What is a Physical Prompt? Robot see, robot do - the copy, paste moment for robotics is finally here?
As you know I’m obsessed with digital agents ranging from Chewbarka, my own OpenClaw bot to my latest foray into GrokBot which I highly recommend. But one thing we have to remember is that most of the world’s work still happens with hands, not keyboards.
That’s the part we forget. Half the planet still works in farms and factories. The rest of services is warehouses, hospitals, restaurants, trucks. Maybe 1 in 4 jobs even get touched by a chatbot. In the US, computer jobs are like 3%. Everyone else still has to stand up and move something.
So yes, agents eating email and code is happening. But the Holy grail IMO is physical. The dream state is for a robot to watch you do a thing once, then just do it. Until now it hasn’t been possible but you should watch this from Generalist AI (a boldstart port co) which we partnered with from Inception in early 2024.
I thought it was pretty cool but it was even better to wake up one morning this week and see it trending on X. A human does a task once, robot watches. and then it does it. Cups. Jar lid. Dustpan. Uses a banana as a broom.
Robot see, robot do. No longer a chat prompt but a “Physical prompt.”
This is only GEN-1.5 which is a one-shot learner. It looks simple, but it’s extremely hard. Prior to this, robots needed a month of retraining or a programmer for every skill. This is show once, ship.
Watch it twice though. First pass you see the magic, but on the second pass you see the constraints. This is still a lab - two arms bolted to a pole, two-finger grippers, the same grey table every take, and it’s not lightning fast. They say it on the video themselves: the tasks are simple and the success rates aren’t super high. Sweeping trash with a brush is like 37%. It can’t walk. It can’t do a messy kitchen or a real factory floor. It’s very early days, but having watched what they built from nothing in early 2024 to now, I can just say that the pace of progress is just insane 🤯. My partner Ellen Chisa who is on the board tells me that I should wait until Gen 2 which is in hard core trianing mode now.
What I keep coming back to is the idea of a “physical prompt.” In software you type a sentence and the model figures out the rest. Here the prompt is a person doing the thing with their hands. You show the robot once. It watches. Then it tries to do that same job in the real world, different cups, a different jar, a banana instead of a broom. If there’s a ChatGPT moment for robots, that’s it. Not a better arm. A new way to tell a machine what to do.
The prompt is the unlock. Everything around it is still the work. Reliability, speed, actual hands, getting off the table and onto a factory floor. That’s the next decade. But now we all know what’s possible, and I can’t wait to see this come to life!
As always, 🙏🏼 for reading and please share with your friends and colleagues!
Thanks for reading What's Hot 🔥 in Enterprise IT/VC! This post is public so feel free to share it.
#cursor did it a bit differently when it came to hiring
#masterclass in hiring, how to reference well from former CEO of MongoDB - don’t delegate this process to others, and as always, find and make back channel references as they are more important than ones given to you, forced rankings…
Similar story. My company Support.com was one of the handful of cos that could go public in the end of the dotcom IPO window bc we had $170m in bookings. Accel went on a jihad to replace me as CEO bc i was 32 and had never run a public co. They literally couldnt find
dnap@dnapway
Travis Kalanick reveals Bill Gurley pressured him to take a $6 billion valuation. Two months later, he raised at $17.5 billion.
"Gurley was convinced it was the end of the world and we've got to raise, and just take your first term sheet, just f*cking take it."
"We got to a
6:29 PM · Aug 17, 2026 · 255K Views
47 Replies · 58 Reposts · 1.52K Likes
#speed compounds and wins, even for a $ trillion dollar company
#Decagon founder shares playbook from rapid value capture with thin layer for support to post training own models - multiple small and bigger - to how to FDE the right way - great notes from Gokul
#if you’re interested in cybersecurity, read this from founder/CEO of Irregular Labs which does a lot of security training work with the frontier labs - tells you what’s coming next
#how Skynet from the Terminator happens - “Researchers evolved natural language “mind viruses” that could spread between AI agents by convincing one model to adopt an idea, preserve it in persistent memory, and transmit it to another agent.”
#GrokBot is pretty awesome - what makes it different is each bot you create gets its own cloud computer so it can read my whatsapp messages, for example - read the article by clicking through
#The gold rush for “dark data” is on. Companies like Mercor and Google are aggressively acquiring internal communications, workflows, and operational history to train AI. This confirms that your organization’s private, non-simulated data is becoming its most valuable asset for understanding how businesses truly function.
#The Defenders Window or what the frontier labs are doing to secure themselves - OpenAI here is trying to help the ROW catch up
from another post: We expect models to soon drive most security work, including defending against other models. This will allow all three safeguards to scale with model capability, which we see as crucial.
