adarsh

/goal build agi. make no mistakes

its funny being in sf that i can see launch posts on twitter and say hey i met that shithead yesterday

@DeltaClimbs @soren_ma @rgury wanna try it out? i think you'll like it

@ak_xcix @soren_ma @rgury Nice! I got the Hover Air, and it has a bunch of modes other than follow me and remote control, but no way am I going through a menu like that!

Low key, isn't the way to fix this cultivating an instagram culture of voice AI controlled selfie videos to drive volume? I am deadly serious @soren_ma @rgury you are both building the factories, but can you generate the culture that creates the customer you truly need?

he is right, the ability to design custom silicon for drones enables dji to not have trade offs that using an off the shelf mcu/sbc would this also enables them to add capabilities which are damn near impossible in similar sized drones, all downstream of having control over both software and hardware and ability to tailor it to their needs

@BigTechAlert @ChatGPTapp @adilmania yoooooo

🆕 @chatgptapp has started following @adilmania

hmmmm

@pham_blnh that's cool!

i bet you haven't seen a SO101 mounted on a wall like this before if you want to do the same, here is LeSlider: github.com/pham-tuan-binh… i built it cause i wanted something that can cover my whole desk for tasks like organizing and cleaning i originally wanted to have a belt system like what 3D printers have, but i was too lazy and used a pinion/track with another sts3215 so: > the extra motor shares the same bus as the rest of SO101 > you can have arbitrary length of track > really cheap and easy to assemble and control it turned out better than expected with this, i'm gonna train a model to pick random stuff up across my table and put it into a bin at the end of table (realistically using yolo to scan table, two policies, one for picking up objects, one for dropping)

@usr_bin_roygbiv @typesfast did you call a350 old?

@typesfast I would far rather they just put the model so I could avoid the old ass a350/380s and take all the 787 flights with good air pressure

The lowest emissions planes are the newest / nicest ones

@lanceyyan fantastic timing!

we're subleasing the place dm me if you need a place to stay in SF starting May

@IlirAliu_ @OrangewoodLabs @Abhindas1 was fun meeting you!

My first visit to @OrangewoodLabs, and Co-Founder and CEO @Abhindas1. 00:00 Intro to Orangewood and Founders in Garage. 00:27 Building industrial robot arms you can talk to. 01:01 A look at the robotics technology. 01:24 Demoing the software and training the models. 01:53 The origin story: from a basement to a business. 02:34 The future of humanoid robots and small-scale manufacturing.

@robotsailor i respect your art

@ak_xcix i made it unnecessarily painful to make things look good where no one will even see them

everytime i pop the lid on the robot i get really happy seeing how clean everything is inside

@NimaZeighami damn, i’ve always seen you guys working incredibly hard until so many midnights at the garage on the robots. best of luck on whatever comes next

I was recently fired from REK, a company I co-founded. At REK I helped deliver the world’s first dozen VR-controlled humanoid robot fights. I still own a stake and wish them the best. Now taking time to travel and explore my interests.

@SpaceX @cursor_ai whoa

SpaceXAI and @cursor_ai are now working closely together to create the world’s best coding and knowledge work AI. The combination of Cursor’s leading product and distribution to expert software engineers with SpaceX’s million H100 equivalent Colossus training supercomputer will allow us to build the world’s most useful models. Cursor has also given SpaceX the right to acquire Cursor later this year for $60 billion or pay $10 billion for our work together.

@pratty_agi @natashamalpani @OrangewoodLabs @AgentOpsAI yeah we have continued the work on models for observability and self improvement in policy training

@natashamalpani @OrangewoodLabs Somewhat. I could tell when the robotic arm would start to fail and should back off instead of retrying. I shortly switched to @AgentOpsAI after.

the most valuable training data in physical AI is sitting in deployed robots right now. unlogged. the field obsesses over successful trajectories. teleoperation hours. demonstration pipelines. simulation runs. but almost no one is systematically capturing failure data. a policy trained only on successful demonstrations learns what good execution looks like under nominal conditions. but in production, conditions are almost never fully nominal. recovery behavior is the hardest problem in deployed manipulation. three companies are working on this: @agibot released world 2026: one of the only open datasets with explicit error-recovery trajectory annotation. hierarchical labeling from high-level task to low-level action. corrective priors built in. @physical_int shifted toward autonomously collected experiential data including suboptimal failures. the explicit logic: policies need to correct the mistakes they actually make in deployment. not just replicate the successes they were shown in training. @machinalabs captures deviation data in production. real-time. closed loop. every deviation from expected is a labeled training signal. acted on immediately. the signal here is in the hardware. it is not being captured because failure data accumulates only through deployment: it cannot be synthesized. it cannot be purchased. it is unglamorous, operationally complex, and deeply valuable. the company that builds failure event detection, multimodal logging, structured annotation, taxonomy standards builds something every serious physical AI company needs. and none of them want to build themselves.

:(

@adilmania happy birthday!

today i turn 29 🎉 today is also 4/20 so i'm dropping a new song. it's called... White Mirror enjoy!

@IlirAliu_ @visitdreampark @brentbushnell :)

A bit scared to actually do some video content... 🙈 When I was in SF a couple of weeks ago, I visited a few startups vlog-style. Hope you like it: First stop: @VisitDreamPark, founded by @brentbushnell, Aidan Wolf, and Kevin Habich. The idea and concept is simple: the world's first mixed reality theme park. No venue, no headset, no single attraction ticket. Instead, public spaces, like streets, parks, malls and so one, become playgrounds layered with augmented reality. I am an 80s kid, so this playing SuperMario in real life. Their flagship experience is called Super Adventure Land. You collect coins, dodge obstacles, interact with virtual objects... all through your phone, in the middle of the real world. The model scales without expensive infrastructure. No real estate, no construction. Just software dropped onto existing spaces. They've been on Shark Tank, and are currently expanding to Seattle, Santa Monica, and Long Island. More of these visits are coming. Thanks for having me, Kevin Habich! 👋 And visit their website visitdreampark.com!! Let me know what you think. 😛 —— Weekly robotics and AI insights. Subscribe free: 22astronauts.com

if you could visualize vibecoded architectures