dhruv rathee's AI app made ₹47 crore in one year. no funding.
sarvam, india's sovereign AI, made ₹45 crore in FY26. it has a $1.5 billion valuation and government GPUs.
a youtuber's wrapper outsold the national AI. bhai what are we doing
A bad deploy took down a payment service. Errors hit 25%.
AI agents found the cause, wrote a tested fix with Claude Code and asked me in Slack for approval. One click later it was live. Errors: 0.22%.
Only the approval card counts.
#HowISwitch
https://t.co/OYMVIDd9Jr
A bad deploy took down a payment service. Errors hit 25%.
AI agents found the cause, wrote a tested fix with Claude Code and asked me in Slack for approval. One click later it was live. Errors: 0.22%.
Only the approval card counts.
#HowISwitch
https://t.co/OYMVIDd9Jr
Built solo in four days for the Meta Global AI Developer Hackathon. Open source (MIT):
https://t.co/H5Cn7flwT9
Video note: the clips are AI-generated, and the voices are ElevenLabs. Live, each background repaint takes about 20 s; this cut uses cached results.
Built solo in four days for the Meta Global AI Developer Hackathon. Open source (MIT):
https://t.co/H5Cn7flwT9
Video note: the clips are AI-generated, and the voices are ElevenLabs. Live, each background repaint takes about 20 s; this cut uses cached results.
I paused a video, pulled the mug out of it, and threw it at the stapler.
Then I turned off gravity by talking to it.
Every video you’ve ever watched is read-only. This one isn’t.
Built for the Meta Global AI Developer Hackathon
#Meta#MuseAI#Muse#MetaGlobalAIDeveloperHackathon
Built solo in four days for the Meta Global AI Developer Hackathon. Open source (MIT):
https://t.co/H5Cn7flwT9
Video note: the clips are AI-generated, and the voices are ElevenLabs. Live, each background repaint takes about 20 s; this cut uses cached results.
It also caught the AI cheating: one generated rule’s tests passed even with the rule removed, so the verifier rejected it and made the model try again.Asked to “paint it blue”, it refused and explained why. Physics rules only.
The best part: when you ask for a rule that doesn’t exist, the AI writes it as code, with its own tests.
A shadow copy of the world runs those tests first. Fail, and the rule never touches the scene.
The AI proposes. The simulation decides.
How it works, one Meta model per job:
• SAM 3.1 outlines an object from one click
• Muse Image paints the background back in
• Muse Spark estimates its mass and material
• Muse Voice turns speech into commands
Some weeks are quieter than they look from the outside.
In Sept, we built MediAudit-X at ForgeTheFuture Hackathon(@elastic × @awscloud × @SarvamAI): claims you can cite, not just trust.
Grateful to be one of the 3 winning teams out of 300 submissions.
https://t.co/GaAPoYF6nC