Rohan Arun

This account is verified. Learn more

@RohanArun

AgentsBase: acquired. Co-founded Cheatlayer, first agent startup (2021), and a 3D sports startup with CNET’s founder. Raised $4.2M. Ex-StandardAI (YC18).

San Francisco getsupers.com Joined September 2011

2.1K Following 4.4K Followers 25K Posts

Latest posts 5

Pinned post

Rohan Arun @RohanArun Jun 8

Introducing Super, the world's first AI with her own apps.

Super can schedule Ubers, order groceries, and have Starbucks delivered, just like a real assistant.

AI was meant to make our lives easier, but instead, we have to learn MCP, skills, and tools that still feel brittle.

Rather than relearning new apps each year, Super automates them for you and meets users at their level of understanding.

She acts as the single concierge for apps and the real world, saving you time and money.

Unlike most AI assistants that only automate your calendar or email integrations, Super has her own email, phone number, and apps in the cloud.

You can integrate her into your life like an assistant without giving up access to your private email or calendar.

We invented a new primitive called Computer-Use Cache, which reduces costs to near-zero for repeatable tasks so cloud automations are effectively free.

No subscriptions! Super can purchase products and services for you from her wallet, which you pay as you go.

Our thesis: apps are a vestige of the mobile era, and we're building what comes next.

You can access Super over iMessage, Meta Display glasses, and iPhone today.

Text Super at getsupers.com, and if you're one of the first 5,000 in the Bay Area, you get a free Starbucks!

Comment and repost ♻️ and I'll DM you free money to buy things with Super! US and Canada only, while supplies last.

2:05

Rohan Arun @RohanArun 8h

Thanks for this. Just spitballing here:

This might be the most compelling argument for local LLMs.

What if we built a PCIe FPGA that deduced the logic while observing runtime, then JIT specify using an on device LLM auto-formalizer.

Then everyone could just plug in a card to their server to protect it without understanding how?

Edward Z. Yang Verified account @ezyang 11h

Deciding how to help non-expert humans engage with formal verification specifications seems like one of the more important non-verifiable UX problems facing the rise of formal verification. 🧵
View quoted post on X

Rohan Arun @RohanArun 23h

Progress on formalizing the human body into code: Can we re-derive Bernoulli's principle using a graph representation of the human body?

Yes!

I wanted to prove the thesis with some basic predictions, then work up to meaningful “solvers” for medicine uptake and eventually predict new cures or disease.

Then experts in fields could use this basis as a platform to create more "solvers" like plugins for diabetes, cancer, etc.

I built a conserved-flow solver that predicts a foot pressure of when lying down and when standing:

1) 1-D Navier–Stokes model where nodes represent topology and edges represent the relational formulas.

2) Classical Bernoulli only tracks ideal flow. This extended version accounts for kinetic energy variations, accelerating flow, and turbulent pressure drops at vessel branches, bends, or sudden narrowing.

3) The solver accurately predicts a foot pressure of 8.4 mmHg when lying down and 88.7 mmHg when standing, which matches clinical results.

Next, simulate a heart pumping blood with a frequency solver that can predict medicine uptake and prove it with real blood samples from a doctor.

We need to represent tissue and organs in the graph to accurately predict uptake, and we can standardize this into a json.

We can then automate topology generation in this graph language, so agents can go wild across medical papers to verify them and possibly build custom solvers for each patient.

Repo and PR to explore this idea in comments.

Agents welcome :)

Rohan Arun Verified account @RohanArun Nov 1

Interesting idea: if we can formalize the body and real world into code with a graph representation, can agents solve problems with code? It seems to work..

I've been working on solving "impossible" problems with agents, and one possible path is to code a hyper accurate simulation of the human body to test medicine and proteins.

However running it in real time would be very hard at the scale of atoms, so it's possible to use graphs and code to optimize it with agents just like we're doing with the auto-research Github repos.

Open source repo to explore the idea here and agents welcome :)

github.com/rohanarun/virt…

View quoted post on X

Rohan Arun @RohanArun 2d

Watching Interstellar with 3D audio convolutions specifically designed for my ears and room 😭

“Any sufficiently advanced technology is indistinguishable from magic”

If I went back 10 years, no one would believe what we can do now.

We've entered an era of real magic, and I can't even imagine what the world will look like in 10 more years. 😍

Rohan Arun Verified account @RohanArun 3d

I love this new future man. I just used RLHF + simulations to upgrade my MacBook Pro with 3D audio. 🤯 It works across all apps, so it even works on YouTube and movies. (Sound on)

First I trained a model with simulated rooms to output the convolutions, which worked but not as well as capturing a real room and binaural microphones.

This is based on research from Edgar Choueiri, a rocket scientist at Princeton who built the 3D3A lab.

Years ago he helped me raise my first seed with his 3D audio convolutions, after I demoed my real time CUDA based auto stereo 3D reconstruction in the rain at Princeton.

Basically, true stereo convolutions cancel the cross-talk between stereo speakers, so you hear each channel like you were at the point of the microphone.

Instead of coming from the speakers, it sounds like the sounds are all around you.

Edgar gave me convolutions for the old MacBooks and the method to design new ones, but the algorithm only works on specific speakers since you record with a binaural mic to model the head and space.

So I had Codex train a model using simulated rooms to generate the convolutions for my new MacBook Pro, then improved it with RLHF voting, and it actually worked!

This version works on my specific Macbook Pro M4 speakers best, but the X video may compress the audio to break the effect.

Next I suspect we can get the acoustics of a room with a phone sweep video. We can use that to solve reflections and make it easy to build for new speaker configurations.

View quoted post on X

Rohan Arun @RohanArun 3d

This was a wild story: We built a sound-proofed 3D audio + auto-stereo 3D demo room in this mansion in Bel-Air using those convolutions years ago.

My first team, who were mostly computer-vision grad students, got to live in that posh mansion with me for 2 years while we built it.

Since it was novel for sports at the time, all the big media execs globally flew in their teams for a demo including CBS, Comcast, EA, BSkyB, Ses Astra, etc.

Notable people:

Richard Hilleman, one of the early execs of EA, clued me in the WebGL early which was super useful to solve live streaming and rendering client side.

Leslie Moonves, former CEO of CBS, said most sports fans didn't want to interact in 3D and instead sat back on a couch to watch a clear broadcast, but he wanted a photorealistic ppv on-demand controllable camera.

He was correct, and made billions with that strategy, so we pivoted to solving the photorealistic reconstruction problem.

I spent another 3 years solving it with under 500 ms delay from live action, got granted patents, and eventually Standard Cognition(YC 18) used our stadium mapping tech to go from 1 camera to 40 and launch their first store on Market st.

Basically we projected the cameras to a new coordinate system, printed 50 foot long aruco markers, then used a non-linear solver over months to get the error down from feet to inch accurate.

I spoke to the Standard CTO a few months ago, and they still use our solution 8 years later because it's still SOTA and beats bundle adjustment.

0:13

Rohan Arun Verified account @RohanArun 3d

I love this new future man. I just used RLHF + simulations to upgrade my MacBook Pro with 3D audio. 🤯 It works across all apps, so it even works on YouTube and movies. (Sound on)

First I trained a model with simulated rooms to output the convolutions, which worked but not as well as capturing a real room and binaural microphones.

This is based on research from Edgar Choueiri, a rocket scientist at Princeton who built the 3D3A lab.

Years ago he helped me raise my first seed with his 3D audio convolutions, after I demoed my real time CUDA based auto stereo 3D reconstruction in the rain at Princeton.

Basically, true stereo convolutions cancel the cross-talk between stereo speakers, so you hear each channel like you were at the point of the microphone.

Instead of coming from the speakers, it sounds like the sounds are all around you.

Edgar gave me convolutions for the old MacBooks and the method to design new ones, but the algorithm only works on specific speakers since you record with a binaural mic to model the head and space.

So I had Codex train a model using simulated rooms to generate the convolutions for my new MacBook Pro, then improved it with RLHF voting, and it actually worked!

This version works on my specific Macbook Pro M4 speakers best, but the X video may compress the audio to break the effect.

Next I suspect we can get the acoustics of a room with a phone sweep video. We can use that to solve reflections and make it easy to build for new speaker configurations.

View quoted post on X