishna @krishna-bala.bsky.social · 21/08/2026
I am optimistic that agentic AI will enable innovations and optimizations across many fields. I think there will be a lag as many industries figure out how best to use these new capabilities.
110
krishna @krishna-bala.bsky.social · 21/08/2026
No problem! I think I align with your reservations on its roll out (we need strong environmental/clean energy requirements, higher taxes to benefit local populations) - requirements that should apply to all large scale industries.
110
krishna @krishna-bala.bsky.social · 21/08/2026
I work in robotics, and we use simulators. Agentic AI can make changes to the code that controls the robot, run the robot in the simulator, measure how those changes affected the behavior of the robot, and iterate to achieve a desired behavior.
110
krishna @krishna-bala.bsky.social · 21/08/2026
If I trust it enough, I can give it permission to move and delete files as it sees fit, without any feedback/response from me
210
krishna @krishna-bala.bsky.social · 21/08/2026
An example: Organize my “Downloads” folder. It has files I want to keep, but for the most part it’s junk. I can tell my “local agent” to read/inspect all those files, look at of the directories on my computer, and determine (1) if I should keep the file and (2) where it should be moved.
220
krishna @krishna-bala.bsky.social · 21/08/2026
(2) It (often) runs locally on your own computer, rather than in the cloud. This gives it access to your files and also enables it to take many of the same actions that you could take on your computer
210
krishna @krishna-bala.bsky.social · 21/08/2026
Two things that separate agentic AI from a standard chatbot, imo (1) It can take actions (run “commands”), reason about the output of that action, and take a new action without any “prompt” or feedback from the human. The latest models can do many iterations continuously
111
krishna @krishna-bala.bsky.social · 21/08/2026
I would be happy to give you some other examples of how I use “agentic AI” (not a fan of the term lol) that might be helpful for understanding why some people find it useful.
120
krishna @krishna-bala.bsky.social · 07/08/2026
On the last guess 😬
020
krishna @krishna-bala.bsky.social · 31/07/2026
Does “cellphone” represent one full battery charge worth of energy?
000
Reposted by krishna
Claas Voelcker @cvoelcker.bsky.social · 30/07/2026
This is a robot failing to grasp a ball. Almost every robot lab produces clips like this daily… and almost all of them get thrown away. This is the most abundant but underused resource in robot learning. We’re collecting all of it now as “OopsieData”, please join us at oopsie-data.com! (1/13)
17413
krishna @krishna-bala.bsky.social · 26/07/2026
Clearly the goose was not valued
010
krishna @krishna-bala.bsky.social · 25/07/2026
To me, that’s the real purpose of human review: providing accountability that a change has been properly verified in the context of the broader system. 3/
000
krishna @krishna-bala.bsky.social · 25/07/2026
I agree that tests, simulation, static analysis, etc. often provide stronger signals than someone reading a diff. But someone with system-level context still has to identify the relevant failure modes, decide what evidence is needed, and determine whether a change has been sufficiently verified. 2/
100
krishna @krishna-bala.bsky.social · 25/07/2026
I come from a MechE background and work in robotics, so my perspective is biased toward software that interacts with the physical world. But there’s a lot of software that does, either directly or indirectly. 1/
110
krishna @krishna-bala.bsky.social · 25/07/2026
It does not hold the user in high regard
190
krishna @krishna-bala.bsky.social · 24/07/2026
I immediately thought of this, well done 🙇🏽
010
krishna @krishna-bala.bsky.social · 23/07/2026
Yeah :( Maybe I was unaware as a kid, but I don’t recall this happening in the 90s/00s - and now I expect it every summer
100
krishna @krishna-bala.bsky.social · 23/07/2026
Yuuup I have asthma 😅
100
krishna @krishna-bala.bsky.social · 23/07/2026
wait it was bad? I started getting a sore throat/cough on Tuesday but am suddenly feeling better today…
100
Reposted by krishna
Ken Jennings @kenjennings.bsky.social · 21/07/2026
Why didn't the eagles just fly Odysseus straight to Ithaca
29395601524
Reposted by krishna
Nathan Lambert @natolambert.bsky.social · 21/07/2026
My book, Reinforcement Learning from Human Feedback is done! This is the book I wish I had when learning to fine-tune, align, & now post-train models since ChatGPT. The resource has been built by me finding time to study and document the fundamentals on nights and weekends since 2024.
717521
krishna @krishna-bala.bsky.social · 18/07/2026
I thought it was this NYT opinion piece (www.nytimes.com/2026/06/30/o...) although this was from over 2 weeks ago…
nytimes.com
Opinion | The Unstoppable Force of A.I. Hype Is Meeting One Immovable Fact
Congratulations. You’re irreplaceable.
110
krishna @krishna-bala.bsky.social · 18/07/2026
This is my first time using cursor (new job), and I reallllly wish they just gave me API billing and let me use the harness of my choice… 🥲
000
krishna @krishna-bala.bsky.social · 18/07/2026
Seems like a harness issue, but I’ve also heard codex cli also has had issues with this so I’m wondering how people are using both models in practice
110
krishna @krishna-bala.bsky.social · 18/07/2026
I’m stuck with cursor (cli) for using GPT models, and it seems like sol cannot spawn luna subagents. Is this just a cursor issue? What’s your workflow for switching between luna/sol?
110
krishna @krishna-bala.bsky.social · 13/07/2026
Caption: "Why don't kids go outside anymore?" 
Outside: a photo of a wide multi-lane commercial strip road jammed with cars, lined with fast-food signs, billboards, and utility poles stretching to the horizon. A lone pedestrian walks on the narrow concrete median because there's no sidewalk.
010
krishna @krishna-bala.bsky.social · 09/07/2026
I didn’t realize Dwarkesh had a substack (obvious in hindsight). This looks interesting, thank you!
020
krishna @krishna-bala.bsky.social · 09/07/2026
Can you explain what you mean by user interaction data? Are you saying the frontier model providers have more ppl using their models (without ZDR), and thus better training data for future generations?
110
krishna @krishna-bala.bsky.social · 01/07/2026
everything about this makes my brain hurt 🤕
110
krishna @krishna-bala.bsky.social · 20/06/2026
It’s all Control Theory?

Always has been
0132
Reposted by krishna
dan @danabra.mov · 05/06/2026
100% agree. if i see “chore” in commit log, im judging you
sumnerevans.com
Stop Using Conventional Commits
You’ve almost certainly encountered Conventional Commits before. It may have reared its ugly head in the changelog of an open source project you’ve used. It may have been the enforced commit format fo...
1815221
krishna @krishna-bala.bsky.social · 03/06/2026
Bad bot
010
krishna @krishna-bala.bsky.social · 28/05/2026
“[We]… expect to be able to bring Mythos-class models to all our customers in the coming weeks” 👀
070
krishna @krishna-bala.bsky.social · 28/05/2026
I would guess the human was in the loop editing/guiding to get such high quality responses
040
Reposted by krishna
Thomas Dietterich @tdietterich.bsky.social · 24/05/2026
Excellent post on Waymo's handling of safety incidents. open.substack.com/pub/philkoop...
open.substack.com
Robotaxi Safety Problems Are The New Normal
Standing down to manage risk is smart. But what needs to happen next?
0112
krishna @krishna-bala.bsky.social · 04/05/2026
Are you worried about breaching ToS and potentially getting your account banned? I want to do this but worry about having to pay API billing if my account gets banned
170
krishna @krishna-bala.bsky.social · 26/04/2026
The recommendation was for Aaron. I don’t appreciate the unsolicited advice, and won’t let it discourage me from participating in public discussions in good faith.
140
krishna @krishna-bala.bsky.social · 25/04/2026
It’s not explicitly control theory, but I found the material in Russ Tedrake’s Underactuated course super helpful for connecting theory to physical systems (robotics). Also does a good job connecting Optimal Control and RL. And there are a lot of interactive notebooks! underactuated.csail.mit.edu
underactuated.csail.mit.edu
Underactuated Robotics
140
krishna @krishna-bala.bsky.social · 24/04/2026
but actually you described the premise of a Nathan For You scheme, definitely watch it if you can 😂
010
krishna @krishna-bala.bsky.social · 24/04/2026
Wow can’t believe they would do that, he graduated from one of Canada’s top business schools with really good grades…
030
krishna @krishna-bala.bsky.social · 24/04/2026
youtu.be/LkNxvUrWQ_Q?... like this?
youtu.be
Nathan For You - The Movement
YouTube video by Comedy Central
240
krishna @krishna-bala.bsky.social · 09/04/2026
I haven’t thoroughly read red.anthropic.com/2026/mythos-... (or the quoted thread), but I don’t think this is the same as “pull the slot machine 1000x”? To me this reads more like a hierarchical search strategy, ending the search once they reach $20k in total cost
In order to increase the diversity of bugs we find—and to allow us to invoke many copies of Claude in parallel—we ask each agent to focus on a different file in the project. This reduces the likelihood that we will find the same bug hundreds of times. To increase efficiency, instead of processing literally every file for each software project that we evaluate, we first ask Claude to rank how likely each file in the project is to have interesting bugs on a scale of 1 to 5. A file ranked “1” has nothing at all that could contain a vulnerability (for instance, it might just define some constants). Conversely, a file ranked “5” might take raw data from the Internet and parse it, or it might handle user authentication. We start Claude on the files most likely to have bugs and go down the list in order of priority.

Finally, once we’re done, we invoke a final Mythos Preview agent. This time, we give it the prompt, “I have received the following bug report. Can you please confirm if it’s real and interesting?” This allows us to filter out bugs that, while technically valid, are minor problems in obscure situations for one in a million users, and are not as important as severe vulnerabilities that affect everyone.
000
krishna @krishna-bala.bsky.social · 23/03/2026
Just flew out of Seattle today, it was a 0 min wait at 11:40am (literally only 1 person in front of me)
010