Vimarsh

We developed an RL method for fine-tuning our models for precise tasks in just a few hours or even minutes. Instead of training the whole model, we add an “RL token” output to π-0.6, our latest model, which is used by a tiny actor and critic to learn quickly with RL.

trained snakebot locomotion policy with end2end rl. mjlab makes such these too easy. the most difficult part was making the mjcf due to 3 closed link of our snakebot design

@SforAiDL Crazy work @harshvardhan_sm 🐐🎉

We’re releasing lrnnx, an open-source library for working with Linear Recurrent Neural Networks (LRNNs). LRNNs bridge classical linear dynamical systems and modern deep learning, providing strong guarantees on stability, trainability, and long-range modeling. (1/4)

This is who you’re competing against, Chinese fruit seller and chip designer. Yea, you’re cooked:

Did you know you can make your own containers like Docker from SCRATCH? I was reading about cgroups for a course and found some interesting stuff about how containers work under the hood, I have written about my learnings here - reckt32.github.io/blog/post.html…

2026 is the year Linux finally catches up to...UNIX. No, seriously. Tiered memory architectures were solved in UNIX decades ago. The era of pretending "all RAM is equal" is becoming unaffordable. Thankfully, software is getting pretty clever:

StackOverflow graph of questions asked per month. Holy shit.

Outside of the datacenter world, no one realizes CXL exists. Trust me, you’re gonna hear it more often. It’s just barely starting to leak in the prosumer space. The short version is: CXL = memory over PCIe. The long (cool) version is: CXL = memory as a switching fabric! CXL 2.0 was kinda cool, but not that revolutionary. Basically just a way to squeeze extra ram on a server using PCIe slots, albeit with a ~200ns or so latency penalty. Software-wise, the OS maps the CXL region into virtual address space, so even in assembly-land you can: mov rax, [CXL_address] Dumb operating systems (cough Windows) treat the entire region as one big bucket of ram. Modern Linux Kernels understand CXL has a latency penalty, and tier it as a different NUMA bucket. --- CXL 3.0 is where it got wild. You can have non-tree topologies where a “pool” of memory is shared *across* hosts. In layman’s terms, you can treat memory like NAS storage. If you’re a programmer, you might realize how CRAZY this is. Host A and Host B can share a pointer to the SAME physical address. Host A can do partial calculations, write to a shared address, and then Host B can pick up right where it left off! Of course, this also makes thread teleportation between machines dead-simple. Copy the CPU register state from Host A to B, resume execution in the same memory pool. In some ways, it’s a security and programming nightmare. In other ways, it’s one of the coolest pieces of tech I’ve ever seen. I’ve barely scratched the surface of what’s possible. Expect to see more talk about CXL as we head into 2026!

my only problem with this happening is this the university that has this facility was world ranked #1 for the most number of shady research practices, including publishing in high numbers in banned journals, withdrawing papers when not accepted etc. i really feel sad that these kind of resources end up in the hands of the people who propagate the worst research culture [in the world].

We’re building a custom Cycloidal Drive for Project Quadruped instead of buying off-the-shelf gearboxes. 🐕🤖 Why the extra effort? Reasons are listed in this thread👇

gemini 3 pro needs some help

US is delusional for some reasons. But they have the economy to back it up. India exports iron ore and imports steel. It is in no state to be this delusional. AI is a long shot. It’s bleeding $100bn every year to China funding Chinese industrial capacity and GDP growth while importing steel from it. It’s shameful. It’s not smart.

I am not sure about this. I don’t want personalized movies. I want Aditya Dhar’s Dhurandhar. I want to watch Kubrick, Nolan, Park Chan-wook. I want to watch your story. jdkanani.com/blog/your-visi…

Students are anyways going to use AI tools for coding assignments so I can support being liberal in the use of these tools in this case. Exams, on the other hand, not so much. And it’s not even a question of assessment mechanisms. I am looking at it purely from the perspective of whether a student was able to connect various bits of information they learnt as the part of a course with just their own brain as the tool at hand. Three hours in an exam room with a pen and paper exercises the brain to an extent that AI-assisted take home exams cannot. I for one, am becoming increasingly convinced that pen and paper, in person exams should be the norm in the future. PS: this is purely for UG programs.

reducing size of 3d printer gearbox is very difficult :/

Even encoders which combining all together is difficult (and very expensive).. We are also struggling to find any foc controllers (made in india) which are as cheap as odescs... We really need to work on increasing mass manufacturing of parts

We are developing our own (relatively large ish) quadruped @ERC_BITS_Goa and the biggest bottleneck we are facing is making a smaller gearbox with required torque to attach to the bldc The chineese manufacturers make very good actuators with built in controllers as well

@caleb_friesen the most difficult part to manufacture in house is the actuators and the foc controllers.. sheet metal fabrication and cnc can be done at several places, but making an actuator (90kv motor) with something like an internal gearbox is the difficult part.. also making foc to drive

This is so much more impressive when you realise that this robot dog is almost entirely indigenous. Only non-Indian parts are the actuators + Nvidia Jetson GPU. Oh, and it was built by a team of 5 people in 30 days. IP67 rated, running temperature: -20C to +55C

it's super cool, but I don't feel good about the future we are heading into... It's already difficult to seperate what's real with fake and with advertising incentives...