Once made a robot move when I moved my arms (patent, 2019, past life). Now builds AI and 3D stuff for homes. Two daughters, one Dota account, zero free weekends
Gurugram, India
Joined September 2008
- Tweets1.8K
- Following281
- Followers224
- Likes3.7K
1/ This is Gurugram.
Not a game studio's trailer. One person, open data, and a browser tab.
5 km × 5 km of the city, rebuilt in 3D. You can walk it or drive it. 🛺
How it's made 🧵
1
3
146
@deepigoyal launched dark.ai and a theory of AI agents in the same earnings letter.
Agents can reorder your atta. They can't pick your biryani. 65%+ of Zomato opens are people just browsing.
Most exposed to agents: District. Concert seats don't need a middleman.
Cheat sheet 👇
1
45
I turned two weeks of briefing Opus 5.5 on games, 3D, sprites and coded video into one free, interactive field guide.
Every line you write shows up on screen. Every line you leave out gets the model's most average guess.
What's inside:
Meta’s Project Phoenix leaks look dope 👀 I’d rather keep my current phone and get this than buy an iPhone Duo just for a bigger screen. Give me a floating cinema, not just a phone that unfolds. @finkd
104
I asked Astra for an NFS-inspired Delhi racer starring the Mahindra XEV 9e.
It built Delhi Afterhours. Nearly 16 km, rain, traffic, drifting and boost. In a browser.
Play the prototype: delhi-afterhours-astra.verce…
Ray tracing changed how games simulate light.
Neural rendering may change how much they need to simulate at all.
I open-sourced an unofficial DLSS 5 video player to explore that shift:
github.com/2600th/dlss5-vide…
Verified on RTX 4080 + 5090.
1
1
268
DLSS 5 Video Player v0.14.1 is live.
It now works with photos + animated GIFs, exports PNG/JPEG/GIF/MP4/MKV, adds portable caching, smarter fullscreen UI, and 6 new AAA trailer samples.
Built for RTX 40/50.
github.com/2600th/dlss5-vide…
118
Made this 30s sci-fi fight with MiniMax H3 running locally on an RTX 5090 + Seedance 2.5.
The leap isn’t prettier frames. It’s control.
Production is becoming less of a gatekeeper.
Story, taste, direction and the agency to finish are becoming the moat.
#AIVideo
Watched Operation Safed Sagar on #Netflix.
A few days later, we had this running in a browser. 🇮🇳
A small AI-built 3D experience inspired by the courage, ingenuity and flying missions of the Kargil War.
Here’s what we made 👇
Update: some of you managed to crash my AI-built ocean 😄
So I went back in. It now loads faster, uses far less memory, and should work on many more devices.
If it failed before, give it another shot 👇
web-ocean-3d.vercel.app/
#GameDev #VibeCoding #ThreeJS
I wanted to test a simple question:
How far can we push today’s AI models on a real-time 3D experience?
So I gave Opus + Sol a target: build toward a Black Flag-inspired ocean + ship experience, running entirely in the browser.
🎮 web-ocean-3d.vercel.app/ teapawt.net/2600th/status/20857884…
The experiment was inspired by @mattshumer_
and his work around pushing AI coding agents harder. The shift for me: stop treating the model like a one-shot code generator. Build → run → inspect → critique → fix → repeat. Each loop raises the bar a little more.
1
3
152
ARC-AGI is one of the hardest AI benchmarks. It tests pure abstract reasoning, not memorization. Human baseline is around 95.4%.
Prime Intellect’s new open-source Prime Agent just scored 95.5% on ARC-AGI-3 with no special training.
Pretty wild.
github.com/PrimeIntellect-ai…
1
96
Most AI demos show what the model can do. Terry Tao’s Jacobian Conjecture session shows what expert judgment adds: better questions, sharper rejection, and verification. chatgpt.com/share/6a5fdc7a-d… #AIEngineering #AIAgents
🤖 Made with AI
71
Was using claude fable and burnt through the Max 20x limit in under 30 min. RCA: fable spins up fable sub-agents for various tasks. Fix: Add this to you CLAUDE.md:
1
120
"Before spawning a subagent, choose the correct model tier for its task to optimize cost: use a small/fast model (e.g. Haiku-class) for search, summarization, and mechanical work; a mid-tier model (e.g. Sonnet-class) for routine implementation; and reserve the strongest model for complex reasoning, architecture, or debugging. Use the agent type's recommended default when one exists; do not pass the most expensive model by default."
1
81
Plot twist: the AI safety lab is now powered by Elon's rockets. He always wins in the end. @elonmusk
We’ve agreed to a partnership with @SpaceX that will substantially increase our compute capacity.
This, along with our other recent compute deals, means that we’ve been able to increase our usage limits for Claude Code and the Claude API.
93
The Jensen × Dwarkesh interview drops CoWoS, HBM, NVL72, systolic arrays, and offtake agreements in the first ten minutes.
If that sentence lost you, I built something for you: a 92-term companion deck, every entry linked to the exact timestamp.
2600th.github.io/jensen-dwar…
@dwarkesh_sp @nvidia #AI
148