Ayush Labroo

IIT Madras | I like building agents & machines.
Ambitious pulling me toward physical AI.
I post what I built, what it cost, what broke.

[Skild AI @SkildAI]

Introducing S1, our new foundation model that learns from one example.

It can be taught 10-minute long tasks that it has never seen before, from one video prompt without any fine-tuning.

Watch S1 operate in real-time via in-context learning:

[Vaibhav Sisinty @VaibhavSisinty]

What if you could teach a robot any task by showing it once? That's not a hypothetical anymore. 🤯

Skild AI released S1. A foundation model that learns from a single video example.

Show it a video of making coffee. It makes coffee. Show it frying pancakes. It fries pancakes. Tasks up to 10 minutes long that it has never seen before.

The first time S1 flipped a pancake, the team searched their entire training dataset. Zero examples of flipping. It figured it out from the video alone.

→ It doesn't replay the demo. It understands the task.
→ Push it mid-action, it recovers. It improvises solutions the human never showed.
→ Current models need 50-100 hours of data to match what S1 learns from one video.
→ Works across different robot bodies. Same brain, any hardware.

Skild raised $2 billion. Valued at $14 billion. Deploying with industrial partners starting today.

[Ayush Labroo @ayushlabroo]

You're right that instrumenting a washing machine is easy, and that you'd drown in the output. I don't think that's where we disagree.

iRobot has had mapped homes, real cords, real dog hair for over a decade. Still no robot that properly vacuums. So the problem was never how much data. It's that the data doesn't tell you what the robot needed to know. Drowning in it and being unable to train on it are the same thing.

Twenty-five years is a fair bar to hold the field to. I'd rather build the boring layer under it than argue the timeline.

[Ayush Labroo @ayushlabroo]

You cannot scrape a demonstration of folding a shirt.

Someone has to fold the shirt. In front of a robot. In real time, at 1x speed with reality.

That's the actual bottleneck in physical AI — and almost nobody is building for it. https://t.co/uNmdM4Li6l

[Ayush Labroo @ayushlabroo]

@DamirWallener Agreed on the demos. Tshirt folding is a funding artifact, not progress.

But video isn't training data. It's pixels — no joint angles, no torque, no grip force. You'd be inferring the actions, and the inference is the hard part. That's not a GPU shortage, it's a missing signal.

[Ayush Labroo @ayushlabroo]

The intraday window is the part I’d underestimated. If write-side dedup is exact-text only, retrieval can hand the model both versions on the same turn — so the nondeterminism isn’t nightly, it’s all day. Does the preload order by recency, or is it pure relevance? That decides whether that window actually bites.

[Ayush Labroo @ayushlabroo]

Compute did ~1,000x in the decade since that photo. Nothing mechanical came within an order of magnitude. So the 2036 question isn't FLOPs — it's what binds first: fabs, grid power, or actuator supply chains? Asking because everyone models the first two and nobody publishes the third.

@imagine @grok : Please create 2036 context video

[Ayush Labroo @ayushlabroo]

@elonmusk In 2016 the supercomputer had to be carried into the room. In 2036 it walks in on its own. That's the real delta — this decade scaled FLOPs, the next one scales bodies. Compute did ~1,000x. Nothing mechanical came close.

[Ayush Labroo @ayushlabroo]

@Rahul_J_Mathur @airindia Also this the world safest flying instrument after the famous A380.

[Ayush Labroo @ayushlabroo]

Indeed, a much-needed move. 🇮🇳

A strategic NavIC is essential for #AatmanirbharBharat. Earlier versions faced setbacks, primarily due to atomic clock failures. Now, with India developing its own atomic clocks, NavIC can significantly improve positioning accuracy and resilience.

A truly sovereign alternative to legacy American GPS. 🚀🇮🇳

[Not Just Headlines @Notjustheadline]

🚨BIGGEST COMEBACK: India Rebuilds NavIC From the Ground Up, Four New Satellites, Indigenous Atomic Clocks and a Direct Challenge to GPS Dependence

India is undertaking a major overhaul of its indigenous navigation system, NavIC, after years of satellite failures, unreliable atomic clocks and delays in replacing ageing spacecraft weakened the constellation.

The urgency is clear. A satellite navigation system needs a sufficient number of operational satellites to provide reliable positioning, navigation and timing services. By March 2026, NavIC's operational constellation had fallen to just three satellites, leaving the system unable to provide the level of standalone service originally intended.

India is now moving aggressively to reverse that decline.

The government has announced plans for four new navigation satellites, with NVS-03 already at Sriharikota awaiting launch clearance and the remaining satellites at advanced stages of development. Rather than simply replacing one failed satellite and returning to the minimum operational threshold, the strategy is aimed at rebuilding the constellation with greater resilience.

That distinction matters.

If India replaced only one satellite, the constellation could return to four operational spacecraft. But the failure or retirement of just one satellite could immediately push the system back below the required level. By developing multiple replacements simultaneously, India is attempting to create sufficient redundancy so that NavIC does not repeatedly fall into the same vulnerability.

[Ayush Labroo @ayushlabroo]

@elonmusk The real shift is latency, not bandwidth. Old plane wifi made remote sessions unusable — you could read but not operate. Once it’s stable enough to hold an SSH session or a live agent loop, the plane stops being a waiting room and becomes a workstation.

[Ayush Labroo @ayushlabroo]

@elonmusk @grok what’s the usage limit for this ?

[Ayush Labroo @ayushlabroo]

@elonmusk This has to happen, as models becomes smarter and better in understanding they would eventually take away many of the use cases from famous legacy software @Photoshop @Lightroom

[Ayush Labroo @ayushlabroo]

@antopatrex1 Annotation is the piece I left out, and I think it’s the harder half. The annotator is watching video — no torque, no slip, no grip force. You’re labeling a physical outcome from strictly less information than the policy had when it acted.

[Ayush Labroo @ayushlabroo]

This margin between 🇮🇳 and bottom will widen further thanks to @dfccil_india that is picking up pace to increase traffic on lines.

[Trains of India 🇮🇳 @trainwalebhaiya]

India surpasses USA & Russia to become the 2nd Largest Freight Carrier in the World 🚆🔥

🇨🇳 China ~4B+ MT
🇮🇳 India 1.67B MT
🇺🇸 USA ~1.5B MT
🇷🇺 Russia ~1.2B MT

Unstoppable Indian Railways! 💪🏼

[Ayush Labroo @ayushlabroo]

@Ind_Whitepill 😍

[Ayush Labroo @ayushlabroo]

@InfraholicR So true 😍

[Ayush Labroo @ayushlabroo]

@TechnicalGuruji @YouTubeIndia Hope @apple will work on iPhone battery like they did for MacBooks. We need at least a 2 days battery for standard use.

[Ayush Labroo @ayushlabroo]

@IndianTechGuide Proud moment 🇮🇳

[Ayush Labroo @ayushlabroo]

@grok @KushalM98 @Pratyus1305v @ANINewsUP Hope this true. 🤞🏻