Physical AI / Robotics GTM, Tech enthusiast/inventor, and avid hiker.
San Diego, CA
Joined April 2009
- Tweets4.3K
- Following579
- Followers1.1K
- Likes4.5K
Anthony Duca retweeted
If you use @AwakenTax to check your crypto, we used to have an app.
Now it’s back.
It has a lot more stuff. And it’s faster.
apps.apple.com/us/app/awaken…
Anthony Duca retweeted
Announcing the beta release of the Pi API. We’ve partnered with companies across diverse industries to adapt our models to their robots and applications, from harvesting strawberries to assembling servers. Now we’re opening access to more teams.
Anthony Duca retweeted
The next trillion dollar AI business will be an AI game studio
The video game industry is bigger than you can imagine. Larger than movies and music COMBINED
For YEARS now video games have absolutely SUCKED
The economics of video games became broken. They became too expensive to make
Developers demand hundreds of thousands of dollars a year, then refuse to use AI
It's completely killed the entire industry
People don't remember this, but there was only 1 year between Grand Theft Auto 3 and Grand Theft Auto Vice City. 362 days to be exact.
It's been THIRTEEN YEARS since the last Grand Theft Auto
From 1 year to 13 years. A completely broken industry
AI fixes this. AI allows one developer to have the power of a team of literally 1,000 developers
While the video game industry absolutely REFUSES to embrace innovation, that gives you the opportunity to strike
If you can make a game of any sort of quality right now with GPT 6 Astra, you can take a bite of a $213 BILLION dollar a year industry that would rather their company go bankrupt than use AI
Video games being all over your timeline the last 24 hours is a MAJOR signal
An entire industry is broken and begging to by disrupted by innovation
I'm legitimately starting my own game dev studio, and you 100% should consider it too
Anthony Duca retweeted
GPT-6 Astra did it. Asked it take an old 3d scan of my parents drawing room and model it from scratch in Blender. It didn't download any assets and sourced textures from my photogrammetry scan and make procedural shaders. Man did it do a nice job!
Anthony Duca retweeted
ChatGPT 6 Astra is the holy grail
It is by far the greatest AI model ever made, and can seriously be considered AGI
It one-shot the top down multiplayer extraction shooter for me. All graphics, sounds, gameplay done through Blender and Unity
Here's the most impressive part:
1. It's faster than GPT 5.6 sol
2. It did this in less credits
3. It did it with significantly higher quality
Cheaper per task, faster, significantly better.
Basically no trade offs. The trifecta. The holy grail.
It's picking up all my existing work and just destroying it. Crushing everything in its way. Finding novel insights, solutions no model has ever thought of.
Here's how I'd use it if I were you:
Include as much context, with as little direction as you can when prompting the model
What I mean by that is, let the model decide how to take action. Let the model breathe. Let it be autonomous. Let is find novel solutions.
Let's face it, the model is smarter than us. Us telling the model how to do work is stupid. It's like me telling Michael Jordan how to take shots.
Ask more questions than directions you give. You should be asking the model 10x more questions than directions.
Questions should be:
• How do you approach this problem?
• What are different solutions we can use?
• What are things I'm not thinking about here?
I call it reverse prompting. You should be relentlessly reverse prompting
What an incredible time to be alive
Anthony Duca retweeted
A New World Record 🤯
8.8 → 8.6 💨
in less than 3 days we’ve gone from +11 seconds to 8.6 seconds on the 100m sprint for humanoid robots.
Wild times.
Anthony Duca retweeted
Introducing Index
Today we're coming out of stealth with Index, the largest & most diverse robot dataset in the world
→ 30min of video uploads/sec
→ 16M video uploads
→ Paid $15M to date
→ 264k downloads
We're committed to spending $1B the next 12 months on data & compute
Anthony Duca retweeted
I love reading entrepreneur founding stories, and here’s Unitree’s from the perspective of the first angel institutional investor.
Unitree just had an explosive Chinese IPO but more than that, they are a world leader in humanoid robotics.
Many Americans assume Chinese industry is nurtured via central politburo edicts and cash. That is at best a very outdated view.
China has no less an important and vibrant Venture Capital community than we have in the US.
The story below documenting Unitree’s early funding, running on fumes, finding early traction, steering away from distractions, and making key pivots is identical to what you’d read about a Silicon Valley startup.
"Very often, it is not the founder catching up to the era. It is the era finally catching up to the founder."
- Jack Zhang (@Jackzzzz), recapping GeekPark's early bet on Unitree, nine years ago, on the day of its IPO.
Anthony Duca retweeted
Awaken is a seasonal SaaS product.
But I can already see it in the numbers.
This year, more people will use Awaken through Claude/Codex. Than web.
People are subscribing to higher $ plans just to use our APIs.
Aka our ARPU goes up bc you can use us through agents.
Future of software.
Anthony Duca retweeted
Trained a humanoid entirely in a 3D scan of the office.
Zero real-world fine-tuning. It just walked in and worked.
RL needs hundreds of thousands of attempts, and real robots can't afford to crash. A misjudged gap or a glass door collision breaks hardware and costs hours resetting. So you train in a sim.
But sim policies usually train on randomized, untextured geometry; depth is easy to fake. The robot learns structure, not the real world: no materials, no lighting, no idea what anything actually is. RGB cameras carry all of that but training RGB policies in generic fake worlds won’t generalize to the real world.
@NianticSpatial Scaniverse reconstructs your scan of the real deployment site. One 360° camera walkthrough → photorealistic 3D Gaussian splat at metric scale → collision mesh pulled from the same reconstruction, so vision and physics match exactly. Drops straight into NVIDIA Isaac Sim/Lab, no manual conversion.
@FlexionAI simulation-first approach then seamlessly enables the training of RGB-only nav policies inside that reconstruction. With added domain randomization + large image encoders for robustness, this deploys straight to hardware.
No real-world fine-tuning. Deployment: months of on-site adaptation → days.
Tune into the NVIDIA livestream on 12 August to hear how these companies are closing the sim2real gap: addevent.com/event/gvb8rkdrc…
@NVIDIARobotics
~~
♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com
Anthony Duca retweeted
Boring Company is about to explode.
They just closed about $3B in new funding and they are now aggressively hiring tons of people including business development (sales). This tells me they’ve reached the level of technical maturity they want so they can blitz municipalities worldwide to build Loop projects.
We currently know about Las Vegas (12 stations, 8 more by end of year, 104 within a few years), Nashville (40+ stations), and Dubai (20), but other cities are coming.
They currently have 7 tunnel boring machines in operation, with an additional 4 more anticipated to be put into service by the end of 2026. They currently build a new machine every two months.
Loop is Boring’s big urban transit innovation. It is a hybrid system, allowing EVs to pop out of tunnels and drive on surface streets for a bit. Stations can be small surface parking lots (one of Vegas’ new stations will be at a Mexican restaurant), or be placed beside a hotel’s entrance portico.
Small cheap stations mean you can build lots of stations resulting in a design rule of less than 100 feet for customers to walk as opposed to a subway’s 1/4 mile design target.
And unlike a subway, Loop is a point-to-point express system. No stopping at intermediate stations along the way or changing train lines in the Loop system.
AI? Absolutely. The boring machines are mostly autonomous now and the EVs driving passengers will be autonomous soon. Obviously these systems will eventually tightly integrate into Tesla’s Robotaxi network.
Being a Musk company, Boring has wrung tons of cost out of the system, making Boring more than 10x cheaper than legacy tunnel boring. Among other things, this has resulted in both Las Vegas and Nashville systems being completely privately funded through a combination of hotels paying for their own Loop station and rider fees that still cost less than an equivalent (but slower) taxi ride.
Any city investigating ways to lighten urban traffic load should be calling up Boring.
Anthony Duca retweeted
It’s become obvious that humans will not be reviewing code in the future.
We aren’t as good at it as top models.
The only reason it stays around is accountability.
But if you are slower bc you care about accountability.
And a competitor is faster (and the code is high quality so very few issues).
You lose.
Speed with quality trumps accountability in most industries.
Engineers and CTOs on X: I wrote this for you. yegge.ai/essays/the-shape-of…
Models and devs on X: I wrote this for you both. yegge.ai/essays/model-welfar…
Enjoy. Or not. Some of you definitely won't. But I invite you to debate it. The world's changing very fast now.
Anthony Duca retweeted
A foundational principle for automating complex machines operating in unstructured environments is what I call Tele Op-to-Autonomy.
All this means is that you should start off with 100% remote control (tele operation) and progressively add automations until the human is eventually 100% removed.
Why this works so well:
1) It allows you to get a product into the field crazy fast, which means you can start learning where all of your wrong assumptions are much sooner. This is by far the fastest path to test out product market fit, test tech efficacy, etc.
2) There is significant overlap in the tech needed to teleop and automate so the amount of “wasted work” towards the ultimate goal of full automation is actually quite low. In both cases, you will require:
- A fully drive by wire system
- A sensing suite capable of giving full situational awareness
- A remote support function for when things inevitably go wrong.
3) Development/iteration speed skyrockets when devs don’t have to eat the entire elephant at once. Imagine being a developer asked to automate some function for machine that is mission critical for the customer and you have the following two options:
Scenario 1 - Focus on automation first:
The devs are well aware that if they don’t deliver a solution with 99.9…% (pick your number of 9s) the customer will be impacted, their work will have caused relationship damage, and they effectively will have failed. As a result, devs start to focus on solving for the long tail of exceptions which exponentially increases time to market.
Scenario 2 - Focus on awesome teleop first:
You built your remote support systems first and thoroughly enough that someone anywhere in the world can strap into an awesome UI, get full situational awareness through strategically located streaming camera and sensor feeds, and take control over every aspect of the operation.
If you did this right there is also a mechanism to alert someone that a machine or process is “stuck” so the person can just get an alert, switch the UI over to that machine and seamlessly take control. In this scenario, someone watching the machine wouldn’t even know that control was handed off to a remote operator as the system keeps operating without the need for any local interventions.
In scenario 2 the automation strategy can start to be viewed through the metric of “# of support personnel per machine”. On day one that ratio will likely be 1:1. As you incrementally add automations that ratio can go to 1:2, the. 1:5, then 1:20, etc. and it can happen fast because each step of the way the developers writing the automations know that there is a graceful fallback plan for any exceptions so they don’t need to solve the long tail of potential issues before releasing the new feature.
Anthony Duca retweeted
Google DeepMind just introduced Gemini Robotics 2.
It's a single VLA that unlocks physical dexterity across different end effectors, hands or grippers, from one model checkpoint.
Apptronik's Apollo 2 humanoid, with the 22-DoF SharpaWave hand, ties knots and seals a ziplock bag
Anthony Duca retweeted
1900: Cars aren’t practical, roads are bad and there’s no place to refuel them.
1910: How is a single seater airplane going to catch on? You can’t transport enough people that way.
1920: Electric washing machines for the home don’t make sense, hardly anyone has a fuse box large enough for one.
1930: Television is way too expensive for the average household.
1940: Airplanes are fine for short distances, but they’ll never replace ships for crossing the oceans.
1960: Only big businesses will ever be able to afford a computer.
1970: The worldwide market for computers numbers in the thousands.
1980: The Internet is fine for academics, but it’ll never reach into the home.
1990: 10 Mbps Ethernet and 10 MB hard drives are plenty (as is 640k RAM).
This is a post about AI, data centers in space, Boring company, electric airplanes, self driving cars, humanoid robots, genomics, and all other early stage technologies.
They don’t all survive, but many grow up to change the world.
Anthony Duca retweeted
It’s our time.
